Expose tray runtime libraries in the Nix development shell

The development shell inherited build dependencies but Cargo-built tray
binaries lack the installed package's wrapper. AppIndicator was present in
the store yet its dlopen lookup failed because the runtime library path was
missing. The same problem affected dynamically loaded graphics libraries.

Expose the package's existing trayRuntimeLibs through passthru and reuse that
list in mkShell packages and its LD_LIBRARY_PATH hook. Include the existing
NixOS OpenGL driver path and preserve any caller-supplied library paths.
The development shell and installed tray now share one dependency list.
Document direct execution of Cargo-built tray binaries inside nix develop.
No agent, protocol, package version or system configuration changes.

Validation on Linux:
- Reproduced the missing AppIndicator dlopen failure before the change.
- Loaded libayatana-appindicator3.so.1, libxkbcommon-x11.so.0, libEGL.so.1,
  libGL.so.1 and libwayland-client.so.0 inside the updated shell.
- Verified AppIndicator still loads with an inherited LD_LIBRARY_PATH and
  that the existing value is preserved. The first probe used an older host
  Python with an incompatible glibc; probes using Python from the pinned
  nixpkgs input passed.
- Evaluated the x86_64-linux CLI package and shared runtime dependency list,
  and the aarch64-linux development shell; no aarch64 execution claimed.
- nixfmt --check flake.nix nix/package.nix
- cargo fmt --all -- --check
- cargo clippy --locked --workspace --all-targets -- -D warnings
- cargo test --locked --workspace --all-targets: 566 passed, 3 ignored
Cargo checks ran through nix develop, with TSUNAGI_CONTROL_SOCKET unset for
Clippy and tests. Existing agent and tray processes were not restarted.

.cargo/config.toml

@@ -0,0 +1,5 @@
+# Windows: link the C runtime statically, so the agent and the tray start on a
+# machine that has no Visual C++ redistributable installed. The installer ships
+# nothing but our own files and Wintun.
+[target.x86_64-pc-windows-msvc]
+rustflags = ["-C", "target-feature=+crt-static"]

.gitea/workflows/release.yml

@@ -0,0 +1,350 @@
+name: Release
+
+# Build the agent for every supported target on a version tag, leave each build
+# as a run artifact, and then — only after every target has built — attach the
+# archives to the release. Linux is a fully static musl binary (built with
+# cargo-zigbuild) so one artifact runs on any distro and can be repackaged
+# however a distro prefers.
+#
+# Targeted at Gitea Actions: artifacts use @v3 (the v4 backend is not available
+# on Gitea/GHES), and the release is filled through Gitea's own API with curl
+# rather than `gh`, which speaks only to github.com.
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+  CARGO_NET_RETRY: "10"
+
+jobs:
+  build:
+    name: ${{ matrix.asset }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            asset: tsunagi-linux-x86_64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: amd64
+            pkg_arch: x86_64
+            gnu_target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            asset: tsunagi-linux-aarch64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: arm64
+            pkg_arch: aarch64
+          # 32-bit ARM with hardware float (ARMv7, `v7l`): single-board and edge
+          # devices. Statically linked like the others.
+          - os: ubuntu-latest
+            target: armv7-unknown-linux-musleabihf
+            asset: tsunagi-linux-armv7
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: armhf
+            pkg_arch: armv7h
+          - os: macos-latest
+            target: aarch64-apple-darwin
+            asset: tsunagi-macos-aarch64
+            archive: tar.gz
+            zig: false
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            asset: tsunagi-windows-x86_64
+            archive: zip
+            zig: false
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+
+      # aws-lc-sys (iroh's rustls backend) and the bundled sqlite are C, and the
+      # musl targets have no prebuilt bindings, so bindgen needs libclang and the
+      # C build needs cmake. zig is the cross C compiler cargo-zigbuild drives.
+      - name: Install musl build and packaging tooling
+        if: matrix.zig
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake clang libclang-dev libarchive-tools zstd
+      - name: Install zig
+        if: matrix.zig
+        uses: mlugg/setup-zig@v2
+        with:
+          # Pin a real, archived Zig release. Left unpinned the action resolves
+          # to the newest version it knows of (a yet-unreleased 0.16.0-dev),
+          # whose tarball 404s on ziglang.org while the community mirrors flake.
+          version: 0.14.1
+      - name: Install cargo-zigbuild
+        if: matrix.zig
+        uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-zigbuild
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ matrix.zig }}" == "true" ]]; then
+            # zig's linker ignores the `-O1` GNU-ld optimization flag Rust passes
+            # and says so on stderr, which rustc now surfaces as a linker_messages
+            # warning. It is cosmetic — the binary is correct — so quiet just that
+            # one lint to keep the musl build log clean. A real link error still
+            # fails the build.
+            export RUSTFLAGS="${RUSTFLAGS:-} -A linker_messages"
+            cargo zigbuild --release --locked --target "${{ matrix.target }}" --bin tsng
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsng
+          fi
+
+      # The tray GUI. On Linux it is built against glibc (an old baseline for
+      # portability), not musl: a GUI must dlopen libGL/X11/Wayland at runtime,
+      # which a static musl binary cannot. macOS and Windows build it natively.
+      - name: Build the tray GUI
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              # tray-icon links GTK and libappindicator through pkg-config, which
+              # has no cross story under zig, so the Linux GUI is built natively
+              # (x86_64 only) against the runner's glibc.
+              sudo apt-get update
+              sudo apt-get install -y --no-install-recommends \
+                libgtk-3-dev libayatana-appindicator3-dev libxdo-dev
+              cargo build --release --locked \
+                --target "${{ matrix.gnu_target }}" --bin tsunagi-tray
+            else
+              echo "no GUI build for ${{ matrix.target }}"
+            fi
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsunagi-tray
+          fi
+
+      - name: Package
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          version="${GITHUB_REF_NAME#v}"
+          stage="dist/pkg/${{ matrix.asset }}"
+          mkdir -p "$stage"
+
+          bin="tsng"
+          [[ "${{ runner.os }}" == "Windows" ]] && bin="tsng.exe"
+          cp "target/${{ matrix.target }}/release/${bin}" "$stage/"
+
+          # The tray GUI alongside it, where it was built (Linux: x86_64 only).
+          tray="tsunagi-tray"
+          [[ "${{ runner.os }}" == "Windows" ]] && tray="tsunagi-tray.exe"
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              cp "target/${{ matrix.gnu_target }}/release/${tray}" "$stage/"
+            fi
+          else
+            cp "target/${{ matrix.target }}/release/${tray}" "$stage/"
+          fi
+
+          # The top-level README, plus the per-OS install notes and integration
+          # files, so the archive carries what the platform needs to set up.
+          cp README.md "$stage/"
+          case "${{ runner.os }}" in
+            Linux)
+              cp dist/linux/tsunagi.service dist/linux/install.sh \
+                 dist/linux/50-tsunagi-resolved.rules "$stage/"
+              cp dist/linux/README.md "$stage/INSTALL.md"
+              ;;
+            macOS)
+              cp dist/macos/cy.hexor.tsunagi.plist "$stage/"
+              cp dist/macos/README.md "$stage/INSTALL.md"
+              # The tray as an application bundle, for the Homebrew cask.
+              bash dist/macos/build-app.sh \
+                "target/${{ matrix.target }}/release/tsunagi-tray" \
+                "$version" "$stage/Tsunagi.app"
+              ;;
+            Windows)
+              cp dist/windows/README.md "$stage/INSTALL.md"
+              ;;
+          esac
+
+          archive="${{ matrix.asset }}-${version}.${{ matrix.archive }}"
+          if [[ "${{ matrix.archive }}" == "zip" ]]; then
+            (cd dist/pkg && 7z a "../../${archive}" "${{ matrix.asset }}")
+          else
+            tar -C dist/pkg -czf "${archive}" "${{ matrix.asset }}"
+          fi
+
+      # The Windows installer: the agent as a service, Wintun, the tray and the
+      # PATH entry in one setup program. Wintun is the signed library from
+      # wintun.net, pinned by hash so a changed download fails the build.
+      - name: Build the Windows installer
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $version = $env:GITHUB_REF_NAME.TrimStart('v')
+          Invoke-WebRequest -Uri https://www.wintun.net/builds/wintun-0.14.1.zip -OutFile wintun.zip
+          $hash = (Get-FileHash wintun.zip -Algorithm SHA256).Hash
+          if ($hash -ne '07C256185D6EE3652E09FA55C0B673E2624B565E02C4B9091C79CA7D2F24EF51') {
+            throw "wintun.zip has an unexpected hash: $hash"
+          }
+          Expand-Archive wintun.zip -DestinationPath wintun-dl
+          choco install innosetup -y --no-progress
+          $iscc = Join-Path ${env:ProgramFiles(x86)} 'Inno Setup 6\ISCC.exe'
+          $stage = Join-Path $PWD 'dist/pkg/${{ matrix.asset }}'
+          & $iscc "/DAppVersion=$version" "/DSourceDir=$stage" `
+            "/DWintunDir=$(Join-Path $PWD 'wintun-dl/wintun/bin/amd64')" `
+            "/O$PWD" dist/windows/tsunagi.iss
+          if ($LASTEXITCODE -ne 0) { throw "ISCC failed with $LASTEXITCODE" }
+
+      # A dependency-free .deb for the x86_64 and aarch64 Linux builds: it drops
+      # the static binary, the systemd service and the polkit rule. Built with
+      # dpkg-deb (present on the runner); see dist/deb/build-deb.sh.
+      - name: Build the Debian package
+        if: matrix.deb
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" \
+            "$version" \
+            "tsunagi_${version}_${{ matrix.deb_arch }}.deb"
+
+      # A dependency-free Arch package (.pkg.tar.zst), assembled with bsdtar and
+      # zstd (no makepkg); same binary, service and polkit rule. See
+      # dist/arch/build-pkg.sh.
+      - name: Build the Arch package
+        if: matrix.pkg_arch
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" \
+            "$version" \
+            "tsunagi-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst"
+
+      # The GUI variants: the same agent plus the tray binary and a desktop
+      # launcher, named tsunagi-gui (conflicting with the headless packages).
+      - name: Build the GUI packages
+        if: matrix.gnu_target
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          tray="target/${{ matrix.gnu_target }}/release/tsunagi-tray"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" "$version" \
+            "tsunagi-gui_${version}_${{ matrix.deb_arch }}.deb" "$tray"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" "$version" \
+            "tsunagi-gui-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst" "$tray"
+
+      # @v3 on purpose: Gitea does not support the v4 artifact backend. This is
+      # what makes the archive (and the .deb, where built) show up under the run
+      # and hands it to `publish`.
+      - name: Upload the build outputs as run artifacts
+        uses: actions/upload-artifact@v3
+        with:
+          name: ${{ matrix.asset }}
+          path: |
+            ${{ matrix.asset }}-*.*
+            *.deb
+            *.pkg.tar.zst
+            tsunagi-setup-*.exe
+          if-no-files-found: error
+
+  publish:
+    name: Publish release
+    needs: build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Download every build artifact
+        uses: actions/download-artifact@v3
+        with:
+          path: artifacts
+
+      - name: Show what was collected
+        run: find artifacts -type f
+
+      - name: Create the release and upload the archives
+        shell: bash
+        env:
+          # Gitea injects this with access to the current repository. If the
+          # release calls below return 403, the repo/runner is configured to
+          # hand Actions a read-only token — grant it write access to releases.
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -u
+
+          tag="${GITHUB_REF_NAME}"
+          # Gitea sets GITHUB_API_URL to its own API root (…/api/v1); fall back
+          # to building it from the server URL if it is absent.
+          base="${GITHUB_API_URL:-${GITHUB_SERVER_URL}/api/v1}"
+          api="${base}/repos/${GITHUB_REPOSITORY}"
+          auth="Authorization: token ${TOKEN}"
+          echo "release API: $api"
+
+          prerelease=false
+          case "$tag" in *-*) prerelease=true ;; esac
+
+          # The release id for this tag, if it already exists.
+          code=$(curl -sS -o resp.json -w '%{http_code}' -H "$auth" "$api/releases/tags/$tag" || true)
+          echo "get release by tag: HTTP $code"
+          id=$(grep -o '"id":[0-9]*' resp.json 2>/dev/null | head -n1 | cut -d: -f2 || true)
+
+          if [ -z "$id" ]; then
+            code=$(curl -sS -o resp.json -w '%{http_code}' -X POST -H "$auth" \
+              -H 'Content-Type: application/json' \
+              -d "{\"tag_name\":\"$tag\",\"name\":\"tsunagi $tag\",\"prerelease\":$prerelease}" \
+              "$api/releases" || true)
+            echo "create release: HTTP $code"
+            id=$(grep -o '"id":[0-9]*' resp.json 2>/dev/null | head -n1 | cut -d: -f2 || true)
+            if [ -z "$id" ]; then
+              echo "could not create the release; response was:" >&2
+              cat resp.json >&2 || true
+              exit 1
+            fi
+          fi
+          echo "release id: $id"
+
+          # download-artifact@v3 nests each artifact in a directory named after
+          # its upload, with the single archive inside. The read loop runs in
+          # this shell (process substitution, not a pipe) so a failed upload
+          # actually fails the job.
+          rc=0
+          while IFS= read -r -d '' f; do
+            name=$(basename "$f")
+            code=$(curl -sS -o up.json -w '%{http_code}' -X POST -H "$auth" \
+              -F "attachment=@${f}" \
+              "$api/releases/$id/assets?name=${name}" || true)
+            echo "upload $name: HTTP $code"
+            if [ "$code" -lt 200 ] || [ "$code" -ge 300 ]; then
+              cat up.json >&2 || true
+              rc=1
+            fi
+          done < <(find artifacts -type f \( -name '*.tar.gz' -o -name '*.zip' -o -name '*.deb' -o -name '*.pkg.tar.zst' -o -name 'tsunagi-setup-*.exe' \) -print0)
+          exit "$rc"

.github/workflows/release.yml

@@ -0,0 +1,322 @@
+name: Release
+
+# GitHub version of the release pipeline. The build half mirrors the Gitea one
+# in .gitea/workflows/release.yml (static musl Linux via cargo-zigbuild, native
+# macOS and Windows); only the publish half differs: it uses the current
+# artifact actions (Node 24) and fills the release with `gh`.
+#
+# Every job is gated on `github.server_url == 'https://github.com'`, so this file
+# runs only on GitHub. Gitea runs its own copy in .gitea/workflows (which pins
+# the older @v3 artifact actions it supports); should a Gitea instance also read
+# .github/workflows, the guard skips these jobs there so they never fail on the
+# artifact backend or `gh`.
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+  CARGO_NET_RETRY: "10"
+
+jobs:
+  build:
+    name: ${{ matrix.asset }}
+    if: ${{ github.server_url == 'https://github.com' }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            asset: tsunagi-linux-x86_64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: amd64
+            pkg_arch: x86_64
+            gnu_target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            asset: tsunagi-linux-aarch64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: arm64
+            pkg_arch: aarch64
+          # 32-bit ARM with hardware float (ARMv7, `v7l`): single-board and edge
+          # devices. Statically linked like the others.
+          - os: ubuntu-latest
+            target: armv7-unknown-linux-musleabihf
+            asset: tsunagi-linux-armv7
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: armhf
+            pkg_arch: armv7h
+          - os: macos-latest
+            target: aarch64-apple-darwin
+            asset: tsunagi-macos-aarch64
+            archive: tar.gz
+            zig: false
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            asset: tsunagi-windows-x86_64
+            archive: zip
+            zig: false
+
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+
+      # aws-lc-sys (iroh's rustls backend) and the bundled sqlite are C, and the
+      # musl targets have no prebuilt bindings, so bindgen needs libclang and the
+      # C build needs cmake. zig is the cross C compiler cargo-zigbuild drives.
+      - name: Install musl build and packaging tooling
+        if: matrix.zig
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake clang libclang-dev libarchive-tools zstd
+      - name: Install zig
+        if: matrix.zig
+        uses: mlugg/setup-zig@v2
+        with:
+          # Pin a real, archived Zig release. Left unpinned the action resolves
+          # to the newest version it knows of (a yet-unreleased 0.16.0-dev),
+          # whose tarball 404s on ziglang.org while the community mirrors flake.
+          version: 0.14.1
+      - name: Install cargo-zigbuild
+        if: matrix.zig
+        uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-zigbuild
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ matrix.zig }}" == "true" ]]; then
+            # zig's linker ignores the `-O1` GNU-ld optimization flag Rust passes
+            # and says so on stderr, which rustc now surfaces as a linker_messages
+            # warning. It is cosmetic — the binary is correct — so quiet just that
+            # one lint to keep the musl build log clean. A real link error still
+            # fails the build.
+            export RUSTFLAGS="${RUSTFLAGS:-} -A linker_messages"
+            cargo zigbuild --release --locked --target "${{ matrix.target }}" --bin tsng
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsng
+          fi
+
+      # The tray GUI. On Linux it is built against glibc (an old baseline for
+      # portability), not musl: a GUI must dlopen libGL/X11/Wayland at runtime,
+      # which a static musl binary cannot. macOS and Windows build it natively.
+      - name: Build the tray GUI
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              # tray-icon links GTK and libappindicator through pkg-config, which
+              # has no cross story under zig, so the Linux GUI is built natively
+              # (x86_64 only) against the runner's glibc.
+              sudo apt-get update
+              sudo apt-get install -y --no-install-recommends \
+                libgtk-3-dev libayatana-appindicator3-dev libxdo-dev
+              cargo build --release --locked \
+                --target "${{ matrix.gnu_target }}" --bin tsunagi-tray
+            else
+              echo "no GUI build for ${{ matrix.target }}"
+            fi
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsunagi-tray
+          fi
+
+      - name: Package
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          version="${GITHUB_REF_NAME#v}"
+          stage="dist/pkg/${{ matrix.asset }}"
+          mkdir -p "$stage"
+
+          bin="tsng"
+          [[ "${{ runner.os }}" == "Windows" ]] && bin="tsng.exe"
+          cp "target/${{ matrix.target }}/release/${bin}" "$stage/"
+
+          # The tray GUI alongside it, where it was built (Linux: x86_64 only).
+          tray="tsunagi-tray"
+          [[ "${{ runner.os }}" == "Windows" ]] && tray="tsunagi-tray.exe"
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              cp "target/${{ matrix.gnu_target }}/release/${tray}" "$stage/"
+            fi
+          else
+            cp "target/${{ matrix.target }}/release/${tray}" "$stage/"
+          fi
+
+          # The top-level README, plus the per-OS install notes and integration
+          # files, so the archive carries what the platform needs to set up.
+          cp README.md "$stage/"
+          case "${{ runner.os }}" in
+            Linux)
+              cp dist/linux/tsunagi.service dist/linux/install.sh \
+                 dist/linux/50-tsunagi-resolved.rules "$stage/"
+              cp dist/linux/README.md "$stage/INSTALL.md"
+              ;;
+            macOS)
+              cp dist/macos/cy.hexor.tsunagi.plist "$stage/"
+              cp dist/macos/README.md "$stage/INSTALL.md"
+              # The tray as an application bundle, for the Homebrew cask.
+              bash dist/macos/build-app.sh \
+                "target/${{ matrix.target }}/release/tsunagi-tray" \
+                "$version" "$stage/Tsunagi.app"
+              ;;
+            Windows)
+              cp dist/windows/README.md "$stage/INSTALL.md"
+              ;;
+          esac
+
+          archive="${{ matrix.asset }}-${version}.${{ matrix.archive }}"
+          if [[ "${{ matrix.archive }}" == "zip" ]]; then
+            (cd dist/pkg && 7z a "../../${archive}" "${{ matrix.asset }}")
+          else
+            tar -C dist/pkg -czf "${archive}" "${{ matrix.asset }}"
+          fi
+
+      # The Windows installer: the agent as a service, Wintun, the tray and the
+      # PATH entry in one setup program. Wintun is the signed library from
+      # wintun.net, pinned by hash so a changed download fails the build.
+      - name: Build the Windows installer
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $version = $env:GITHUB_REF_NAME.TrimStart('v')
+          Invoke-WebRequest -Uri https://www.wintun.net/builds/wintun-0.14.1.zip -OutFile wintun.zip
+          $hash = (Get-FileHash wintun.zip -Algorithm SHA256).Hash
+          if ($hash -ne '07C256185D6EE3652E09FA55C0B673E2624B565E02C4B9091C79CA7D2F24EF51') {
+            throw "wintun.zip has an unexpected hash: $hash"
+          }
+          Expand-Archive wintun.zip -DestinationPath wintun-dl
+          choco install innosetup -y --no-progress
+          $iscc = Join-Path ${env:ProgramFiles(x86)} 'Inno Setup 6\ISCC.exe'
+          $stage = Join-Path $PWD 'dist/pkg/${{ matrix.asset }}'
+          & $iscc "/DAppVersion=$version" "/DSourceDir=$stage" `
+            "/DWintunDir=$(Join-Path $PWD 'wintun-dl/wintun/bin/amd64')" `
+            "/O$PWD" dist/windows/tsunagi.iss
+          if ($LASTEXITCODE -ne 0) { throw "ISCC failed with $LASTEXITCODE" }
+
+      # A dependency-free .deb for the x86_64 and aarch64 Linux builds: it drops
+      # the static binary, the systemd service and the polkit rule. Built with
+      # dpkg-deb (present on the runner); see dist/deb/build-deb.sh.
+      - name: Build the Debian package
+        if: matrix.deb
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" \
+            "$version" \
+            "tsunagi_${version}_${{ matrix.deb_arch }}.deb"
+
+      # A dependency-free Arch package (.pkg.tar.zst), assembled with bsdtar and
+      # zstd (no makepkg); same binary, service and polkit rule. See
+      # dist/arch/build-pkg.sh.
+      - name: Build the Arch package
+        if: matrix.pkg_arch
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" \
+            "$version" \
+            "tsunagi-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst"
+
+      # The GUI variants: the same agent plus the tray binary and a desktop
+      # launcher, named tsunagi-gui (conflicting with the headless packages).
+      - name: Build the GUI packages
+        if: matrix.gnu_target
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          tray="target/${{ matrix.gnu_target }}/release/tsunagi-tray"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" "$version" \
+            "tsunagi-gui_${version}_${{ matrix.deb_arch }}.deb" "$tray"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" "$version" \
+            "tsunagi-gui-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst" "$tray"
+
+      - name: Upload the build outputs as run artifacts
+        uses: actions/upload-artifact@v7
+        with:
+          name: ${{ matrix.asset }}
+          path: |
+            ${{ matrix.asset }}-*.*
+            *.deb
+            *.pkg.tar.zst
+            tsunagi-setup-*.exe
+          if-no-files-found: error
+
+  publish:
+    name: Publish release
+    needs: build
+    if: ${{ github.server_url == 'https://github.com' }}
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Download every build artifact
+        uses: actions/download-artifact@v8
+        with:
+          path: artifacts
+
+      - name: Show what was collected
+        run: find artifacts -type f
+
+      - name: Create the release and upload the archives
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+          TAG: ${{ github.ref_name }}
+        run: |
+          set -euo pipefail
+
+          # download-artifact nests each artifact under its own directory, so
+          # the archives are artifacts/<asset>/<file>.
+          mapfile -t files < <(find artifacts -type f \( -name '*.tar.gz' -o -name '*.zip' -o -name '*.deb' -o -name '*.pkg.tar.zst' -o -name 'tsunagi-setup-*.exe' \))
+          printf 'asset: %s\n' "${files[@]}"
+
+          prerelease=()
+          case "$TAG" in *-*) prerelease=(--prerelease) ;; esac
+
+          # Create the release on the first run; on a re-run, add/replace assets
+          # on the release that is already there.
+          if gh release view "$TAG" --repo "$GITHUB_REPOSITORY" >/dev/null 2>&1; then
+            gh release upload "$TAG" "${files[@]}" --repo "$GITHUB_REPOSITORY" --clobber
+          else
+            gh release create "$TAG" "${files[@]}" \
+              --repo "$GITHUB_REPOSITORY" \
+              --title "tsunagi $TAG" \
+              --generate-notes \
+              --verify-tag \
+              "${prerelease[@]}"
+          fi

.gitignore

@@ -1 +1,2 @@
 /target
+/.claude

AGENTS.md

@@ -8,36 +8,96 @@ before touching the code. Do not restate the full design here; follow the links.
 A proof-of-concept agent library for small private mesh networks. A user
 supplies a network name and one shared secret; agents derive the same network
 space independently, find each other, prove membership and exchange control
-messages. Scope and non-scope are in [README.md](README.md).
+messages. What it does and does not do is in [docs/architecture.md](docs/architecture.md) and [docs/threat-model.md](docs/threat-model.md).
 
 There is no central server, no owner, no registration and no majority vote.
 Design accordingly: a majority is not a root of trust.
 
+## User experience
+
+User convenience is a core requirement. The CLI should feel logical and
+predictable to an experienced administrator: use consistent selectors and
+flags, sensible defaults, and actionable errors. Let users refer to configured
+networks by their exact name as well as by id or a unique id prefix; do not
+make them copy an opaque identifier when the stored name identifies one network.
+When a selector is ambiguous, list the matching names and ids and change
+nothing. Never guess which resource a destructive command should affect.
+Keep help text and examples in sync with behavior, and preserve existing
+scripts when improving the interactive workflow.
+
 ## Architectural boundaries
 
 Keep these separate. Crossing them is the main thing to review for.
 
 - **Control plane vs data plane.** The separation is **logical, not physical**.
-  The control protocol in `src/proto/` knows nothing about packets, and
-  `src/dataplane/` knows nothing about the control protocol; either can be
-  replaced on its own. Both may ride on iroh — refusing to would throw away
-  iroh's NAT traversal and force the data plane to reimplement it. They use
-  different ALPNs and different connections, so a busy or broken data plane
-  cannot disturb control traffic.
+  The control protocol in `crates/tsunagi-core/src/proto/` knows nothing about
+  packets, and a protocol crate knows nothing about the control protocol;
+  either can be replaced on its own. Both may ride on iroh — refusing to
+  would throw away iroh's NAT traversal and force the data plane to
+  reimplement it. They use different ALPNs and different connections, so a
+  busy or broken data plane cannot disturb control traffic.
+- **One agent, one interface.** It belongs to the system level, along with
+  the addresses on it and the decision of whose packet is whose. Several
+  protocols may be carrying traffic at once and none of them owns the thing
+  they carry it for. A protocol is handed a routed packet and hands back a
+  decrypted one; it never creates an interface and never picks an address.
+- **One state directory, one identity, one live agent — many networks.** A
+  network is added to the agent that is already running, never by starting a
+  second one on the same directory. A second agent is a second identity and
+  is isolated: its own directories, its own interface, its own runtime. Do
+  not add anything that lets two agents share a directory, and do not make a
+  network's lifetime depend on the process that happened to start it.
+- **A protocol is a separate crate with its own version.** The version peers
+  compare is the *wire* version, never the software version: two peers on
+  different releases work together for as long as the bytes between them
+  have not changed. Nothing negotiated may be derived from anything that
+  moves with a release.
+- **A protocol never sees the network secret.** It proves who is at the other
+  end of a tunnel; proving membership of a network stays in the core, which
+  is why the authenticated transport does too.
+- **A relay carries bytes, never packets.** A datagram passed on for two
+  other peers goes link in, link out: it is never written to the middle's
+  interface or decrypted there. Multihop routing has a bounded hop limit;
+  equal-cost next hops are chosen per flow. Reachability is first-hand and
+  volatile: each authenticated member advertises its own direct links, of
+  whatever kind. A link is a way to reach a neighbour and a protocol is an
+  end-to-end tunnel to a peer; the two meet only in the envelope, so a
+  tunnel never depends on which transports carry it. Build routing tables on topology changes, never per packet. Transit
+  must not acquire a routing mutex or wait for a protocol/TUN reader.
+- **Broadcast fanout belongs at local IP ingress.** Participation is local to
+  each network, enabled by default, persisted and advertised to authenticated
+  peers. Resolve one source domain and target only its willing members. Remote
+  delivery never triggers another fanout. Physical LAN exports must extend
+  explicit ingress and source/destination admission policies; never bypass
+  ownership checks or add broadcast flooding to encrypted transit.
 - **Plugins never learn reachability.** An `IpPlugin` is handed a `PacketLink`
   per peer and moves datagrams over it. Addresses, hole punching and relays
-  belong to `src/dataplane/transport/`. A plugin announcement says *who*, never
+  belong to `crates/tsunagi-core/src/dataplane/transport/`. A plugin announcement says *who*, never
   *where*.
-- **The core never parses a plugin payload.** See `src/dataplane/mod.rs`. Only
-  `src/dataplane/wireguard/announcement.rs` interprets WireGuard payloads, and
+- **The core never parses a plugin payload.** See `crates/tsunagi-core/src/dataplane/mod.rs`. Only
+  `crates/tsunagi-wg-quic/src/announcement.rs` interprets `wg-quic` payloads, and
   only after bounding every field. A data plane failure must never stop the
   control plane.
 - **Derived, not claimed.** A peer's overlay address is derived from its public
   key. Outbound packets are routed to the owner of the destination address;
   inbound packets are dropped unless their source is the address derived for
   the peer that sent them. Never trust an address a peer announces.
+- **Signed state is the only durable agreement.** A fact that must survive a
+  participant being away goes in `crates/tsunagi-core/src/state/` as a record signed by its
+  author, never in a session. Merging is deterministic, an older version never
+  rolls back a newer one, and absence from a snapshot is not deletion. Never
+  add a vote or a quorum: a majority is not a trust root here, and it would
+  stall with one peer online.
+- **A record and the author's own version counter commit together**, in one
+  transaction, **before** the record is announced.
 - **Plugins own their system objects.** A plugin creates and removes its own
   interface and nothing else. Never touch routing, DNS or firewall settings.
+  The one exception is the system level, not a plugin: while a network's
+  broadcast is on, `overlay::hostrules` may install exactly a
+  `255.255.255.255/32` route and an inbound UDP firewall allowance (overlay range
+  only) for the interface the agent owns. Both are tagged `tsunagi:<interface>`,
+  reversible, removed on broadcast off, leave and shutdown, and computed from
+  local state, never from a remote announcement.
 - **Device identity vs network identity.** The iroh endpoint id is the device's
   public key. `NetworkId` is derived from name + secret only. Never conflate
   them, and never let one change the other.
@@ -57,7 +117,7 @@ Keep these separate. Crossing them is the main thing to review for.
    normal path *and* the important failures. See [docs/testing.md](docs/testing.md).
    Do not add tests for getters or to move a coverage number.
 2. **Never change `IDENTITY_SCHEME`, the derivation labels, or the transcript
-   encoding** in `src/identity/network.rs` and `src/proto/handshake.rs` without
+   encoding** in `crates/tsunagi-core/src/identity/network.rs` and `crates/tsunagi-core/src/proto/handshake.rs` without
    treating it as an incompatible protocol change. Bumping the crate version or
    the control protocol version must not change an existing `NetworkId`.
 3. **Secrets never leak.** Not into logs, not into `Debug`, not into status
@@ -68,7 +128,7 @@ Keep these separate. Crossing them is the main thing to review for.
    the library; tests opt out explicitly at the top of each file.
 5. **Bounds before allocation.** Frame lengths are checked against
    `Limits::max_frame_len` before a buffer is allocated. Every string, list and
-   queue has a limit in `src/config.rs`.
+   queue has a limit in `crates/tsunagi-core/src/config.rs`.
 6. **Failure is contained.** A bad signature, wrong secret, malformed packet or
    unknown version rejects one message or one session. It never stops another
    network and never stops the agent. There is no irreversible global error
@@ -85,24 +145,35 @@ Keep these separate. Crossing them is the main thing to review for.
    up shutdown. Wind tasks down with a grace period and then abort.
 10. **Nothing from a remote announcement becomes a shell command, a filesystem
     path or an OS setting.** The agent never touches interfaces or OS settings
-    it did not create.
+    it did not create (the tagged broadcast host rules above are for its own
+    interface).
 
 ## Where things live
 
 | path                | responsibility |
 |---------------------|----------------|
-| `src/identity/`     | device identity; deterministic network space identity and derived keys |
-| `src/storage/`      | `state.sqlite`, `cache.sqlite`, directory ownership lock |
-| `src/discovery.rs`  | candidate sources; test and static backends |
-| `src/proto/`        | framing, message formats, membership handshake |
-| `src/net.rs`        | iroh endpoint adapter and observability snapshots |
-| `src/agent/`        | agent lifecycle, per-network runtimes, sessions, events, status |
-| `src/dataplane/`    | the plugin contract, the packet transport, and the WireGuard plugin |
-| `src/bin/tsunagi.rs`| the command line agent; the only place that owns a runtime, a logger and signals |
+| `crates/tsunagi-core/src/identity/`     | device identity; deterministic network space identity and derived keys |
+| `crates/tsunagi-core/src/storage/`      | `state.sqlite`, `cache.sqlite`, directory ownership lock |
+| `crates/tsunagi-core/src/discovery.rs`  | candidate sources; test and static backends |
+| `crates/tsunagi-core/src/proto/`        | framing, message formats, membership handshake |
+| `crates/tsunagi-core/src/net.rs`        | iroh endpoint adapter and observability snapshots |
+| `crates/tsunagi-core/src/agent/`        | agent lifecycle, per-network runtimes, sessions, events, status |
+| `crates/tsunagi-core/src/state/`        | signed records that outlive a session, their merge rules and address allocation |
+| `crates/tsunagi-core/src/overlay/`   | the one interface an agent owns: provisioning, the TUN, routing, source checks, broadcast host rules |
+| `crates/tsunagi-core/src/dataplane/` | the protocol contract and the authenticated packet transport |
+| `crates/tsunagi-core/src/dns/`       | the DNS view of a network: zone, server, resolver publication |
+| `crates/tsunagi-wg-quic/`       | the `wg-quic` protocol: its keys, its announcement, its tunnels |
+| `crates/tsunagi/` | the command line agent, built as `tsng`; the only place that owns a runtime, a logger and signals |
 | `tests/`            | integration tests; `tests/common/` is the shared harness |
 
 Add abstractions only at real substitution or testing boundaries. Do not add a
-trait per struct. Prefer one crate with clear modules over many small crates.
+trait per struct.
+
+**The system level and its plugins are separate crates**, so that boundary is
+checked by the compiler and not by discipline: a plugin can reach only what
+`tsunagi` makes public, and carries its own version. Everything else stays
+one crate with clear modules — do not split further without a reason of that
+kind.
 
 ## Testing rules
 
@@ -124,7 +195,26 @@ trait per struct. Prefer one crate with clear modules over many small crates.
 - Running several library instances in one process is not a test of several
   system processes; do not describe it as one.
 
-## Before you open a change
+## Commit every completed change
+
+Agents must create a Git commit before finishing any task that changes this
+repository. Do not leave the user to make the commit. Inspect recent commit
+messages and follow their style: a short English subject stating the change,
+then a useful body explaining the problem, final behavior, architectural
+decisions, compatibility/migration effects, and checks actually performed.
+Write enough context for a future maintainer or LLM to understand why the code
+exists. Record failed or skipped checks honestly; never claim unrun tests.
+
+Before committing, inspect the diff and stage only this task's changes. Preserve
+unrelated user changes and never commit secrets or generated build artifacts.
+If Git has no effective author name or email, set only the missing repository-
+local values to `AB` and `ab@hexor.cy`. Do not override an existing identity or
+change global Git configuration. Report the resulting commit hash.
+
+Version bumps, release tags, and pushing commits/tags belong to the user unless
+they explicitly ask the agent to do them. Read-only tasks need no empty commit.
+
+## Required validation
 
 ```bash
 cargo fmt --all -- --check

Cargo.toml

@@ -1,64 +1,42 @@
-[package]
-name = "tsunagi"
-version = "0.1.0"
+# The system level and its plugins are separate crates, so the boundary
+# between them is checked by the compiler rather than by discipline: a plugin
+# can reach only what the core makes public, and carries its own version.
+[workspace]
+resolver = "3"
+members = ["crates/*"]
+exclude = ["vendor/mainline", "vendor/iroh"]
+
+# Mainline 8.0.0 treats Windows UDP read timeouts as socket failures.
+# Keep the small receive-loop correction reproducible until an upstream fix.
+[patch.crates-io]
+mainline = { path = "vendor/mainline" }
+# Backport bounded path retries until an iroh release includes the OOM fix.
+iroh = { path = "vendor/iroh" }
+
+[workspace.package]
 edition = "2024"
 rust-version = "1.91"
-license = "MIT OR Apache-2.0"
-description = "Proof-of-concept library for small private mesh networks: persistent agent identity, deterministic network spaces, iroh-based control plane."
-repository = "https://github.com/tsunagi-net/tsunagi"
-readme = "README.md"
-keywords = ["mesh", "p2p", "iroh", "networking"]
-categories = ["network-programming"]
+license = "WTFPL"
+repository = "https://github.com/house-of-vanity/tsunagi"
 
-[features]
-default = ["cli"]
-# The `tsunagi` command line binary. Library users can opt out.
-cli = ["dep:clap", "dep:tracing-subscriber", "tokio/signal", "tun-device"]
-# A real TUN device, so the WireGuard plugin can carry actual IP traffic.
-# Needs CAP_NET_ADMIN at run time; without it the plugin still runs and its
-# in-memory device can be used for tests.
-tun-device = ["dep:tun"]
-
-[[bin]]
-name = "tsunagi"
-path = "src/bin/tsunagi.rs"
-required-features = ["cli"]
-
-[dependencies]
-clap = { version = "4.5", features = ["derive", "env"], optional = true }
-tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
-iroh = { version = "1.2", default-features = false, features = ["tls-ring"] }
-tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros"] }
-rusqlite = { version = "0.40", features = ["bundled"] }
-hkdf = "0.13"
-hmac = "0.13"
-sha2 = "0.11"
-subtle = "2.6"
-zeroize = { version = "1.9", features = ["derive"] }
-rand = "0.10"
-serde = { version = "1.0", features = ["derive"] }
-postcard = { version = "1.1", default-features = false, features = ["use-std"] }
-data-encoding = "2.11"
-hex = "0.4"
-thiserror = "2.0"
-tracing = "0.1"
-fs4 = { version = "1.1", features = ["sync"] }
-directories = "6.0"
-netwatch = "0.19.3"
-bytes = "1.12.1"
-boringtun = { version = "0.7.1", default-features = false }
-tun = { version = "0.8", features = ["async"], optional = true }
-
-[dev-dependencies]
-tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros", "process"] }
-tempfile = "3.24"
-tracing-subscriber = { version = "0.3", features = ["env-filter"] }
-
-[lints.rust]
+[workspace.lints.rust]
 missing_docs = "warn"
-unsafe_code = "forbid"
+unsafe_code = "deny"
 
-[lints.clippy]
+[workspace.lints.clippy]
 unwrap_used = "warn"
 expect_used = "warn"
 panic = "warn"
+
+[workspace.dependencies]
+iroh = { version = "1.2", default-features = false, features = ["tls-ring", "portmapper"] }
+tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros"] }
+serde = { version = "1.0", features = ["derive"] }
+postcard = { version = "1.1", default-features = false, features = ["use-std"] }
+bytes = "1.12.1"
+thiserror = "2.0"
+tracing = "0.1"
+hex = "0.4"
+data-encoding = "2.11"
+tempfile = "3.24"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }

LICENSE

@@ -0,0 +1,13 @@
+            DO WHAT THE FUCK YOU WANT TO PUBLIC LICENSE
+                    Version 2, December 2004
+
+ Copyright (C) 2004 Sam Hocevar <sam@hocevar.net>
+
+ Everyone is permitted to copy and distribute verbatim or modified
+ copies of this license document, and changing it is allowed as long
+ as the name is changed.
+
+            DO WHAT THE FUCK YOU WANT TO PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. You just DO WHAT THE FUCK YOU WANT TO.

README.md

@@ -1,256 +1,254 @@
-# tsunagi
+<p align="center">
+  <img src="dist/linux/tsunagi.svg" width="96" alt="">
+</p>
 
-A proof-of-concept Rust library for **small private mesh networks** — a handful
-of friends, home machines, a few servers. Units to dozens of participants, not
-thousands.
+<h1 align="center">tsunagi</h1>
 
-The end user configures exactly two things:
+<p align="center">
+  <b>A mesh VPN with no server.</b><br>
+  Pick a network name, share a secret, and your machines are on one private network.
+</p>
 
-```text
-network_name
-secret          # one shared secret; "password" and "secret" mean the same value
+<p align="center">
+  <a href="https://github.com/house-of-vanity/tsunagi/releases/latest"><img alt="release" src="https://img.shields.io/github/v/release/house-of-vanity/tsunagi?include_prereleases"></a>
+  <a href="https://github.com/house-of-vanity/tsunagi/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/house-of-vanity/tsunagi/actions/workflows/ci.yml/badge.svg"></a>
+  <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-WTFPL-blue"></a>
+</p>
+
+```console
+laptop$ tsng join -n home
+joined `home`
+  secret  tsn1u7c…
+
+server$ tsng join -n home -s tsn1u7c…
+server$ tsng id hostname server
+
+laptop$ ssh server.home
 ```
 
-From those, every agent independently derives the same network space. There is
-no central server, no network owner with special powers, no registration and no
-majority vote. Anyone who knows the parameters can join; nobody has to trust
-anybody else.
+No account, no coordination server, no admin console, no keys to approve.
+Anyone who knows the **name** and the **secret** is in. Nobody else can find
+the network, let alone enter it.
 
-## What this proof of concept actually does
+## Why
 
-A working library with **real iroh connections** and integration tests:
+Tailscale, Headscale and ZeroTier are great, but each needs a control server:
+theirs, or one you have to run and keep alive. tsunagi doesn't. Machines find
+each other through the public BitTorrent DHT, connect directly (hole punching,
+with relay fallback), and keep the network state among themselves.
 
-- persistent device identity stored in SQLite, stable across restarts;
-- several independent networks at once in one agent;
-- deterministic network identity derived from name + secret;
-- candidates supplied by a replaceable discovery component;
-- real iroh connections plus an explicit mutual proof of network membership;
-- a small versioned control protocol: handshake, hostname/capability
-  announcement, ping/pong;
-- automatic reconnect with bounded exponential backoff and jitter;
-- status snapshots, an event stream and honest diagnostics;
-- configuration restored after a restart;
-- correct behaviour when the disposable cache is missing or corrupt;
-- a **WireGuard data plane**, in userspace: its own key per network,
-  deterministic IPv6 overlay addressing, real tunnels carried over iroh, and
-  address ownership enforced rather than believed;
-- a **command line agent**, `tsunagi`.
+|                        | Tailscale          | Headscale        | ZeroTier            | **tsunagi**         |
+| ---------------------- | ------------------ | ---------------- | ------------------- | ------------------- |
+| Control server         | hosted by vendor   | you run it       | hosted or self-run  | **none**            |
+| Account / sign-in      | yes                | yes              | yes                 | **no**              |
+| To join a network      | sign in, get approved | pre-auth key  | network ID, approval | **name + secret**  |
+| Direct connections     | yes                | yes              | yes                 | yes                 |
+| Exit nodes             | yes                | yes              | via routes          | yes                 |
 
-### What it deliberately does **not** do
+## Features
 
-Not implemented, and not pretended to be: Mainline DHT, DNS, routing through
-intermediate participants, a full CRDT, dynamically loaded plugins, a system
-service, a complete CLI, or a local control socket. Snapshot synchronisation
-and signed revocations are designed for but not implemented — see
-[docs/sync-model.md](docs/sync-model.md). The WireGuard plugin's own limits,
-including that its system backend is Linux-only, are in
-[docs/wireguard.md](docs/wireguard.md#limits-and-future-work).
+- **Two things to know:** a network name and a secret. A secret is generated for you when you make a network.
+- **Direct, encrypted links** between members (WireGuard cryptography), with relay fallback behind strict NATs.
+- **Names for machines:** `server.home` just works, even while `server` is switched off. A name belongs to whoever claimed it first; a later device with the same name keeps working by address and is told so.
+- **Exit nodes:** send all your internet traffic through any member that offers one.
+- **Multi-hop:** members that can't reach each other directly are routed through a member that can, over whatever kind of link each hop has. A member reached that way is online, and shown as relayed; traffic moves to a direct link by itself the moment one comes up.
+- **LAN games:** UDP broadcasts are relayed, so "LAN" lobbies show up across the mesh.
+- **Several networks at once** on one device.
+- **Tray app** for Windows, macOS and Linux; a command line for everything.
 
-**Control and data are separated logically, not physically.** Both ride on
-iroh, on different ALPNs and different connections, so the data plane inherits
-iroh's hole punching and relay fallback instead of reimplementing them — while
-the control protocol still knows nothing about packets and can keep a different
-transport underneath it later. Filtering user traffic remains the operating
-system's and the user's responsibility, not this library's.
+## Install
 
-## Requirements
+Grab a build from the [latest release](https://github.com/house-of-vanity/tsunagi/releases/latest).
 
-- Rust 1.91 or newer (iroh 1.2 requires it), edition 2024. Pinned dependencies
-  in `Cargo.lock`.
-- No internet, no DHT, no public relay, no administrator rights and no changes
-  to OS network settings are needed to build or test.
-- WireGuard runs in userspace (boringtun): **no kernel module and no `wg`
-  tool**. Only creating a real network interface needs `CAP_NET_ADMIN`, and
-  `--no-tun` skips even that.
+| Platform              | How                                                                              |
+| --------------------- | -------------------------------------------------------------------------------- |
+| **Windows**           | Run `tsunagi-setup-<version>-x86_64.exe`. Installs the service, the tray and `tsng`. |
+| **macOS**             | `brew tap house-of-vanity/tap && brew install --cask tsunagi-gui`<br>CLI only: `brew install tsunagi` |
+| **Debian / Ubuntu**   | `sudo apt install ./tsunagi-gui_<version>_amd64.deb` (`tsunagi_…` for no tray)   |
+| **Arch**              | `sudo pacman -U tsunagi-gui-<version>-1-x86_64.pkg.tar.zst`                      |
+| **NixOS**             | [Flake and module](#nixos) below                                                 |
+| **Any other Linux**   | Unpack the tarball, then `sudo ./install.sh && sudo systemctl enable --now tsunagi` |
 
-## Trying it on two machines
+Linux packages run the agent as an unprivileged service user. Add yourself to
+the `tsunagi` group (`sudo usermod -aG tsunagi $USER`, then log in again) to
+use `tsng` and the tray without `sudo`.
 
-On the first machine:
+## Quick start
 
-```bash
-cargo build --release
-./target/release/tsunagi secret          # prints tsn1...; share it privately
-./target/release/tsunagi doctor          # what this host can and cannot do
-
-./target/release/tsunagi up --network lab --secret "$SECRET" --wireguard
+```sh
+tsng join -n home                 # first machine: makes the network and prints a secret
+tsng join -n home -s tsn1…        # every other machine
+tsng status                       # who is online, over which path
 ```
 
-It prints its endpoint id and then waits. On the second machine, pass that id:
+Every new network gets a stable pseudorandom `/24` in `10.0.0.0/8`, derived
+from its network id. Its members agree on that range even when starting offline,
+and each gets a name like `<host>.home`. Existing signed ranges are preserved.
 
-```bash
-./target/release/tsunagi up --network lab --secret "$SECRET" --wireguard \
-  --peer <endpoint-id-from-the-first-machine>
+If two configured networks have overlapping ranges, the agent runs only one.
+`tsng status`, `tsng network` and the tray show which network blocks the
+other. `tsng network start <name-or-id>` or the tray’s “Switch to this network”
+button suspends the conflicting networks. They resume automatically when the
+selected network stops; the selection survives restarting the agent. Explicitly
+stopping a suspended network cancels its automatic resume.
+
+Machines are called `tsunagi-<id>` until you rename them: `tsng id hostname laptop`.
+
+### Virtual services
+
+```sh
+tsng vip create home api
+tsng vip port add home api tcp/443 --backend server1:6443 --backend server2:7443
+tsng vip status home api --watch
 ```
 
-Within a few seconds both print something like:
+Use `api.home:443`. Each client selects the available backend with the lowest
+measured data-plane RTT. A VIP can have multiple independent TCP and UDP ports.
+Every member may edit it; the tray's **VIP** button opens a draft editor with
+explicit Save. [Commands, health, metrics and limits](docs/vip.md).
 
-```text
-  + peer b47c958462 connected over Direct rtt=Some(4.5ms)
-  + data link to b47c958462 for wireguard: Direct via Ip(…), datagram 1382
+### Exit nodes
 
---- status ---
-control: 1 peer(s), 0 dial failure(s), 0 handshake failure(s)
-wireguard: tsunkkcp43lmdje on fd15:1d9e:fa21:f201:…/64 mtu 1100, 1/1 tunnel(s) established
-  4jO4kx9Z fd15:1d9e:fa21:f201:… handshake 3s ago tx=0 rx=0 dropped=0 path=Direct via Ip(…)
+```sh
+tsng network exit-node home on    # on the machine that should be the exit
+tsng exit-node                    # on another one: who offers an exit
+tsng exit-node laptop             # send everything through it
+tsng exit-node off
 ```
 
-`1/1 tunnel(s) established` means a real WireGuard handshake completed. Then
-`ping6` the peer's overlay address.
+The tray does the same with a checkbox. An exit node sees everything you send
+through it, so use only members you trust.
+[Details and per-OS notes](docs/exit-node.md).
 
-Notes:
+The default data transport order is `wg,awg,wg-quic,tcp-tls`. AWG negotiates a
+fresh masking profile and outer keys for each pair connection attempt, including
+reentry after fallback. Sustained data loss triggers switching independently of
+the iroh control connection. [Configuration and limits](docs/amnezia.md).
 
-- Only one side needs `--peer`; the link is bidirectional.
-- The default `--transport n0` uses iroh's public address lookup and relays, so
-  two machines behind NAT find each other. `--transport local` keeps everything
-  on the local network. See *How peers find each other* below — it is worth
-  understanding what gets published.
-- Without `CAP_NET_ADMIN`, add `--no-tun`: the mesh, the data links and the
-  WireGuard handshakes all still run and are visible in the status output, only
-  traffic does not reach the operating system. That is the quickest way to
-  confirm the network forms.
-- Run as root (or grant `CAP_NET_ADMIN`) to get a real interface.
+Change a network's protocol selection without editing its service or restarting:
 
-## Checks
-
-```bash
-cargo fmt --all -- --check
-cargo clippy --locked --workspace --all-targets -- -D warnings
-cargo test --locked --workspace --all-targets
+```sh
+tsng protocols                         # available protocols and settings
+tsng protocols home                    # this device's choice for home
+tsng protocols home awg,wg-quic         # prefer AWG, retain QUIC as fallback
+tsng protocols home none                # control plane only
 ```
 
-The whole suite runs offline on loopback. Set `TSUNAGI_TEST_LOG=tsunagi=debug`
-to see agent logs while a test runs.
+Exact network names, full ids and unique id prefixes work here. The choice is
+local to this device, saved per network and applied immediately; an offline
+change applies on the next start. In the tray, expand **protocols** on a network
+tile to toggle protocols or change their preference order. Both endpoints need
+a common protocol at the same wire version. `tsng up --protocol` supplies the
+initial defaults for networks without a saved choice; it does not restrict
+which built-in protocols can be enabled later.
 
-There are also two runnable demos, which are demos and not substitutes for the
-tests:
+Remove an obsolete offline device from a network's shared history:
 
-```bash
-cargo run --example two_agents       # control plane only
-cargo run --example wireguard_mesh   # a WireGuard overlay carrying a real packet
+```sh
+tsng network forget-node home old-laptop
+tsng network forget-node home 10.13.37.143
 ```
 
-Both run with no privileges and change nothing on the host.
+Networks accept exact names, full ids or unique id prefixes; nodes accept those
+id selectors, exact hostnames or overlay IPv4 addresses. Ambiguous selectors
+change nothing. Start the network first if it is stopped. In **Show devices**,
+click **Forget** beside an offline node and confirm its identity.
 
-## Usage
+Every member has the same cleanup right. The signed cleanup frees the observed
+claim's name and IPv4 reservation, survives restarts and reaches offline replicas
+when they reconnect. It does not ban the device: a newer signed claim can return.
+Online nodes and the currently selected exit node cannot be forgotten.
+[State and limits](docs/sync-model.md#offline-node-cleanup).
 
-```rust,no_run
-use std::sync::Arc;
-use tsunagi::config::{AgentConfig, StoragePaths, TransportPolicy};
-use tsunagi::discovery::SharedMemoryDiscovery;
-use tsunagi::identity::{NetworkName, NetworkSecret};
-use tsunagi::proto::ControlMessage;
-use tsunagi::{Agent, Result};
+## Platforms
 
-// The library never starts a runtime, installs a logger, handles signals,
-// forks, or calls process::exit. The binary owns all of that.
-#[tokio::main]
-async fn main() -> Result<()> {
-    let config = AgentConfig::new(StoragePaths::user_default()?)
-        .with_transport(TransportPolicy::N0Defaults)
-        .with_discovery(Arc::new(SharedMemoryDiscovery::new()));
+| Platform                  | Build        | CLI `tsng` | Tray | Exit node: offer / use | Packages                            |
+| ------------------------- | ------------ | :--------: | :--: | :--------------------: | ----------------------------------- |
+| Linux x86_64              | release      |     ✔      |  ✔   |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball, Nix |
+| Linux aarch64             | release      |     ✔      |  –¹  |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball, Nix |
+| Linux armv7               | release      |     ✔      |  –¹  |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball      |
+| macOS, Apple silicon      | release      |     ✔      |  ✔   |        ✔ / ✔²          | Homebrew, tarball                    |
+| macOS, Intel              | from source  |     ✔³     |  ✔³  |        ✔ / ✔²³         | –                                    |
+| Windows x86_64            | release      |     ✔      |  ✔   |        ✔ / ✔²          | installer, zip                       |
 
-    let agent = Agent::spawn(config).await?;
+<sup>¹ Not in the release; builds from source (and through Nix).
+² Newer and less battle-tested than on Linux.
+³ Builds from source, not covered by CI.</sup>
 
-    let name = NetworkName::new("kitchen-table")?;
-    let secret = NetworkSecret::generate();       // 32 random bytes
-    println!("share this: {}", secret.encode().as_str());
+The Linux CLI is a single static binary: it runs on any distribution.
 
-    let network = agent.join_network(&name, &secret).await?;
+## NixOS
 
-    let mut events = agent.subscribe();
-    tokio::spawn(async move {
-        while let Ok(event) = events.recv().await {
-            println!("{event:?}");
+The repository is a flake with a package and a module.
+
+```nix
+# flake.nix
+{
+  inputs.tsunagi.url = "github:house-of-vanity/tsunagi";
+
+  outputs = { nixpkgs, tsunagi, ... }: {
+    nixosConfigurations.myhost = nixpkgs.lib.nixosSystem {
+      modules = [
+        tsunagi.nixosModules.default
+        {
+          services.tsunagi = {
+            enable = true;
+            tray.enable = true;      # optional, for desktops
+            users = [ "alice" ];     # may use tsng and the tray without sudo
+          };
         }
-    });
-
-    for peer in agent.network_status(network).await?.connected_peers() {
-        agent
-            .send(network, peer, ControlMessage::Ping { seq: 1, payload: vec![] })
-            .await?;
-    }
-
-    agent.shutdown().await;
-    Ok(())
+      ];
+    };
+  };
 }
 ```
 
-`TransportPolicy::LocalOnly` is the default, so a plain `AgentConfig::new` never
-reaches the internet by accident. Opt into `DirectOnly` or `N0Defaults`
-explicitly.
+Just trying it: `nix run github:house-of-vanity/tsunagi#tsunagi-cli -- --help`.
+`packages.<system>.tsunagi` (with the tray) and `tsunagi-cli` are also exported,
+along with `overlays.default`.
 
-## How peers find each other
+For development from a checkout, `nix develop` also exposes the tray's native
+runtime libraries, including AppIndicator and graphics/windowing libraries.
+Cargo-built binaries such as `target/debug/tsunagi-tray` can run directly inside
+this shell, without installing the packaged wrapper.
 
-Two different lookups are involved, and only one of them is this project's:
+## Build from source
 
-**1. Resolving one endpoint's address — iroh's, and it works today.**
-With `--transport n0` or `--transport direct`, iroh publishes a signed record
-of this endpoint's addresses, keyed by its endpoint id, to Number 0's public
-service (`dns.iroh.link`, over pkarr and DNS) and resolves other endpoints the
-same way. That is why `--peer <endpoint-id>` works with no address attached:
-iroh looks it up. None of that code is ours.
+Rust 1.91 or newer.
 
-**2. Finding who is in a network — ours, and it is still manual.**
-`NetworkDiscovery` maps a secret-derived `DiscoveryKey` to a set of *candidate*
-members. Two backends exist: `StaticBootstrap` (what `--peer` feeds) and an
-in-memory one for tests. The planned Mainline DHT backend, which would let
-members find each other from the network secret alone, is **not implemented**.
-So today you bootstrap by passing one peer's id; after that the mesh is
-whatever those agents reach.
+```sh
+cargo build --release --bin tsng               # the agent and CLI
+cargo build --release --bin tsunagi-tray       # the tray (Linux needs GTK 3 and libayatana-appindicator)
+```
 
-What this means in practice:
+Windows needs [Wintun](https://www.wintun.net) next to `tsng.exe`.
+Per-platform notes: [Linux](dist/linux/README.md), [Arch](dist/arch/README.md),
+[macOS](dist/macos/README.md), [Windows](dist/windows/README.md).
 
-- With `n0` or `direct`, **your endpoint id and IP addresses are published to a
-  public third-party service.** They are not secret, and the network secret is
-  never published, but an observer of that service learns that your endpoint
-  exists and where it is. `--transport local` publishes nothing.
-- A relay, when one is needed, sees the volume and timing of your traffic — not
-  its contents.
+## How it works
 
-## Storage
+- Your name and secret derive a network identity. Members publish signed
+  contact records to the public Mainline DHT under it, and find each other there.
+- Every connection proves knowledge of the secret before it's accepted.
+  Traffic between members is end-to-end encrypted, relays included.
+- Addresses and names are claimed with each device's own key and replicated
+  between members. There is no one in charge, so nothing to be taken down.
 
-Two physically separate SQLite files, placed wherever the library's
-configuration says (`StoragePaths`). A future system service supplies its own
-paths; tests always use temporary directories.
-
-| file           | holds                                            | when damaged            |
-|----------------|--------------------------------------------------|-------------------------|
-| `state.sqlite` | device identity, network configuration, hostname | clear error, never reset |
-| `cache.sqlite` | address hints and other recoverable data          | discarded and recreated  |
-
-The WireGuard plugin keeps its own keys in its own `wireguard.sqlite`, wherever
-its configuration points, because plugin keys are neither the iroh identity nor
-the network secret.
-
-The command line agent puts everything under the platform's per-user
-directories by default; `--state-dir` and `--cache-dir` override them.
-
-One state directory belongs to one live agent, enforced with a real OS file
-lock rather than an existence check.
+The secret is the only key, so make it a good one (the generated ones are).
+The [threat model](docs/threat-model.md) says what this does and doesn't protect against.
 
 ## Documentation
 
-- [docs/architecture.md](docs/architecture.md) — module boundaries and runtime.
-- [docs/wireguard.md](docs/wireguard.md) — the WireGuard plugin: overlay
-  addressing, announcements, backends, reconciliation.
-- [docs/protocol.md](docs/protocol.md) — identity derivation, framing, handshake.
-- [docs/sync-model.md](docs/sync-model.md) — the planned signed-state model and
-  what is deliberately not built yet.
-- [docs/threat-model.md](docs/threat-model.md) — threat model and known limits.
-- [docs/testing.md](docs/testing.md) — what the suite covers and what it does not.
-- [AGENTS.md](AGENTS.md) — rules for anyone (human or agent) changing this repo.
-
-## Security in one paragraph
-
-Membership is proved by an HMAC over a transcript keyed by a value derived from
-the shared secret, bound to the specific iroh connection through the TLS
-exporter, to the network id, to both endpoint identities and to distinct role
-labels. This targets high-entropy secrets: there is no PAKE here, so a short
-human passphrase is guessable offline by anyone who can reach the handshake.
-Anyone who knows the secret is a full participant and can create many
-identities. Read [docs/threat-model.md](docs/threat-model.md) before relying on
-any of this.
+[Architecture](docs/architecture.md) ·
+[Protocol](docs/protocol.md) ·
+[WireGuard data plane](docs/wireguard.md) ·
+[Pairwise Amnezia transport](docs/amnezia.md) ·
+[Routing](docs/routing.md) ·
+[Exit nodes](docs/exit-node.md) ·
+[Virtual services](docs/vip.md) ·
+[Threat model](docs/threat-model.md) ·
+[Contributing](AGENTS.md)
 
 ## Licence
 
-MIT OR Apache-2.0.
+[WTFPL](LICENSE): do what the fuck you want to.

crates/tsunagi-awg/Cargo.toml

@@ -0,0 +1,30 @@
+[package]
+name = "tsunagi-awg"
+version = "0.1.0"
+description = "Ephemeral pairwise Amnezia transport for Tsunagi."
+edition.workspace = true
+rust-version.workspace = true
+license = "WTFPL AND Apache-2.0"
+repository.workspace = true
+
+[dependencies]
+tsunagi-core = { path = "../tsunagi-core" }
+tsunagi-wg-quic = { path = "../tsunagi-wg-quic" }
+tsunagi-wireguard = { path = "../tsunagi-wireguard" }
+iroh.workspace = true
+tokio = { workspace = true, features = ["net", "time", "sync", "macros"] }
+serde.workspace = true
+postcard.workspace = true
+bytes.workspace = true
+tracing.workspace = true
+rand = "0.10"
+zeroize = { version = "1.9", features = ["derive"] }
+boringtun = { version = "0.7.1", default-features = false }
+chacha20 = "0.9"
+
+[dev-dependencies]
+tsunagi-core = { path = "../tsunagi-core", features = ["testing"] }
+tempfile.workspace = true
+
+[lints]
+workspace = true

crates/tsunagi-awg/LICENSE-AmneziaWG-RS

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Friderika Coulson (NanoBee)
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

crates/tsunagi-awg/UPSTREAM.md

@@ -0,0 +1,21 @@
+# Amnezia packet engine
+
+`src/masking.rs` is adapted from `src/core/amnezia.rs` in
+https://github.com/RunaRoo/AmneziaWG-RS at
+`51579a7d6e8882d1d8d6d00822f09e18a9423b73`, under Apache-2.0.
+The upstream license is retained in `LICENSE-AmneziaWG-RS`.
+
+The adaptation replaces configuration-file loading with bounded typed pair
+profiles, updates RNG calls to rand 0.10, removes an unnecessary unwrap, and
+does not embed upstream's Node, global executors, OS interface operations or
+detached tasks. Noise/WireGuard is supplied by boringtun, already used by
+Tsunagi. Each pair owns one socket and one cancellable packet-engine task.
+
+This is Tsunagi AWG wire version 1, not a claim of compatibility with external
+Amnezia servers. The masking engine substitutes message headers outside the
+standard WireGuard state machine and restores them before Noise validation.
+There are no remote file paths, shell hooks, or persisted pair profiles.
+
+The adapted masking source remains Apache-2.0. Newly written Tsunagi code
+follows the repository's WTFPL. This package declares both licenses in Cargo
+metadata; the workspace license alone does not describe its copied source.

crates/tsunagi-awg/src/lib.rs

@@ -0,0 +1,16 @@
+//! Pairwise, ephemeral Amnezia packet transport with independent end-to-end tunnels.
+//! All negotiation travels over authenticated Tsunagi control sessions.
+mod masking;
+mod plugin;
+pub mod profile;
+mod session;
+mod stun;
+mod transport;
+
+pub use plugin::AwgPlugin;
+pub use transport::{AwgTransport, PairOverview};
+
+/// Protocol identifier and default second-choice direct transport.
+pub const AWG_PROTOCOL: &str = "awg";
+/// Tsunagi AWG wire version, independent of software releases.
+pub const AWG_WIRE_VERSION: u16 = 1;

crates/tsunagi-awg/src/masking.rs

@@ -0,0 +1,336 @@
+// Adapted from RunaRoo/AmneziaWG-RS, commit 51579a7d6e8882d1d8d6d00822f09e18a9423b73.
+// Apache-2.0; see LICENSE-AmneziaWG-RS and UPSTREAM.md.
+#![allow(dead_code, missing_docs)]
+// AmneziaWG configuration, header protection, and traffic shaping.
+
+use bytes::{Bytes, BytesMut};
+use chacha20::{
+    ChaCha20,
+    cipher::{KeyIvInit, StreamCipher},
+};
+use rand::RngExt;
+
+pub const HEADER_PROTECTION_NONCE_SIZE: usize = 12;
+pub const DEFAULT_UDP_WINDOW: usize = 500;
+
+#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
+pub struct Range {
+    pub min: u32,
+    pub max: u32,
+}
+
+impl Range {
+    pub fn contains(&self, val: u32) -> bool {
+        val >= self.min && val <= self.max
+    }
+
+    pub fn random(&self) -> u32 {
+        if self.min >= self.max {
+            self.min
+        } else {
+            rand::rng().random_range(self.min..=self.max)
+        }
+    }
+
+    pub fn is_zero(&self) -> bool {
+        self.min == 0 && self.max == 0
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum AwgKind {
+    Initiation,
+    Response,
+    Cookie,
+    Transport,
+}
+
+impl AwgKind {
+    pub fn fixed_size(self) -> usize {
+        match self {
+            Self::Initiation => 148,
+            Self::Response => 92,
+            Self::Cookie => 64,
+            Self::Transport => 32,
+        }
+    }
+
+    fn standard_type(self) -> u32 {
+        match self {
+            Self::Initiation => 1,
+            Self::Response => 2,
+            Self::Cookie => 3,
+            Self::Transport => 4,
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct DecodedPacket {
+    pub kind: AwgKind,
+    pub message: Bytes,
+    pub wire_len: usize,
+}
+
+pub struct AmneziaConfig {
+    pub jc: u32,
+    pub jmin: usize,
+    pub jmax: usize,
+    pub s1: usize,
+    pub s2: usize,
+    pub s3: usize,
+    pub s4: usize,
+    pub h1: Range,
+    pub h2: Range,
+    pub h3: Range,
+    pub h4: Range,
+    pub i1: String,
+    pub i2: String,
+    pub i3: String,
+    pub i4: String,
+    pub i5: String,
+    pub header_protection_key: Option<[u8; 32]>,
+    pub content_padding_addition: Range,
+    pub random_trailers: bool,
+    pub mtu: usize,
+}
+
+impl AmneziaConfig {
+    pub fn random_header(&self, range: &Range) -> u32 {
+        range.random()
+    }
+
+    pub fn random_padding(&self, len: usize) -> Vec<u8> {
+        let mut padding = vec![0u8; len];
+        if len > 0 {
+            rand::fill(&mut padding);
+        }
+        padding
+    }
+
+    fn shape(&self, kind: AwgKind) -> (usize, Range) {
+        match kind {
+            AwgKind::Initiation => (self.s1, self.h1),
+            AwgKind::Response => (self.s2, self.h2),
+            AwgKind::Cookie => (self.s3, self.h3),
+            AwgKind::Transport => (self.s4, self.h4),
+        }
+    }
+
+    pub fn encode(&self, kind: AwgKind, payload: Bytes, udp_window: usize) -> Bytes {
+        let (padding_len, _) = self.shape(kind);
+        let trailer_len = if kind != AwgKind::Transport && self.random_trailers {
+            let base = padding_len.saturating_add(payload.len());
+            let room = udp_window.max(DEFAULT_UDP_WINDOW).saturating_sub(base);
+            if room == 0 {
+                0
+            } else {
+                rand::rng().random_range(0..room)
+            }
+        } else {
+            0
+        };
+
+        let mut out = BytesMut::zeroed(padding_len + payload.len() + trailer_len);
+        if padding_len > 0 {
+            rand::fill(&mut out[..padding_len]);
+        }
+        out[padding_len..padding_len + payload.len()].copy_from_slice(&payload);
+        if trailer_len > 0 {
+            rand::fill(&mut out[padding_len + payload.len()..]);
+        }
+
+        if self.header_protection_key.is_some() {
+            let protected_len = if kind == AwgKind::Transport {
+                16
+            } else {
+                kind.fixed_size()
+            };
+            self.apply_header_cipher(&mut out, padding_len, protected_len);
+        }
+
+        out.freeze()
+    }
+
+    pub fn decode(&self, datagram: &Bytes) -> Option<DecodedPacket> {
+        if datagram.len() < 4 {
+            return None;
+        }
+
+        for kind in [
+            AwgKind::Initiation,
+            AwgKind::Response,
+            AwgKind::Cookie,
+            AwgKind::Transport,
+        ] {
+            let (padding_len, headers) = self.shape(kind);
+            let minimum = padding_len + kind.fixed_size();
+            let size_ok = match kind {
+                AwgKind::Transport | AwgKind::Initiation => datagram.len() >= minimum,
+                AwgKind::Response | AwgKind::Cookie if self.random_trailers => {
+                    datagram.len() >= minimum
+                }
+                AwgKind::Response | AwgKind::Cookie => datagram.len() == minimum,
+            };
+            if !size_ok || datagram.len() < padding_len + 4 {
+                continue;
+            }
+
+            let protected_header =
+                u32::from_le_bytes(datagram[padding_len..padding_len + 4].try_into().ok()?);
+            let candidate = if self.header_protection_key.is_some() {
+                protected_header ^ self.header_type_hash(datagram)?
+            } else {
+                protected_header
+            };
+            if !headers.contains(candidate) {
+                continue;
+            }
+
+            let message_len = match kind {
+                AwgKind::Transport | AwgKind::Initiation => datagram.len() - padding_len,
+                AwgKind::Response | AwgKind::Cookie => kind.fixed_size(),
+            };
+            let mut payload = BytesMut::from(&datagram[padding_len..padding_len + message_len]);
+            if self.header_protection_key.is_some() {
+                let protected_len = if kind == AwgKind::Transport {
+                    16
+                } else {
+                    kind.fixed_size()
+                };
+                self.apply_header_cipher_with_nonce(
+                    &mut payload,
+                    0,
+                    protected_len,
+                    &datagram[..HEADER_PROTECTION_NONCE_SIZE],
+                );
+            }
+
+            return Some(DecodedPacket {
+                kind,
+                message: payload.freeze(),
+                wire_len: datagram.len(),
+            });
+        }
+
+        None
+    }
+
+    pub fn prepare_transport_plaintext(&self, packet: Bytes, udp_window: usize) -> Bytes {
+        if packet.is_empty() {
+            return packet;
+        }
+
+        let requested = if !self.content_padding_addition.is_zero() {
+            self.content_padding_addition.random() as usize
+        } else if self.random_trailers {
+            let wire_without_padding = self.s4 + AwgKind::Transport.fixed_size() + packet.len();
+            let room = udp_window
+                .max(DEFAULT_UDP_WINDOW)
+                .saturating_sub(wire_without_padding);
+            if room == 0 {
+                0
+            } else {
+                rand::rng().random_range(0..room)
+            }
+        } else {
+            let aligned = (packet.len() + 15) & !15;
+            aligned.saturating_sub(packet.len())
+        };
+
+        let padding = self.clamp_padding_to_mtu(packet.len(), requested);
+        if padding == 0 {
+            return packet;
+        }
+
+        let mut out = BytesMut::with_capacity(packet.len() + padding);
+        out.extend_from_slice(&packet);
+        out.resize(packet.len() + padding, 0);
+        out.freeze()
+    }
+
+    pub fn transport_content_length(&self, plaintext: &[u8]) -> Option<usize> {
+        if plaintext.is_empty() || plaintext[0] == 0 {
+            return Some(0);
+        }
+        match plaintext[0] >> 4 {
+            4 if plaintext.len() >= 20 => {
+                let total = u16::from_be_bytes([plaintext[2], plaintext[3]]) as usize;
+                (total >= 20 && total <= plaintext.len()).then_some(total)
+            }
+            6 if plaintext.len() >= 40 => {
+                let payload = u16::from_be_bytes([plaintext[4], plaintext[5]]) as usize;
+                let total = 40 + payload;
+                (total <= plaintext.len()).then_some(total)
+            }
+            _ => None,
+        }
+    }
+
+    fn clamp_padding_to_mtu(&self, packet_len: usize, requested: usize) -> usize {
+        if self.mtu == 0 {
+            return requested;
+        }
+        let normalized = if packet_len > self.mtu {
+            packet_len % self.mtu
+        } else {
+            packet_len
+        };
+        requested.min(self.mtu.saturating_sub(normalized))
+    }
+
+    fn header_type_hash(&self, datagram: &Bytes) -> Option<u32> {
+        if datagram.len() < HEADER_PROTECTION_NONCE_SIZE {
+            return None;
+        }
+        let mut hash = [0u8; 4];
+        self.apply_keystream(&mut hash, &datagram[..HEADER_PROTECTION_NONCE_SIZE]);
+        Some(u32::from_le_bytes(hash))
+    }
+
+    fn apply_header_cipher(&self, packet: &mut [u8], offset: usize, len: usize) {
+        if packet.len() < HEADER_PROTECTION_NONCE_SIZE || packet.len() < offset + len {
+            return;
+        }
+        let nonce: [u8; HEADER_PROTECTION_NONCE_SIZE] =
+            match packet[..HEADER_PROTECTION_NONCE_SIZE].try_into() {
+                Ok(nonce) => nonce,
+                Err(_) => return,
+            };
+        self.apply_header_cipher_with_nonce(packet, offset, len, &nonce);
+    }
+
+    fn apply_header_cipher_with_nonce(
+        &self,
+        packet: &mut [u8],
+        offset: usize,
+        len: usize,
+        nonce: &[u8],
+    ) {
+        if packet.len() < offset + len {
+            return;
+        }
+        self.apply_keystream(&mut packet[offset..offset + len], nonce);
+    }
+
+    fn apply_keystream(&self, data: &mut [u8], nonce: &[u8]) {
+        let Some(key) = self.header_protection_key.as_ref() else {
+            return;
+        };
+        if nonce.len() != HEADER_PROTECTION_NONCE_SIZE {
+            return;
+        }
+        let Ok(nonce) = <&[u8; HEADER_PROTECTION_NONCE_SIZE]>::try_from(nonce) else {
+            return;
+        };
+        let mut cipher = ChaCha20::new(key.into(), nonce.into());
+        cipher.apply_keystream(data);
+    }
+}
+
+impl Drop for AmneziaConfig {
+    fn drop(&mut self) {
+        use zeroize::Zeroize;
+        self.header_protection_key.zeroize();
+    }
+}

crates/tsunagi-awg/src/plugin.rs

@@ -0,0 +1,132 @@
+//! End-to-end encryption remains independent of physical Amnezia pair links.
+use crate::{AWG_PROTOCOL, AWG_WIRE_VERSION, AwgTransport};
+use iroh::EndpointId;
+use std::sync::Arc;
+use tsunagi::dataplane::transport::SharedLink;
+use tsunagi::dataplane::{IpPlugin, PluginCapability, PluginContext, PluginError, ProtocolOption};
+use tsunagi::state::Ipv4Range;
+use tsunagi::{BoxFuture, NetworkId};
+use tsunagi_wg_quic::{NetworkOverview, WireguardConfig, WireguardPlugin};
+
+/// AWG capability and an independent end-to-end WireGuard engine.
+/// Pair transport keys/profiles never enter its durable end-to-end key store.
+#[derive(Debug)]
+pub struct AwgPlugin {
+    inner: Arc<WireguardPlugin>,
+    transport: Arc<AwgTransport>,
+}
+impl AwgPlugin {
+    /// Uses standard tunnel MTU/keepalive options; pair ports and profiles are ephemeral.
+    pub const OPTIONS: &'static [ProtocolOption] = &[
+        ProtocolOption {
+            key: "keepalive",
+            value: "SECONDS",
+            help: "end-to-end tunnel keepalive; 0 disables it",
+            default: Some("25"),
+        },
+        ProtocolOption {
+            key: "mtu",
+            value: "BYTES",
+            help: "maximum overlay packet size",
+            default: Some("1280"),
+        },
+        ProtocolOption {
+            key: "stun",
+            value: "HOST:PORT|off",
+            help: "UDP public mapping discovery; disabled in local reach mode",
+            default: Some("stun.l.google.com:19302"),
+        },
+    ];
+    /// Opens the end-to-end engine, retaining separate pair transport state.
+    pub async fn open(
+        mut config: WireguardConfig,
+        transport: Arc<AwgTransport>,
+    ) -> Result<Arc<Self>, PluginError> {
+        config.protocol = AWG_PROTOCOL;
+        Ok(Arc::new(Self {
+            inner: WireguardPlugin::open(config).await?,
+            transport,
+        }))
+    }
+    /// Reports end-to-end tunnel health and traffic counters.
+    pub fn overview(&self, network: NetworkId) -> Option<NetworkOverview> {
+        self.inner.overview(network)
+    }
+}
+impl IpPlugin for AwgPlugin {
+    fn protocol_id(&self) -> &str {
+        AWG_PROTOCOL
+    }
+    fn protocol_version(&self) -> u16 {
+        AWG_WIRE_VERSION
+    }
+    fn options(&self) -> &'static [ProtocolOption] {
+        Self::OPTIONS
+    }
+    fn attach(&self, context: PluginContext) {
+        self.inner.attach(context);
+    }
+    fn local_capability(
+        &self,
+        network: NetworkId,
+    ) -> Result<Option<PluginCapability>, PluginError> {
+        self.inner.local_capability(network).map(|cap| {
+            cap.map(|mut cap| {
+                cap.protocol = AWG_PROTOCOL.into();
+                cap.version = AWG_WIRE_VERSION;
+                cap
+            })
+        })
+    }
+    fn on_network_activated(&self, network: NetworkId) {
+        self.inner.on_network_activated(network);
+    }
+    fn on_peer_capability(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        capability: &PluginCapability,
+    ) -> Result<(), PluginError> {
+        if capability.protocol != AWG_PROTOCOL
+            || capability.version != AWG_WIRE_VERSION
+            || !capability.enabled
+        {
+            return Err(PluginError::Rejected("unsupported AWG capability".into()));
+        }
+        self.inner.on_peer_capability(network, peer, capability)?;
+        self.transport.allow_peer(network, peer);
+        Ok(())
+    }
+    fn on_address_allocation(
+        &self,
+        network: NetworkId,
+        range: Ipv4Range,
+        allocations: &[(EndpointId, std::net::Ipv4Addr)],
+    ) {
+        self.inner
+            .on_address_allocation(network, range, allocations);
+    }
+    fn carry(&self, network: NetworkId, peer: EndpointId, packet: bytes::Bytes) -> bool {
+        self.inner.carry(network, peer, packet)
+    }
+    fn on_peer_link(&self, network: NetworkId, peer: EndpointId, link: SharedLink) {
+        self.inner.on_peer_link(network, peer, link);
+    }
+    fn on_peer_gone(&self, network: NetworkId, peer: EndpointId) {
+        self.transport.forget_peer(network, peer);
+        self.inner.on_peer_gone(network, peer);
+    }
+    fn on_network_deactivated(&self, network: NetworkId) {
+        self.transport.stop_network(network);
+        self.inner.on_network_deactivated(network);
+    }
+    fn on_network_forgotten(&self, network: NetworkId) {
+        self.inner.on_network_forgotten(network);
+    }
+    fn shutdown(&self) -> BoxFuture<'_, ()> {
+        Box::pin(async move {
+            self.transport.shutdown();
+            self.inner.shutdown().await;
+        })
+    }
+}

crates/tsunagi-awg/src/profile.rs

@@ -0,0 +1,226 @@
+//! Bounded profiles, regenerated for every pair connection attempt.
+use crate::masking::{AmneziaConfig, AwgKind, Range};
+use bytes::Bytes;
+use serde::{Deserialize, Serialize};
+
+/// An ephemeral masking profile. Secret fields are redacted and erased on drop.
+#[derive(Clone, Serialize, Deserialize, zeroize::ZeroizeOnDrop)]
+pub struct PairProfile {
+    id: [u8; 16],
+    prefix: u16,
+    headers: [[u32; 2]; 4],
+    header_key: [u8; 32],
+    pub(crate) psk: [u8; 32],
+    junk_count: u8,
+    junk_min: u16,
+    junk_max: u16,
+    padding: u16,
+    signature: u8,
+}
+impl std::fmt::Debug for PairProfile {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("PairProfile([REDACTED])")
+    }
+}
+impl PairProfile {
+    /// Generates a new safe profile; no deterministic seed or durable state.
+    pub fn generate() -> Self {
+        let headers = std::array::from_fn(|i| {
+            let base = (i as u32) * 0x4000_0000 + rand::random_range(1024..0x3fff_0000);
+            [base, base + rand::random_range(64..1024)]
+        });
+        Self {
+            id: rand::random(),
+            prefix: rand::random_range(12..49),
+            headers,
+            header_key: rand::random(),
+            psk: rand::random(),
+            junk_count: rand::random_range(4..9),
+            junk_min: rand::random_range(24..65),
+            junk_max: rand::random_range(80..161),
+            padding: rand::random_range(16..65),
+            signature: rand::random_range(0..2),
+        }
+    }
+    /// Public attempt identifier, useful for distinguishing retries without revealing secrets.
+    pub fn id(&self) -> [u8; 16] {
+        self.id
+    }
+    /// Header key, for packet engines and protocol tests; never log this value.
+    pub fn header_key(&self) -> &[u8; 32] {
+        &self.header_key
+    }
+    /// Rejects invalid remote profiles before allocating packets or entering the codec.
+    pub fn validate(&self) -> Result<(), &'static str> {
+        if !(12..=48).contains(&self.prefix)
+            || !(4..=8).contains(&self.junk_count)
+            || self.junk_min < 24
+            || self.junk_min > self.junk_max
+            || self.junk_max > 160
+            || self.padding > 64
+            || self.signature > 1
+            || self.header_key == [0; 32]
+            || self.psk == [0; 32]
+        {
+            return Err("invalid AWG profile bounds");
+        }
+        for (i, [lo, hi]) in self.headers.iter().copied().enumerate() {
+            if lo < 5 || lo > hi || hi - lo > 1024 {
+                return Err("invalid AWG header range");
+            }
+            for [other_lo, other_hi] in &self.headers[..i] {
+                if lo <= *other_hi && *other_lo <= hi {
+                    return Err("overlapping AWG header ranges");
+                }
+            }
+        }
+        Ok(())
+    }
+    fn config(&self) -> AmneziaConfig {
+        let range = |i: usize| Range {
+            min: self.headers[i][0],
+            max: self.headers[i][1],
+        };
+        AmneziaConfig {
+            jc: self.junk_count.into(),
+            jmin: self.junk_min.into(),
+            jmax: self.junk_max.into(),
+            s1: self.prefix.into(),
+            s2: self.prefix.into(),
+            s3: self.prefix.into(),
+            s4: self.prefix.into(),
+            h1: range(0),
+            h2: range(1),
+            h3: range(2),
+            h4: range(3),
+            i1: String::new(),
+            i2: String::new(),
+            i3: String::new(),
+            i4: String::new(),
+            i5: String::new(),
+            header_protection_key: Some(self.header_key),
+            content_padding_addition: Range {
+                min: 0,
+                max: self.padding.into(),
+            },
+            random_trailers: true,
+            mtu: 1050,
+        }
+    }
+    /// Masks one standard WireGuard message; invalid types and oversized packets are refused.
+    pub fn encode(&self, raw: &[u8]) -> Option<Bytes> {
+        let kind = match raw.get(..4)? {
+            [1, 0, 0, 0] => AwgKind::Initiation,
+            [2, 0, 0, 0] => AwgKind::Response,
+            [3, 0, 0, 0] => AwgKind::Cookie,
+            [4, 0, 0, 0] => AwgKind::Transport,
+            _ => return None,
+        };
+        if raw.len() < kind.fixed_size() || raw.len() > 1050 {
+            return None;
+        }
+        let config = self.config();
+        let range = match kind {
+            AwgKind::Initiation => config.h1,
+            AwgKind::Response => config.h2,
+            AwgKind::Cookie => config.h3,
+            AwgKind::Transport => config.h4,
+        };
+        let mut payload = raw.to_vec();
+        payload[..4].copy_from_slice(&range.random().to_le_bytes());
+        Some(config.encode(kind, Bytes::from(payload), 1200))
+    }
+    /// Unmasks a packet and restores standard headers before WireGuard validation.
+    pub fn decode(&self, wire: &[u8]) -> Option<Vec<u8>> {
+        if wire.len() > 1200 {
+            return None;
+        }
+        let decoded = self.config().decode(&Bytes::copy_from_slice(wire))?;
+        let (kind, len) = match decoded.kind {
+            AwgKind::Initiation => (1, 148),
+            AwgKind::Response => (2, 92),
+            AwgKind::Cookie => (3, 64),
+            AwgKind::Transport => (4, decoded.message.len()),
+        };
+        let mut message = decoded.message.get(..len)?.to_vec();
+        message[..4].copy_from_slice(&(kind as u32).to_le_bytes());
+        Some(message)
+    }
+    pub(crate) fn pad(&self, packet: Bytes) -> Bytes {
+        self.config().prepare_transport_plaintext(packet, 1200)
+    }
+    /// Generates structured signature packets with fresh fields, never static captures.
+    /// These target the other Tsunagi peer, not DNS or SIP services.
+    pub fn signature_packets(&self) -> Vec<Vec<u8>> {
+        if self.signature == 0 {
+            // A complete DNS query. Counts, label lengths and qtype remain valid.
+            let label = format!("n{:016x}", rand::random::<u64>());
+            let mut packet = Vec::from(rand::random::<u16>().to_be_bytes());
+            packet.extend_from_slice(&[1, 0, 0, 1, 0, 0, 0, 0, 0, 0]);
+            packet.push(label.len() as u8);
+            packet.extend_from_slice(label.as_bytes());
+            packet.extend_from_slice(b"\x07example\x03net\0\0");
+            packet.push(if rand::random() { 1 } else { 28 });
+            packet.extend_from_slice(&[0, 1]);
+            vec![packet]
+        } else {
+            let id = rand::random::<u64>();
+            vec![format!("OPTIONS sip:n{id:x}@example.net SIP/2.0\r\nVia: SIP/2.0/UDP example.net;branch=z9hG4bK{id:x}\r\nMax-Forwards: 70\r\nFrom: <sip:a@example.net>;tag={id:x}\r\nTo: <sip:n{id:x}@example.net>\r\nCall-ID: {id:x}@example.net\r\nCSeq: 1 OPTIONS\r\nContent-Length: 0\r\n\r\n").into_bytes()]
+        }
+    }
+    pub(crate) fn junk_packets(&self) -> Vec<Vec<u8>> {
+        (0..self.junk_count)
+            .map(|_| {
+                let mut packet = vec![0; rand::random_range(self.junk_min..=self.junk_max).into()];
+                rand::fill(&mut packet[..]);
+                packet
+            })
+            .collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[test]
+    fn unsafe_peer_profiles_are_rejected_before_the_codec() {
+        let valid = PairProfile::generate();
+        let mut bad = valid.clone();
+        bad.prefix = 11;
+        assert!(bad.validate().is_err());
+        let mut bad = valid.clone();
+        bad.headers[1] = bad.headers[0];
+        assert!(bad.validate().is_err());
+        let mut bad = valid.clone();
+        bad.junk_max = u16::MAX;
+        assert!(bad.validate().is_err());
+        let mut bad = valid.clone();
+        bad.padding = u16::MAX;
+        assert!(bad.validate().is_err());
+        let mut bad = valid.clone();
+        bad.psk = [0; 32];
+        assert!(bad.validate().is_err());
+        let mut bad = valid.clone();
+        bad.signature = 2;
+        assert!(bad.validate().is_err());
+    }
+    #[test]
+    fn signature_templates_have_consistent_framing_and_fresh_fields() {
+        let mut profile = PairProfile::generate();
+        profile.signature = 0;
+        let dns = profile.signature_packets().remove(0);
+        assert_eq!(&dns[2..12], &[1, 0, 0, 1, 0, 0, 0, 0, 0, 0]);
+        let mut at = 12;
+        while dns[at] != 0 {
+            at += 1 + usize::from(dns[at]);
+        }
+        assert_eq!(at + 5, dns.len());
+        assert_eq!(&dns[dns.len() - 2..], &[0, 1]);
+        profile.signature = 1;
+        let sip = profile.signature_packets().remove(0);
+        assert!(sip.starts_with(b"OPTIONS sip:"));
+        assert!(sip.ends_with(b"Content-Length: 0\r\n\r\n"));
+        assert_ne!(sip, profile.signature_packets().remove(0));
+    }
+}

crates/tsunagi-awg/src/session.rs

@@ -0,0 +1,402 @@
+//! One cancellable actor and UDP socket per pair. All payload and health frames
+//! are inside the Amnezia-masked authenticated WireGuard tunnel.
+use crate::profile::PairProfile;
+use boringtun::noise::{Tunn, TunnResult};
+use bytes::Bytes;
+use iroh::EndpointId;
+use std::collections::{HashMap, VecDeque};
+use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+use tokio::net::UdpSocket;
+use tokio::sync::{mpsc, watch};
+use tsunagi::dataplane::transport::{PacketLink, TransportError};
+use tsunagi::{BoxFuture, NetworkId};
+
+pub(crate) const MAX_PAYLOAD: usize = tsunagi::config::MAX_DATA_DATAGRAM;
+const CHUNK: usize = 700;
+const QUEUE: usize = 64;
+const WINDOW: Duration = Duration::from_secs(20);
+
+pub(crate) struct PairLink {
+    network: NetworkId,
+    peer: EndpointId,
+    pub(crate) attempt: [u8; 16],
+    tx: mpsc::Sender<Bytes>,
+    rx: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    stop: watch::Sender<bool>,
+    closed: AtomicBool,
+    remote: Mutex<Option<SocketAddr>>,
+    reason: Mutex<Option<String>>,
+}
+impl std::fmt::Debug for PairLink {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AwgPair")
+            .field("network", &self.network)
+            .field("peer", &self.peer)
+            .finish_non_exhaustive()
+    }
+}
+impl PairLink {
+    pub(crate) fn has_authenticated_path(&self) -> bool {
+        match self.remote.lock() {
+            Ok(remote) => remote.is_some(),
+            Err(_) => false,
+        }
+    }
+}
+impl PacketLink for PairLink {
+    fn network(&self) -> NetworkId {
+        self.network
+    }
+    fn peer(&self) -> EndpointId {
+        self.peer
+    }
+    fn max_datagram_size(&self) -> usize {
+        MAX_PAYLOAD
+    }
+    fn send(&self, payload: Bytes) -> Result<(), TransportError> {
+        if self.is_closed() {
+            return Err(TransportError::Closed);
+        }
+        if payload.len() > MAX_PAYLOAD {
+            return Err(TransportError::TooLarge {
+                size: payload.len(),
+                limit: MAX_PAYLOAD,
+            });
+        }
+        self.tx
+            .try_send(payload)
+            .map_err(|_| TransportError::Other("AWG send queue full or closed".into()))
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut stop = self.stop.subscribe();
+            if *stop.borrow() {
+                return None;
+            }
+            let mut rx = self.rx.lock().await;
+            tokio::select! {biased; _=stop.changed()=>None,packet=rx.recv()=>packet}
+        })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(async move {
+            let mut stop = self.stop.subscribe();
+            if !*stop.borrow() {
+                let _ = stop.changed().await;
+            }
+        })
+    }
+    fn is_closed(&self) -> bool {
+        self.closed.load(Ordering::Acquire) || *self.stop.borrow()
+    }
+    fn close(&self) {
+        self.stop.send_replace(true);
+    }
+    fn failure_reason(&self) -> Option<String> {
+        match self.reason.lock() {
+            Ok(reason) => reason.clone(),
+            Err(_) => None,
+        }
+    }
+    fn path_description(&self) -> String {
+        match self.remote.lock() {
+            Ok(addr) => match *addr {
+                Some(addr) => format!("awg UDP via {addr}"),
+                None => "awg UDP, establishing".into(),
+            },
+            Err(_) => "awg UDP".into(),
+        }
+    }
+}
+impl Drop for PairLink {
+    fn drop(&mut self) {
+        self.stop.send_replace(true);
+    }
+}
+
+pub(crate) fn start(
+    network: NetworkId,
+    peer: EndpointId,
+    socket: UdpSocket,
+    tunn: Tunn,
+    profile: PairProfile,
+    candidates: Vec<SocketAddr>,
+    parent: watch::Receiver<bool>,
+) -> Arc<PairLink> {
+    let (tx, rx) = mpsc::channel(QUEUE);
+    let (in_tx, in_rx) = mpsc::channel(QUEUE);
+    let (stop, _) = watch::channel(false);
+    let link = Arc::new(PairLink {
+        network,
+        peer,
+        attempt: profile.id(),
+        tx,
+        rx: tokio::sync::Mutex::new(in_rx),
+        stop: stop.clone(),
+        closed: AtomicBool::new(false),
+        remote: Mutex::new(None),
+        reason: Mutex::new(None),
+    });
+    let weak = Arc::downgrade(&link);
+    tokio::spawn(async move {
+        let mut engine = Engine {
+            stop: stop.subscribe(),
+            socket,
+            tunn,
+            profile,
+            candidates,
+            remote: None,
+            inbox: in_tx,
+            rx,
+            counter: rand::random(),
+            sequence: 0,
+            pending: VecDeque::new(),
+            samples: VecDeque::new(),
+            assemblies: HashMap::new(),
+            started: Instant::now(),
+            last_authenticated: Instant::now(),
+        };
+        engine.run(parent, &weak).await;
+        if let Some(link) = weak.upgrade() {
+            link.closed.store(true, Ordering::Release);
+            link.stop.send_replace(true);
+        }
+    });
+    link
+}
+
+struct Assembly {
+    created: Instant,
+    count: u16,
+    chunks: Vec<Option<Bytes>>,
+}
+struct Engine {
+    stop: watch::Receiver<bool>,
+    socket: UdpSocket,
+    tunn: Tunn,
+    profile: PairProfile,
+    candidates: Vec<SocketAddr>,
+    remote: Option<SocketAddr>,
+    inbox: mpsc::Sender<Bytes>,
+    rx: mpsc::Receiver<Bytes>,
+    counter: u64,
+    sequence: u64,
+    pending: VecDeque<(u64, Instant)>,
+    samples: VecDeque<bool>,
+    assemblies: HashMap<u64, Assembly>,
+    started: Instant,
+    last_authenticated: Instant,
+}
+enum ResultOwned {
+    Wire(Vec<u8>),
+    Plain(Vec<u8>),
+    Done,
+}
+fn own(result: TunnResult<'_>) -> ResultOwned {
+    match result {
+        TunnResult::WriteToNetwork(p) => ResultOwned::Wire(p.to_vec()),
+        TunnResult::WriteToTunnelV4(p, _) => ResultOwned::Plain(p.to_vec()),
+        _ => ResultOwned::Done,
+    }
+}
+fn ip(payload: &[u8]) -> Bytes {
+    let mut packet = vec![0; 20];
+    packet[0] = 0x45;
+    packet[8] = 64;
+    packet[9] = 253;
+    packet[12..16].copy_from_slice(&Ipv4Addr::new(192, 0, 2, 1).octets());
+    packet[16..20].copy_from_slice(&Ipv4Addr::new(192, 0, 2, 2).octets());
+    packet[2..4].copy_from_slice(&((20 + payload.len()) as u16).to_be_bytes());
+    let sum = packet
+        .as_chunks::<2>()
+        .0
+        .iter()
+        .map(|b| u32::from(u16::from_be_bytes([b[0], b[1]])))
+        .sum::<u32>();
+    let sum = (sum & 0xffff) + (sum >> 16);
+    packet[10..12].copy_from_slice(&(!(sum as u16)).to_be_bytes());
+    packet.extend_from_slice(payload);
+    Bytes::from(packet)
+}
+impl Engine {
+    async fn run(&mut self, mut stop: watch::Receiver<bool>, link: &std::sync::Weak<PairLink>) {
+        if *self.stop.borrow() || *stop.borrow() {
+            return;
+        }
+        let mut buffer = [0; 1201];
+        let mut scratch = [0; 4096];
+        let mut timer = tokio::time::interval(Duration::from_millis(250));
+        timer.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+        let mut last_probe = Instant::now() - Duration::from_secs(2);
+        loop {
+            tokio::select! {
+                biased;
+                _=stop.changed()=>break,
+                _=self.stop.changed()=>break,
+                _=timer.tick()=>{
+                    let result=own(self.tunn.update_timers(&mut scratch));self.process(result,None).await;
+                    if last_probe.elapsed()>=Duration::from_secs(1) {
+                        last_probe=Instant::now();self.sequence=self.sequence.wrapping_add(1);
+                        self.pending.push_back((self.sequence,Instant::now()));
+                        let mut probe=vec![1];probe.extend_from_slice(&self.sequence.to_be_bytes());
+                        probe.resize(713,0);rand::fill(&mut probe[9..]);self.encrypt(&probe).await;
+                    }
+                    while self.pending.front().is_some_and(|(_,at)|at.elapsed()>=Duration::from_secs(3)) {
+                        self.pending.pop_front();self.samples.push_back(false);
+                    }
+                    while self.samples.len()>20 {self.samples.pop_front();}
+                    self.assemblies.retain(|_,a|a.created.elapsed()<Duration::from_secs(5));
+                    if self.started.elapsed()>WINDOW && (self.last_authenticated.elapsed()>Duration::from_secs(10)
+                        || (self.samples.len()>=15&&self.samples.iter().filter(|ok|!**ok).count()*100>=self.samples.len()*60)) {
+                        if let Some(link)=link.upgrade() && let Ok(mut reason)=link.reason.lock() {*reason=Some("AWG data delivery checks: persistent loss or timeout".into());}
+                        break;
+                    }
+                }
+                packet=self.socket.recv_from(&mut buffer)=>{
+                    if let Ok((len,source))=packet && let Some(message)=self.profile.decode(&buffer[..len]) {
+                            let result=own(self.tunn.decapsulate(None,&message,&mut scratch));
+                            if matches!(result,ResultOwned::Plain(_)) {
+                                self.remote=Some(source);self.last_authenticated=Instant::now();
+                                if let Some(link)=link.upgrade() && let Ok(mut remote)=link.remote.lock(){*remote=Some(source);}
+                            }
+                            self.process(result,Some(source)).await;
+                            for _ in 0..16 {
+                                let result=own(self.tunn.decapsulate(None,&[],&mut scratch));
+                                if matches!(result,ResultOwned::Done){break;}
+                                self.process(result,Some(source)).await;
+                            }
+                    }
+                }
+                packet=self.rx.recv()=>{
+                    let Some(packet)=packet else{break;};
+                    let count=packet.len().max(1).div_ceil(CHUNK) as u16;
+                    self.counter=self.counter.wrapping_add(1);
+                    for (index,chunk) in packet.chunks(CHUNK).enumerate() {
+                        let mut payload=vec![3];payload.extend_from_slice(&self.counter.to_be_bytes());
+                        payload.extend_from_slice(&(index as u16).to_be_bytes());payload.extend_from_slice(&count.to_be_bytes());
+                        payload.extend_from_slice(chunk);self.encrypt(&payload).await;
+                        // Large logical datagrams must let local readers and other
+                        // pairs drain their bounded socket/command queues.
+                        if (index + 1).is_multiple_of(16) {tokio::task::yield_now().await;}
+                    }
+                }
+
+            }
+        }
+    }
+    async fn encrypt(&mut self, payload: &[u8]) {
+        let packet = self.profile.pad(ip(payload));
+        let mut scratch = [0; 4096];
+        let result = own(self.tunn.encapsulate(&packet, &mut scratch));
+        self.process(result, None).await;
+    }
+    async fn wire(&self, raw: &[u8], target: Option<SocketAddr>) {
+        let Some(wire) = self.profile.encode(raw) else {
+            return;
+        };
+        let targets: Vec<_> = target
+            .or(self.remote)
+            .map(|a| vec![a])
+            .unwrap_or_else(|| self.candidates.clone());
+        for addr in targets {
+            if raw.first() == Some(&1) {
+                for signature in self.profile.signature_packets() {
+                    let _ = self.send_wire(&signature, addr).await;
+                }
+                for junk in self.profile.junk_packets() {
+                    let _ = self.send_wire(&junk, addr).await;
+                }
+            }
+            let _ = self.send_wire(&wire, addr).await;
+        }
+    }
+    async fn send_wire(&self, bytes: &[u8], address: SocketAddr) {
+        let mut stop = self.stop.clone();
+        if *stop.borrow() {
+            return;
+        }
+        tokio::select! {biased; _=stop.changed()=>{}, _=tokio::time::timeout(Duration::from_millis(100),self.socket.send_to(bytes,address))=>{}}
+    }
+    async fn process(&mut self, result: ResultOwned, target: Option<SocketAddr>) {
+        match result {
+            ResultOwned::Wire(raw) => self.wire(&raw, target).await,
+            ResultOwned::Plain(packet) => {
+                let Some(payload) = packet.get(20..) else {
+                    return;
+                };
+                let Some(seq) = payload
+                    .get(1..9)
+                    .and_then(|b| b.try_into().ok())
+                    .map(u64::from_be_bytes)
+                else {
+                    return;
+                };
+                match payload.first() {
+                    Some(1) => {
+                        let mut pong = vec![2];
+                        pong.extend_from_slice(&seq.to_be_bytes());
+                        pong.resize(713, 0);
+                        let packet = self.profile.pad(ip(&pong));
+                        let mut scratch = [0; 4096];
+                        if let ResultOwned::Wire(raw) =
+                            own(self.tunn.encapsulate(&packet, &mut scratch))
+                        {
+                            self.wire(&raw, target).await;
+                        }
+                    }
+                    Some(2) => {
+                        if let Some(index) =
+                            self.pending.iter().position(|(pending, _)| *pending == seq)
+                        {
+                            self.pending.remove(index);
+                            self.samples.push_back(true);
+                        }
+                    }
+                    Some(3) => self.fragment(seq, payload),
+                    _ => {}
+                }
+            }
+            ResultOwned::Done => {}
+        }
+    }
+    fn fragment(&mut self, id: u64, payload: &[u8]) {
+        let Some(header) = payload.get(9..13) else {
+            return;
+        };
+        let index = u16::from_be_bytes([header[0], header[1]]);
+        let count = u16::from_be_bytes([header[2], header[3]]);
+        let chunk = &payload[13..];
+        if count == 0
+            || usize::from(count) > MAX_PAYLOAD.div_ceil(CHUNK)
+            || index >= count
+            || chunk.len() > CHUNK
+        {
+            return;
+        }
+        if !self.assemblies.contains_key(&id) && self.assemblies.len() >= 8 {
+            return;
+        }
+        let assembly = self.assemblies.entry(id).or_insert_with(|| Assembly {
+            created: Instant::now(),
+            count,
+            chunks: vec![None; count.into()],
+        });
+        if assembly.count != count {
+            return;
+        }
+        assembly.chunks[usize::from(index)] = Some(Bytes::copy_from_slice(chunk));
+        if assembly.chunks.iter().all(Option::is_some)
+            && let Some(assembly) = self.assemblies.remove(&id)
+        {
+            let mut packet = Vec::new();
+            for chunk in assembly.chunks.into_iter().flatten() {
+                packet.extend_from_slice(&chunk);
+            }
+            if packet.len() <= MAX_PAYLOAD {
+                let _ = self.inbox.try_send(Bytes::from(packet));
+            }
+        }
+    }
+}

crates/tsunagi-awg/src/stun.rs

@@ -0,0 +1,101 @@
+//! Optional, bounded server-reflexive discovery on the pair's own UDP socket.
+//! A STUN answer is a candidate only; WireGuard authentication verifies a path.
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+use std::time::Duration;
+use tokio::net::UdpSocket;
+const COOKIE: [u8; 4] = [0x21, 0x12, 0xa4, 0x42];
+pub(crate) async fn mapped(socket: &UdpSocket, server: &str) -> Option<SocketAddr> {
+    tokio::time::timeout(Duration::from_millis(1200), async {
+        let address = tokio::net::lookup_host(server)
+            .await
+            .ok()?
+            .find(SocketAddr::is_ipv4)?;
+        let transaction: [u8; 12] = rand::random();
+        let mut request = [0; 20];
+        request[..2].copy_from_slice(&[0, 1]);
+        request[4..8].copy_from_slice(&COOKIE);
+        request[8..].copy_from_slice(&transaction);
+        socket.send_to(&request, address).await.ok()?;
+        let mut bytes = [0; 1024];
+        loop {
+            let (len, source) = socket.recv_from(&mut bytes).await.ok()?;
+            if source == address
+                && let Some(mapped) = parse(&bytes[..len], &transaction)
+            {
+                return Some(mapped);
+            }
+        }
+    })
+    .await
+    .ok()
+    .flatten()
+}
+fn parse(bytes: &[u8], transaction: &[u8; 12]) -> Option<SocketAddr> {
+    if bytes.len() < 20
+        || bytes[..2] != [1, 1]
+        || bytes[4..8] != COOKIE
+        || &bytes[8..20] != transaction
+    {
+        return None;
+    }
+    let declared = u16::from_be_bytes(bytes[2..4].try_into().ok()?) as usize;
+    if !declared.is_multiple_of(4) || declared != bytes.len() - 20 {
+        return None;
+    }
+    let mut at = 20;
+    while at + 4 <= bytes.len() {
+        let kind = u16::from_be_bytes(bytes[at..at + 2].try_into().ok()?);
+        let len = u16::from_be_bytes(bytes[at + 2..at + 4].try_into().ok()?) as usize;
+        at += 4;
+        let attribute = bytes.get(at..at.checked_add(len)?)?;
+        if kind == 0x0020 && attribute.len() == 8 && attribute[0..2] == [0, 1] {
+            let port = u16::from_be_bytes([attribute[2] ^ COOKIE[0], attribute[3] ^ COOKIE[1]]);
+            let ip = Ipv4Addr::new(
+                attribute[4] ^ COOKIE[0],
+                attribute[5] ^ COOKIE[1],
+                attribute[6] ^ COOKIE[2],
+                attribute[7] ^ COOKIE[3],
+            );
+            if port == 0 || ip.is_unspecified() || ip.is_multicast() || ip.is_broadcast() {
+                return None;
+            }
+            return Some(SocketAddr::new(IpAddr::V4(ip), port));
+        }
+        at = at.checked_add((len + 3) & !3)?;
+    }
+    None
+}
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[tokio::test]
+    async fn mapping_uses_the_actual_socket_and_rejects_unrelated_or_truncated_answers() {
+        let server = UdpSocket::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
+        let address = server.local_addr().unwrap();
+        let socket = UdpSocket::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
+        let expected = socket.local_addr().unwrap();
+        let task = tokio::spawn(async move {
+            let mut request = [0; 128];
+            let (len, source) = server.recv_from(&mut request).await.unwrap();
+            assert_eq!(len, 20);
+            assert_eq!(source, expected);
+            let tx: [u8; 12] = request[8..20].try_into().unwrap();
+            let mut reply = request[..20].to_vec();
+            reply[..4].copy_from_slice(&[1, 1, 0, 12]);
+            reply.extend_from_slice(&[0, 0x20, 0, 8, 0, 1]);
+            let port = source.port().to_be_bytes();
+            reply.extend_from_slice(&[port[0] ^ COOKIE[0], port[1] ^ COOKIE[1]]);
+            for (i, b) in Ipv4Addr::LOCALHOST.octets().into_iter().enumerate() {
+                reply.push(b ^ COOKIE[i]);
+            }
+            assert!(parse(&reply, &[0; 12]).is_none());
+            for len in 0..reply.len() {
+                assert!(parse(&reply[..len], &tx).is_none());
+            }
+            server.send_to(&reply, source).await.unwrap();
+        });
+        assert_eq!(mapped(&socket, &address.to_string()).await, Some(expected));
+        task.await.unwrap();
+    }
+}

crates/tsunagi-awg/src/transport.rs

@@ -0,0 +1,720 @@
+//! Private pair negotiation. No secret profiles in capabilities or durable state.
+use crate::{
+    AWG_PROTOCOL, AWG_WIRE_VERSION,
+    profile::PairProfile,
+    session::{self, PairLink},
+};
+use boringtun::{
+    noise::Tunn,
+    x25519::{PublicKey, StaticSecret},
+};
+use iroh::EndpointId;
+use serde::{Deserialize, Serialize};
+use std::collections::{HashMap, HashSet};
+use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::{Arc, Mutex, OnceLock, Weak};
+use std::time::Duration;
+use tokio::net::UdpSocket;
+use tokio::sync::{mpsc, oneshot, watch};
+use tsunagi::dataplane::transport::TransportPayload;
+use tsunagi::dataplane::transport::{
+    ControlSender, InboundLink, PacketLink, PacketTransport, SharedLink, TransportError,
+};
+use tsunagi::{BoxFuture, NetworkId};
+
+type Pair = (NetworkId, EndpointId);
+#[derive(Serialize, Deserialize)]
+enum Negotiation {
+    Offer {
+        version: u16,
+        profile: PairProfile,
+        key: [u8; 32],
+        #[serde(deserialize_with = "bounded_addresses")]
+        addresses: Vec<SocketAddr>,
+    },
+    Answer {
+        attempt: [u8; 16],
+        key: [u8; 32],
+        #[serde(deserialize_with = "bounded_addresses")]
+        addresses: Vec<SocketAddr>,
+    },
+}
+fn bounded_addresses<'de, D: serde::Deserializer<'de>>(
+    deserializer: D,
+) -> Result<Vec<SocketAddr>, D::Error> {
+    struct Visitor;
+    impl<'de> serde::de::Visitor<'de> for Visitor {
+        type Value = Vec<SocketAddr>;
+        fn expecting(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            f.write_str("at most 32 UDP candidates")
+        }
+        fn visit_seq<A: serde::de::SeqAccess<'de>>(
+            self,
+            mut sequence: A,
+        ) -> Result<Self::Value, A::Error> {
+            if sequence.size_hint().is_some_and(|n| n > 32) {
+                return Err(serde::de::Error::custom("too many UDP candidates"));
+            }
+            let mut addresses = Vec::with_capacity(sequence.size_hint().unwrap_or(0).min(32));
+            while let Some(address) = sequence.next_element()? {
+                if addresses.len() == 32 {
+                    return Err(serde::de::Error::custom("too many UDP candidates"));
+                }
+                addresses.push(address);
+            }
+            Ok(addresses)
+        }
+    }
+    deserializer.deserialize_seq(Visitor)
+}
+struct Pending {
+    profile: PairProfile,
+    key: StaticSecret,
+    socket: UdpSocket,
+    reply: oneshot::Sender<Result<SharedLink, TransportError>>,
+}
+struct Inner {
+    stopped: watch::Sender<bool>,
+    local: EndpointId,
+    bind: Ipv4Addr,
+    stun: Option<String>,
+    busy: Mutex<HashSet<Pair>>,
+    control: OnceLock<Arc<dyn ControlSender>>,
+    allowed: Mutex<HashSet<Pair>>,
+    pending: Mutex<HashMap<Pair, Pending>>,
+    links: Mutex<HashMap<Pair, Arc<PairLink>>>,
+    inbound: mpsc::Sender<InboundLink>,
+    port_by_pair: Mutex<HashMap<Pair, u16>>,
+}
+/// Public diagnostics for one pair. No key or masking parameters are exposed.
+#[derive(Debug, Clone)]
+pub struct PairOverview {
+    /// Other endpoint's authenticated identity.
+    pub peer: EndpointId,
+    /// Fresh connection-attempt identifier.
+    pub attempt: [u8; 16],
+    /// Last known UDP port; None if diagnostics race with pair retirement.
+    pub local_port: Option<u16>,
+    /// Actual transport path, if established.
+    pub path: String,
+    /// Whether this attempt has been retired.
+    pub closed: bool,
+}
+/// A bounded set of independently configured pair engines.
+pub struct AwgTransport {
+    inner: Arc<Inner>,
+    commands: mpsc::Sender<(NetworkId, EndpointId, Negotiation)>,
+    incoming: tokio::sync::Mutex<mpsc::Receiver<InboundLink>>,
+}
+impl std::fmt::Debug for AwgTransport {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AwgTransport")
+            .field("local", &self.inner.local)
+            .finish_non_exhaustive()
+    }
+}
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(g) => g,
+        Err(p) => p.into_inner(),
+    }
+}
+fn error(text: &str) -> TransportError {
+    TransportError::Other(text.into())
+}
+async fn addresses(inner: &Inner, socket: &UdpSocket) -> Result<Vec<SocketAddr>, TransportError> {
+    let local = socket
+        .local_addr()
+        .map_err(|_| error("AWG socket address unavailable"))?;
+    let mut candidates = if inner.bind.is_loopback() {
+        vec![local]
+    } else {
+        tsunagi_wireguard::local_ip_candidates()
+            .into_iter()
+            .filter(|ip| ip.is_ipv4())
+            .take(31)
+            .map(|ip| SocketAddr::new(ip, local.port()))
+            .collect()
+    };
+    if let Some(server) = &inner.stun
+        && let Some(mapped) = crate::stun::mapped(socket, server).await
+        && !candidates.contains(&mapped)
+    {
+        candidates.push(mapped);
+    }
+    if candidates.is_empty() {
+        return Err(error("AWG has no UDP address candidates"));
+    }
+    Ok(candidates)
+}
+fn validate_endpoint(key: &[u8; 32], addresses: &[SocketAddr]) -> Result<(), TransportError> {
+    if *key == [0; 32]
+        || addresses.is_empty()
+        || addresses.len() > 32
+        || addresses.iter().any(|address| {
+            address.port() == 0
+                || !address.is_ipv4()
+                || address.ip().is_unspecified()
+                || address.ip().is_multicast()
+                || matches!(address.ip(), std::net::IpAddr::V4(ip) if ip.is_broadcast())
+        })
+    {
+        return Err(error("invalid AWG endpoint"));
+    }
+    Ok(())
+}
+
+impl AwgTransport {
+    /// Starts the coordinator. Production binds wildcard IPv4; tests use loopback.
+    pub fn new(local: EndpointId, bind: Ipv4Addr) -> Arc<Self> {
+        Self::new_with_stun(local, bind, None)
+    }
+    /// Enables an optional STUN server (host:port) for this pair's own socket.
+    /// Local-only callers pass None; failure still permits local UDP candidates.
+    pub fn new_with_stun(local: EndpointId, bind: Ipv4Addr, stun: Option<String>) -> Arc<Self> {
+        let (commands, mut rx) = mpsc::channel(64);
+        let (inbound, incoming) = mpsc::channel(64);
+        let (stopped, _) = watch::channel(false);
+        let inner = Arc::new(Inner {
+            stopped,
+            local,
+            bind,
+            stun,
+            busy: Mutex::default(),
+            control: OnceLock::new(),
+            allowed: Mutex::default(),
+            pending: Mutex::default(),
+            links: Mutex::default(),
+            inbound,
+            port_by_pair: Mutex::default(),
+        });
+        let weak = Arc::downgrade(&inner);
+        let mut stopped = inner.stopped.subscribe();
+        tokio::spawn(async move {
+            while let Some((network, peer, message)) =
+                tokio::select! { biased; _=stopped.changed()=>None, message=rx.recv()=>message }
+            {
+                let Some(inner) = weak.upgrade() else {
+                    break;
+                };
+                if !lock(&inner.allowed).contains(&(network, peer))
+                    || !lock(&inner.busy).insert((network, peer))
+                {
+                    continue;
+                }
+                tokio::spawn(async move {
+                    let mut stop = inner.stopped.subscribe();
+                    tokio::select! {biased; _=stop.changed()=>{}, _=async {
+                        if let Err(err)=receive(&inner,(network,peer),message).await {tracing::debug!(%err,"AWG pair negotiation rejected");}
+                    }=>{}}
+                    lock(&inner.busy).remove(&(network, peer));
+                });
+            }
+        });
+        Arc::new(Self {
+            inner,
+            commands,
+            incoming: tokio::sync::Mutex::new(incoming),
+        })
+    }
+    /// Registers a peer after its authenticated capability has been validated.
+    pub fn allow_peer(&self, network: NetworkId, peer: EndpointId) {
+        let mut allowed = lock(&self.inner.allowed);
+        if !*self.inner.stopped.borrow() && allowed.len() < 128 {
+            allowed.insert((network, peer));
+        }
+    }
+    /// Retires a pair when its network membership session goes away.
+    pub fn forget_peer(&self, network: NetworkId, peer: EndpointId) {
+        let pair = (network, peer);
+        lock(&self.inner.allowed).remove(&pair);
+        lock(&self.inner.pending).remove(&pair);
+        lock(&self.inner.port_by_pair).remove(&pair);
+        if let Some(link) = lock(&self.inner.links).remove(&pair) {
+            link.close();
+        }
+    }
+    /// Stops all pair engines in this network.
+    pub fn stop_network(&self, network: NetworkId) {
+        let peers: Vec<_> = lock(&self.inner.allowed)
+            .iter()
+            .filter(|(n, _)| *n == network)
+            .map(|(_, p)| *p)
+            .collect();
+        for peer in peers {
+            self.forget_peer(network, peer);
+        }
+    }
+    /// Stops all sockets and coordinator work without waiting for remote peers.
+    pub fn shutdown(&self) {
+        self.inner.stopped.send_replace(true);
+        lock(&self.inner.allowed).clear();
+        lock(&self.inner.pending).clear();
+        for (_, link) in lock(&self.inner.links).drain() {
+            link.close();
+        }
+        lock(&self.inner.port_by_pair).clear();
+    }
+    /// Receives a successfully negotiated and authenticated inbound pair link.
+    pub async fn next_inbound(&self) -> Option<InboundLink> {
+        let mut stopped = self.inner.stopped.subscribe();
+        if *stopped.borrow() {
+            return None;
+        }
+        let mut incoming = self.incoming.lock().await;
+        tokio::select! {biased; _=stopped.changed()=>None, link=incoming.recv()=>link}
+    }
+    /// Takes a public snapshot of this network's pair attempts.
+    pub fn overview(&self, network: NetworkId) -> Vec<PairOverview> {
+        let links = lock(&self.inner.links);
+        let ports = lock(&self.inner.port_by_pair);
+        links
+            .iter()
+            .filter(|((n, _), _)| *n == network)
+            .map(|((_, peer), link)| PairOverview {
+                peer: *peer,
+                attempt: link.attempt,
+                local_port: ports.get(&(network, *peer)).copied(),
+                path: link.path_description(),
+                closed: link.is_closed(),
+            })
+            .collect()
+    }
+}
+impl PacketTransport for AwgTransport {
+    fn name(&self) -> &str {
+        AWG_PROTOCOL
+    }
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+    fn attach_control(&self, control: Arc<dyn ControlSender>) {
+        let _ = self.inner.control.set(control);
+    }
+    fn on_control(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &str,
+        payload: &[u8],
+    ) -> Result<(), TransportError> {
+        if protocol != AWG_PROTOCOL || !lock(&self.inner.allowed).contains(&(network, peer)) {
+            return Err(TransportError::Declined(protocol.into()));
+        }
+        if payload.len() > 4096 {
+            return Err(error("AWG negotiation too large"));
+        }
+        let message: Negotiation =
+            postcard::from_bytes(payload).map_err(|_| error("malformed AWG negotiation"))?;
+        match &message {
+            Negotiation::Offer {
+                version,
+                profile,
+                key,
+                addresses,
+            } => {
+                if *version != AWG_WIRE_VERSION || peer.as_bytes() >= self.inner.local.as_bytes() {
+                    return Err(error("invalid AWG offer role/version"));
+                }
+                profile.validate().map_err(error)?;
+                validate_endpoint(key, addresses)?;
+            }
+            Negotiation::Answer { key, addresses, .. } => validate_endpoint(key, addresses)?,
+        }
+        self.commands
+            .try_send((network, peer, message))
+            .map_err(|_| error("AWG negotiation queue full"))
+    }
+    fn open<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &'a str,
+    ) -> BoxFuture<'a, Result<SharedLink, TransportError>> {
+        Box::pin(async move {
+            let pair = (network, peer);
+            if protocol != AWG_PROTOCOL
+                || self.inner.local.as_bytes() >= peer.as_bytes()
+                || !lock(&self.inner.allowed).contains(&pair)
+            {
+                return Err(TransportError::Declined(protocol.into()));
+            }
+            let control = self
+                .inner
+                .control
+                .get()
+                .ok_or_else(|| error("AWG control plane unavailable"))?;
+            let socket = UdpSocket::bind((self.inner.bind, 0))
+                .await
+                .map_err(|_| error("AWG UDP bind failed"))?;
+            let profile = PairProfile::generate();
+            let attempt = profile.id();
+            let key = StaticSecret::from(rand::random::<[u8; 32]>());
+            let public = PublicKey::from(&key).to_bytes();
+            let message = Negotiation::Offer {
+                version: AWG_WIRE_VERSION,
+                profile: profile.clone(),
+                key: public,
+                addresses: addresses(&self.inner, &socket).await?,
+            };
+            let (reply, rx) = oneshot::channel();
+            if lock(&self.inner.pending).len() >= 128 {
+                return Err(error("too many AWG pair negotiations"));
+            }
+            lock(&self.inner.pending).insert(
+                pair,
+                Pending {
+                    profile,
+                    key,
+                    socket,
+                    reply,
+                },
+            );
+            // A cancellation guard also retires partially started attempts.
+            let mut guard = AttemptGuard {
+                inner: Arc::downgrade(&self.inner),
+                pair,
+                attempt,
+                completed: false,
+            };
+            let bytes =
+                postcard::to_stdvec(&message).map_err(|_| error("AWG offer encoding failed"))?;
+            control
+                .send(network, peer, AWG_PROTOCOL, TransportPayload::new(bytes))
+                .await?;
+            let result = tokio::time::timeout(Duration::from_secs(15), rx)
+                .await
+                .map_err(|_| error("AWG negotiation/handshake timed out"))?
+                .map_err(|_| error("AWG attempt cancelled"))?;
+            guard.completed = result.is_ok();
+            result
+        })
+    }
+}
+struct AttemptGuard {
+    inner: Weak<Inner>,
+    pair: Pair,
+    attempt: [u8; 16],
+    completed: bool,
+}
+impl Drop for AttemptGuard {
+    fn drop(&mut self) {
+        if let Some(inner) = self.inner.upgrade() {
+            let mut pending = lock(&inner.pending);
+            if pending
+                .get(&self.pair)
+                .is_some_and(|p| p.profile.id() == self.attempt)
+            {
+                pending.remove(&self.pair);
+            }
+            drop(pending);
+            if !self.completed
+                && let Some(link) = lock(&inner.links).get(&self.pair)
+                && link.attempt == self.attempt
+            {
+                link.close();
+            }
+        }
+    }
+}
+async fn receive(
+    inner: &Arc<Inner>,
+    pair: Pair,
+    message: Negotiation,
+) -> Result<(), TransportError> {
+    if !lock(&inner.allowed).contains(&pair) {
+        return Err(error("AWG peer no longer authenticated"));
+    }
+    match message {
+        Negotiation::Offer {
+            profile,
+            key,
+            addresses: candidates,
+            ..
+        } => {
+            if let Some(link) = lock(&inner.links).get(&pair) {
+                if link.attempt == profile.id() {
+                    return Err(error("duplicate AWG offer"));
+                }
+                link.close();
+            }
+            let socket = UdpSocket::bind((inner.bind, 0))
+                .await
+                .map_err(|_| error("AWG UDP bind failed"))?;
+            let local_port = socket
+                .local_addr()
+                .map_err(|_| error("AWG socket address unavailable"))?
+                .port();
+            let local_key = StaticSecret::from(rand::random::<[u8; 32]>());
+            let public = PublicKey::from(&local_key).to_bytes();
+            let answer = Negotiation::Answer {
+                attempt: profile.id(),
+                key: public,
+                addresses: addresses(inner, &socket).await?,
+            };
+            let tunn = Tunn::new(
+                local_key,
+                PublicKey::from(key),
+                Some(profile.psk),
+                Some(1),
+                rand::random(),
+                None,
+            );
+
+            let link = session::start(
+                pair.0,
+                pair.1,
+                socket,
+                tunn,
+                profile,
+                candidates,
+                inner.stopped.subscribe(),
+            );
+            if !lock(&inner.allowed).contains(&pair) {
+                link.close();
+                return Err(error("AWG membership ended"));
+            }
+            lock(&inner.links).insert(pair, link.clone());
+            lock(&inner.port_by_pair).insert(pair, local_port);
+            let control = inner
+                .control
+                .get()
+                .ok_or_else(|| error("AWG control unavailable"))?;
+            if let Err(err) = control
+                .send(
+                    pair.0,
+                    pair.1,
+                    AWG_PROTOCOL,
+                    TransportPayload::new(
+                        postcard::to_stdvec(&answer)
+                            .map_err(|_| error("AWG answer encoding failed"))?,
+                    ),
+                )
+                .await
+            {
+                link.close();
+                return Err(err);
+            }
+            publish_when_up(Arc::downgrade(inner), pair, link, None);
+        }
+        Negotiation::Answer {
+            attempt,
+            key,
+            addresses: candidates,
+        } => {
+            if !lock(&inner.pending)
+                .get(&pair)
+                .is_some_and(|p| p.profile.id() == attempt)
+            {
+                return Err(error("stale AWG answer"));
+            }
+            let pending = lock(&inner.pending)
+                .remove(&pair)
+                .ok_or_else(|| error("AWG attempt cancelled"))?;
+            let local_port = pending
+                .socket
+                .local_addr()
+                .map_err(|_| error("AWG socket address unavailable"))?
+                .port();
+            let tunn = Tunn::new(
+                pending.key,
+                PublicKey::from(key),
+                Some(pending.profile.psk),
+                Some(1),
+                rand::random(),
+                None,
+            );
+
+            let link = session::start(
+                pair.0,
+                pair.1,
+                pending.socket,
+                tunn,
+                pending.profile,
+                candidates,
+                inner.stopped.subscribe(),
+            );
+            if !lock(&inner.allowed).contains(&pair) {
+                link.close();
+                return Err(error("AWG membership ended"));
+            }
+            lock(&inner.links).insert(pair, link.clone());
+            lock(&inner.port_by_pair).insert(pair, local_port);
+            publish_when_up(Arc::downgrade(inner), pair, link, Some(pending.reply));
+        }
+    }
+    Ok(())
+}
+fn publish_when_up(
+    inner: Weak<Inner>,
+    pair: Pair,
+    link: Arc<PairLink>,
+    reply: Option<oneshot::Sender<Result<SharedLink, TransportError>>>,
+) {
+    tokio::spawn(async move {
+        let mut reply = reply;
+        let result = tokio::time::timeout(Duration::from_secs(14), async {
+            loop {
+                if link.is_closed() || reply.as_ref().is_some_and(oneshot::Sender::is_closed) {
+                    return Err(error("AWG pair closed during handshake"));
+                }
+                if link.has_authenticated_path() {
+                    return Ok(());
+                }
+                tokio::time::sleep(Duration::from_millis(25)).await;
+            }
+        })
+        .await;
+        if !matches!(result, Ok(Ok(()))) {
+            link.close();
+            if let Some(reply) = reply.take() {
+                let _ = reply.send(Err(error("AWG authenticated data path timed out")));
+            }
+            return;
+        }
+        let Some(inner) = inner.upgrade() else {
+            link.close();
+            return;
+        };
+        if !lock(&inner.allowed).contains(&pair) {
+            link.close();
+            return;
+        }
+        if link.is_closed()
+            || !lock(&inner.links)
+                .get(&pair)
+                .is_some_and(|current| current.attempt == link.attempt)
+        {
+            link.close();
+            return;
+        }
+        let shared: SharedLink = link.clone();
+        if let Some(reply) = reply {
+            if reply.send(Ok(shared)).is_err() {
+                link.close();
+            }
+        } else if inner
+            .inbound
+            .try_send(InboundLink {
+                network: pair.0,
+                peer: pair.1,
+                protocol: AWG_PROTOCOL.into(),
+                link: shared,
+            })
+            .is_err()
+        {
+            link.close();
+        }
+    });
+}
+impl Drop for AwgTransport {
+    fn drop(&mut self) {
+        self.shutdown();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[tokio::test]
+    async fn concurrent_admission_is_bounded_across_networks_and_retirement_releases_slots() {
+        let identity = tsunagi::identity::DeviceIdentity::generate();
+        let transport = AwgTransport::new(identity.endpoint_id(), Ipv4Addr::LOCALHOST);
+        let peer = tsunagi::identity::DeviceIdentity::generate().endpoint_id();
+        std::thread::scope(|scope| {
+            for thread in 0..8u8 {
+                let transport = &transport;
+                scope.spawn(move || {
+                    for item in 0..32u8 {
+                        let mut id = [0; 32];
+                        id[0] = thread;
+                        id[1] = item;
+                        transport.allow_peer(NetworkId::from_bytes(id), peer);
+                    }
+                });
+            }
+        });
+        assert_eq!(lock(&transport.inner.allowed).len(), 128);
+        let network = lock(&transport.inner.allowed).iter().next().unwrap().0;
+        transport.stop_network(network);
+        let new = NetworkId::from_bytes([255; 32]);
+        transport.allow_peer(new, peer);
+        assert!(lock(&transport.inner.allowed).contains(&(new, peer)));
+        assert_eq!(lock(&transport.inner.allowed).len(), 128);
+        transport.shutdown();
+        assert!(lock(&transport.inner.allowed).is_empty());
+    }
+    #[tokio::test]
+    async fn private_negotiation_bounds_roles_and_membership_before_starting_engines() {
+        let mut ids = [
+            tsunagi::identity::DeviceIdentity::generate().endpoint_id(),
+            tsunagi::identity::DeviceIdentity::generate().endpoint_id(),
+        ];
+        ids.sort_by_key(|id| *id.as_bytes());
+        let peer = ids[0];
+        let local = ids[1];
+        let network = NetworkId::from_bytes([7; 32]);
+        let transport = AwgTransport::new(local, Ipv4Addr::LOCALHOST);
+        let offer = |version, key, addresses| Negotiation::Offer {
+            version,
+            profile: PairProfile::generate(),
+            key,
+            addresses,
+        };
+        let address = SocketAddr::from((Ipv4Addr::LOCALHOST, 12345));
+        let encode = |message: &Negotiation| postcard::to_stdvec(message).unwrap();
+        let valid = encode(&offer(AWG_WIRE_VERSION, [1; 32], vec![address]));
+        assert!(
+            transport
+                .on_control(network, peer, AWG_PROTOCOL, &valid)
+                .is_err()
+        );
+        transport.allow_peer(network, peer);
+        assert!(
+            transport
+                .on_control(network, peer, "wrong-protocol", &valid)
+                .is_err()
+        );
+        assert!(
+            transport
+                .on_control(network, peer, AWG_PROTOCOL, &[0; 4097])
+                .is_err()
+        );
+        for bytes in [
+            encode(&offer(99, [1; 32], vec![address])),
+            encode(&offer(AWG_WIRE_VERSION, [0; 32], vec![address])),
+            encode(&offer(AWG_WIRE_VERSION, [1; 32], vec![])),
+            encode(&offer(AWG_WIRE_VERSION, [1; 32], vec![address; 33])),
+        ] {
+            assert!(
+                transport
+                    .on_control(network, peer, AWG_PROTOCOL, &bytes)
+                    .is_err()
+            );
+        }
+        for len in 0..valid.len() {
+            assert!(
+                transport
+                    .on_control(network, peer, AWG_PROTOCOL, &valid[..len])
+                    .is_err()
+            );
+        }
+        let wrong_role = AwgTransport::new(peer, Ipv4Addr::LOCALHOST);
+        wrong_role.allow_peer(network, local);
+        assert!(
+            wrong_role
+                .on_control(network, local, AWG_PROTOCOL, &valid)
+                .is_err()
+        );
+        transport.forget_peer(network, peer);
+        assert!(
+            transport
+                .on_control(network, peer, AWG_PROTOCOL, &valid)
+                .is_err()
+        );
+        assert!(transport.overview(network).is_empty());
+        transport.shutdown();
+        wrong_role.shutdown();
+    }
+}

crates/tsunagi-awg/tests/pairs.rs

@@ -0,0 +1,773 @@
+//! Real membership, UDP sockets, pair negotiation and nested WireGuard encryption.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+use bytes::Bytes;
+use std::{
+    net::Ipv4Addr,
+    path::Path,
+    sync::{
+        Arc,
+        atomic::{AtomicBool, AtomicU64, Ordering},
+    },
+    time::Duration,
+};
+use tempfile::TempDir;
+use tsunagi::dataplane::IpPlugin;
+use tsunagi::dataplane::transport::{
+    ControlSender, PacketLink, PacketTransport, SharedLink, TransportError,
+};
+use tsunagi::discovery::SharedMemoryDiscovery;
+use tsunagi::overlay::{MemoryTun, MemoryTunFactory};
+use tsunagi::testing::{config_with, network, wait_for_peers, wait_until};
+use tsunagi::{Agent, BoxFuture, NetworkId};
+use tsunagi_awg::{AWG_PROTOCOL, AwgPlugin, AwgTransport};
+use tsunagi_wg_quic::WireguardConfig;
+
+struct Node {
+    agent: Agent,
+    plugin: Arc<AwgPlugin>,
+    quic: Option<Arc<tsunagi_wg_quic::WireguardPlugin>>,
+    transport: Arc<AwgTransport>,
+    tuns: MemoryTunFactory,
+    loss: Arc<AtomicBool>,
+}
+impl Node {
+    async fn open(root: &Path, discovery: &SharedMemoryDiscovery, interface: &str) -> Self {
+        Self::open_with_fallback(root, discovery, interface, false).await
+    }
+    async fn open_with_fallback(
+        root: &Path,
+        discovery: &SharedMemoryDiscovery,
+        interface: &str,
+        fallback: bool,
+    ) -> Self {
+        Self::open_config(root, discovery, interface, fallback, 1280).await
+    }
+    async fn open_config(
+        root: &Path,
+        discovery: &SharedMemoryDiscovery,
+        interface: &str,
+        fallback: bool,
+        mtu: u32,
+    ) -> Self {
+        Self::open_selected(root, discovery, interface, fallback, mtu, None).await
+    }
+    async fn open_selected(
+        root: &Path,
+        discovery: &SharedMemoryDiscovery,
+        interface: &str,
+        fallback: bool,
+        mtu: u32,
+        default_protocols: Option<Vec<String>>,
+    ) -> Self {
+        let identity =
+            tsunagi::storage::StateStore::open(tsunagi::StoragePaths::under(root).state_db())
+                .unwrap()
+                .load_or_create_device_identity()
+                .unwrap();
+        let transport = AwgTransport::new(identity.endpoint_id(), Ipv4Addr::LOCALHOST);
+        let plugin = AwgPlugin::open(
+            WireguardConfig::new(root.join("awg"))
+                .with_mtu(mtu)
+                .with_reconcile(Duration::from_millis(20), Duration::from_millis(250)),
+            transport.clone(),
+        )
+        .await
+        .unwrap();
+        let loss = Arc::new(AtomicBool::new(false));
+        let faults = Arc::new(FaultTransport {
+            inner: transport.clone(),
+            loss: loss.clone(),
+        });
+        let tuns = MemoryTunFactory::new();
+        let mut config = config_with(root, discovery)
+            .with_interface(Arc::new(tuns.clone()), interface, mtu)
+            .with_plugin(plugin.clone() as Arc<dyn IpPlugin>)
+            .with_custom_transport(AWG_PROTOCOL, faults);
+        let mut quic_plugin = None;
+        if fallback {
+            let quic = tsunagi_wg_quic::WireguardPlugin::open(
+                WireguardConfig::new(root.join("wg-quic"))
+                    .with_mtu(mtu)
+                    .with_reconcile(Duration::from_millis(20), Duration::from_millis(250)),
+            )
+            .await
+            .unwrap();
+            config = config.with_plugin(quic.clone());
+            quic_plugin = Some(quic);
+        }
+        config.default_protocols = default_protocols;
+        let agent = Agent::spawn(config).await.unwrap();
+        let inbound_agent = agent.clone();
+        let inbound_transport = transport.clone();
+        let inbound_loss = loss.clone();
+        tokio::spawn(async move {
+            while let Some(mut link) = inbound_transport.next_inbound().await {
+                link.link = impaired(link.link, inbound_loss.clone());
+                let _ = inbound_agent.install_inbound_link(link).await;
+            }
+        });
+        Self {
+            agent,
+            plugin,
+            quic: quic_plugin,
+            transport,
+            tuns,
+            loss,
+        }
+    }
+    async fn ready(&self, id: NetworkId, n: usize) {
+        wait_until("AWG outer paths and end-to-end handshakes", || async {
+            (self
+                .transport
+                .overview(id)
+                .iter()
+                .filter(|p| !p.closed && !p.path.contains("establishing"))
+                .count()
+                == n
+                && self
+                    .plugin
+                    .overview(id)?
+                    .peers
+                    .iter()
+                    .filter(|p| {
+                        p.tunnel
+                            .as_ref()
+                            .is_some_and(|t| t.health.is_up() && t.path.contains("awg UDP via"))
+                    })
+                    .count()
+                    == n)
+                .then_some(())
+        })
+        .await;
+    }
+    fn tun(&self) -> Arc<MemoryTun> {
+        self.tuns
+            .device(&self.agent.overlay().unwrap().interface)
+            .unwrap()
+    }
+    fn addr(&self, id: NetworkId) -> Ipv4Addr {
+        self.plugin
+            .overview(id)
+            .unwrap()
+            .overlay_address_v4
+            .unwrap()
+    }
+}
+fn packet(source: Ipv4Addr, dest: Ipv4Addr, len: usize) -> Bytes {
+    let mut bytes = vec![0; len];
+    bytes[0] = 0x45;
+    bytes[2..4].copy_from_slice(&(len as u16).to_be_bytes());
+    bytes[8] = 64;
+    bytes[9] = 253;
+    bytes[12..16].copy_from_slice(&source.octets());
+    bytes[16..20].copy_from_slice(&dest.octets());
+    rand::fill(&mut bytes[20..]);
+    Bytes::from(bytes)
+}
+async fn exchange(a: &Node, b: &Node, id: NetworkId) {
+    for len in [64, 1097, 1280] {
+        let bytes = packet(a.addr(id), b.addr(id), len);
+        a.tun().push_from_os(bytes.clone());
+        assert_eq!(
+            tokio::time::timeout(Duration::from_secs(5), b.tun().pop_to_os())
+                .await
+                .unwrap()
+                .unwrap(),
+            bytes
+        );
+    }
+}
+
+#[tokio::test]
+async fn a_control_only_agent_enables_real_data_without_restart_and_cancels_stale_attempts() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let a = Node::open_selected(
+        roots[0].path(),
+        &discovery,
+        "initial-none-a",
+        true,
+        1280,
+        Some(vec![]),
+    )
+    .await;
+    let b = Node::open_selected(
+        roots[1].path(),
+        &discovery,
+        "initial-none-b",
+        true,
+        1280,
+        Some(vec![]),
+    )
+    .await;
+    let (name, secret) = network("enable-after-control-only");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&a.agent, id, 1).await;
+    wait_for_peers(&b.agent, id, 1).await;
+    assert!(a.agent.overlay().is_none());
+    assert!(b.agent.overlay().is_none());
+    for node in [&a, &b] {
+        assert!(
+            node.agent
+                .network_status(id)
+                .await
+                .unwrap()
+                .protocols
+                .is_empty()
+        );
+        node.agent
+            .set_protocols(id, vec!["awg".into()])
+            .await
+            .unwrap();
+        assert!(node.agent.overlay().is_some());
+    }
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&a, &b, id).await;
+    let before = a.agent.overlay().unwrap().interface;
+    // Changes race unfinished opens and prepared plugin callbacks; the last
+    // selection must win even if an earlier open has completed in its queue.
+    for selection in [
+        vec!["wg-quic".into()],
+        vec![],
+        vec!["awg".into()],
+        vec![],
+        vec!["awg".into()],
+    ] {
+        a.agent.set_protocols(id, selection).await.unwrap();
+    }
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&a, &b, id).await;
+    assert_eq!(a.agent.overlay().unwrap().interface, before);
+    for node in [&a, &b] {
+        let status = node.agent.network_status(id).await.unwrap();
+        assert_eq!(status.protocols, ["awg"]);
+        assert_eq!(status.metrics.sessions_established, 1);
+        assert_eq!(status.metrics.disconnects, 0);
+    }
+    a.agent.shutdown().await;
+    b.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn runtime_protocol_selection_replaces_real_tunnels_and_survives_restart() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let a = Node::open_with_fallback(roots[0].path(), &discovery, "runtime-a", true).await;
+    let b = Node::open_with_fallback(roots[1].path(), &discovery, "runtime-b", true).await;
+    let (name, secret) = network("runtime-protocols");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&a, &b, id).await;
+
+    let (other_name, other_secret, other_id) = loop {
+        let (name, secret) = network("unchanged-runtime-network");
+        let other_id = tsunagi::identity::NetworkKeys::derive(&name, &secret).network_id();
+        if tsunagi::state::derived_ipv4_range(id) != tsunagi::state::derived_ipv4_range(other_id) {
+            break (name, secret, other_id);
+        }
+    };
+    a.agent
+        .join_network(&other_name, &other_secret)
+        .await
+        .unwrap();
+    b.agent
+        .join_network(&other_name, &other_secret)
+        .await
+        .unwrap();
+    a.ready(other_id, 1).await;
+    b.ready(other_id, 1).await;
+    let other_before = a
+        .agent
+        .network_status(other_id)
+        .await
+        .unwrap()
+        .metrics
+        .data_links_established;
+
+    for node in [&a, &b] {
+        let before = node.agent.network_status(id).await.unwrap().protocols;
+        for invalid in [vec!["unknown".into()], vec!["awg".into(), "awg".into()]] {
+            assert!(node.agent.set_protocols(id, invalid).await.is_err());
+            assert_eq!(
+                node.agent.network_status(id).await.unwrap().protocols,
+                before
+            );
+        }
+    }
+    // Exercise updates initiated by either endpoint-id role, including an
+    // empty set. Membership and the control session stay alive throughout.
+    for node in [&a, &b] {
+        node.agent.set_protocols(id, Vec::new()).await.unwrap();
+        wait_until("peer accepts control-only selection", || async {
+            let other = if node.agent.endpoint_id() == a.agent.endpoint_id() {
+                &b
+            } else {
+                &a
+            };
+            let status = other.agent.network_status(id).await.ok()?;
+            (status.peers.len() == 1
+                && status.peers[0].protocols.is_empty()
+                && matches!(status.peers[0].data_path, tsunagi::agent::DataPath::None))
+            .then_some(())
+        })
+        .await;
+        assert_eq!(node.agent.network_status(id).await.unwrap().peers.len(), 1);
+        node.agent
+            .set_protocols(id, vec!["wg-quic".into(), "awg".into()])
+            .await
+            .unwrap();
+    }
+    a.agent
+        .set_protocols(id, vec!["wg-quic".into()])
+        .await
+        .unwrap();
+    wait_until("real QUIC replaces AWG", || async {
+        let a_status = a.agent.network_status(id).await.ok()?;
+        let b_status = b.agent.network_status(id).await.ok()?;
+        (a_status.peers[0].protocols == ["wg-quic"]
+            && b_status.peers[0].protocols == ["wg-quic"]
+            && !matches!(a_status.peers[0].data_path, tsunagi::agent::DataPath::None)
+            && !matches!(b_status.peers[0].data_path, tsunagi::agent::DataPath::None)
+            && a.transport.overview(id).iter().all(|p| p.closed)
+            && b.transport.overview(id).iter().all(|p| p.closed))
+        .then_some(())
+    })
+    .await;
+    // Get addresses from the system allocation: the disabled AWG plugin no
+    // longer has a live view of this network.
+    for (source, dest) in [(&a, &b), (&b, &a)] {
+        wait_until("QUIC end-to-end handshake", || async {
+            source
+                .quic
+                .as_ref()?
+                .overview(id)?
+                .peers
+                .iter()
+                .any(|p| p.tunnel.as_ref().is_some_and(|t| t.health.is_up()))
+                .then_some(())
+        })
+        .await;
+        let addr = |n: &Node| {
+            n.quic
+                .as_ref()
+                .unwrap()
+                .overview(id)
+                .unwrap()
+                .overlay_address_v4
+                .unwrap()
+        };
+        let bytes = packet(addr(source), addr(dest), 1280);
+        source.tun().push_from_os(bytes.clone());
+        assert_eq!(
+            tokio::time::timeout(Duration::from_secs(5), dest.tun().pop_to_os())
+                .await
+                .unwrap()
+                .unwrap(),
+            bytes
+        );
+    }
+    assert_eq!(
+        a.agent
+            .network_status(other_id)
+            .await
+            .unwrap()
+            .metrics
+            .data_links_established,
+        other_before
+    );
+    exchange(&a, &b, other_id).await;
+    a.agent.shutdown().await;
+    let restarted = Node::open_with_fallback(roots[0].path(), &discovery, "runtime-a", true).await;
+    wait_for_peers(&restarted.agent, id, 1).await;
+    assert_eq!(
+        restarted.agent.network_status(id).await.unwrap().protocols,
+        ["wg-quic"]
+    );
+    assert_eq!(
+        b.agent.network_status(other_id).await.unwrap().protocols,
+        ["awg", "wg-quic"]
+    );
+    restarted.ready(other_id, 1).await;
+    b.ready(other_id, 1).await;
+    exchange(&restarted, &b, other_id).await;
+    restarted
+        .agent
+        .set_protocols(id, vec!["awg".into()])
+        .await
+        .unwrap();
+    restarted.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&restarted, &b, id).await;
+    restarted.agent.shutdown().await;
+    b.agent.shutdown().await;
+}
+#[tokio::test]
+async fn real_pairs_carry_full_size_packets_and_restart_with_fresh_profiles() {
+    tsunagi::testing::init_tracing();
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [
+        TempDir::new().unwrap(),
+        TempDir::new().unwrap(),
+        TempDir::new().unwrap(),
+    ];
+    let a = Node::open(roots[0].path(), &discovery, "awga").await;
+    let b = Node::open(roots[1].path(), &discovery, "awgb").await;
+    let c = Node::open(roots[2].path(), &discovery, "awgc").await;
+    let (name, secret) = network("awg-pairs");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    c.agent.join_network(&name, &secret).await.unwrap();
+    for node in [&a, &b, &c] {
+        wait_for_peers(&node.agent, id, 2).await;
+        node.ready(id, 2).await;
+    }
+    let ids: std::collections::HashSet<_> = [&a, &b, &c]
+        .into_iter()
+        .flat_map(|n| n.transport.overview(id).into_iter().map(|p| p.attempt))
+        .collect();
+    assert_eq!(
+        ids.len(),
+        3,
+        "one independent profile for each of three pairs"
+    );
+    exchange(&a, &b, id).await;
+    exchange(&b, &a, id).await;
+    exchange(&a, &c, id).await;
+    let old_ports: Vec<_> = a
+        .transport
+        .overview(id)
+        .into_iter()
+        .filter_map(|p| p.local_port)
+        .collect();
+    a.agent.shutdown().await;
+    for port in old_ports {
+        wait_until("old pair socket released", || async {
+            tokio::net::UdpSocket::bind((Ipv4Addr::LOCALHOST, port))
+                .await
+                .ok()
+        })
+        .await;
+    }
+    let restarted = Node::open(roots[0].path(), &discovery, "awga").await;
+    restarted.ready(id, 2).await;
+    assert!(
+        restarted
+            .transport
+            .overview(id)
+            .iter()
+            .all(|p| !ids.contains(&p.attempt))
+    );
+    exchange(&restarted, &b, id).await;
+    exchange(&c, &restarted, id).await;
+    restarted.agent.shutdown().await;
+    b.agent.shutdown().await;
+    c.agent.shutdown().await;
+    assert!(
+        tokio::time::timeout(Duration::from_secs(1), restarted.transport.next_inbound())
+            .await
+            .unwrap()
+            .is_none()
+    );
+}
+
+// Drops datagrams at the link boundary. Membership, control, sockets, crypto,
+// addressing and storage remain real. This substitutes a lossy physical path.
+#[derive(Debug)]
+struct FaultTransport {
+    inner: Arc<AwgTransport>,
+    loss: Arc<AtomicBool>,
+}
+impl PacketTransport for FaultTransport {
+    fn name(&self) -> &str {
+        AWG_PROTOCOL
+    }
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+    fn attach_control(&self, sender: Arc<dyn ControlSender>) {
+        self.inner.attach_control(sender);
+    }
+    fn on_control(
+        &self,
+        network: NetworkId,
+        peer: iroh::EndpointId,
+        protocol: &str,
+        bytes: &[u8],
+    ) -> Result<(), TransportError> {
+        self.inner.on_control(network, peer, protocol, bytes)
+    }
+    fn open<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: iroh::EndpointId,
+        protocol: &'a str,
+    ) -> BoxFuture<'a, Result<SharedLink, TransportError>> {
+        Box::pin(async move {
+            Ok(impaired(
+                self.inner.open(network, peer, protocol).await?,
+                self.loss.clone(),
+            ))
+        })
+    }
+}
+fn impaired(raw: SharedLink, loss: Arc<AtomicBool>) -> SharedLink {
+    Arc::new(FaultLink {
+        raw,
+        loss,
+        sequence: AtomicU64::new(0),
+    })
+}
+#[derive(Debug)]
+struct FaultLink {
+    raw: SharedLink,
+    loss: Arc<AtomicBool>,
+    sequence: AtomicU64,
+}
+impl PacketLink for FaultLink {
+    fn network(&self) -> NetworkId {
+        self.raw.network()
+    }
+    fn peer(&self) -> iroh::EndpointId {
+        self.raw.peer()
+    }
+    fn max_datagram_size(&self) -> usize {
+        self.raw.max_datagram_size()
+    }
+    fn send(&self, bytes: Bytes) -> Result<(), TransportError> {
+        if self.loss.load(Ordering::Relaxed)
+            && !self
+                .sequence
+                .fetch_add(1, Ordering::Relaxed)
+                .is_multiple_of(5)
+        {
+            Ok(())
+        } else {
+            self.raw.send(bytes)
+        }
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        self.raw.recv()
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        self.raw.closed()
+    }
+    fn is_closed(&self) -> bool {
+        self.raw.is_closed()
+    }
+    fn close(&self) {
+        self.raw.close();
+    }
+    fn path_description(&self) -> String {
+        self.raw.path_description()
+    }
+}
+#[tokio::test]
+async fn persistent_runtime_loss_retires_the_pair_and_negotiates_fresh_parameters() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let a = Node::open(roots[0].path(), &discovery, "lossa").await;
+    let b = Node::open(roots[1].path(), &discovery, "lossb").await;
+    let (name, secret) = network("awg-loss");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&a, &b, id).await;
+    let old = a.transport.overview(id)[0].attempt;
+    let mut events = a.agent.subscribe();
+    a.loss.store(true, Ordering::Relaxed);
+    tsunagi::testing::wait_event(&mut events, |event| match event {
+        tsunagi::Event::DataLinkDown {
+            network, reason, ..
+        } if *network == id && reason.contains("delivery checks") => Some(()),
+        _ => None,
+    })
+    .await;
+    a.loss.store(false, Ordering::Relaxed);
+    wait_until("fresh AWG profile after runtime loss", || async {
+        a.transport
+            .overview(id)
+            .iter()
+            .any(|p| p.attempt != old && !p.closed && !p.path.contains("establishing"))
+            .then_some(())
+    })
+    .await;
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    let initiator = if a.agent.endpoint_id().as_bytes() < b.agent.endpoint_id().as_bytes() {
+        &a
+    } else {
+        &b
+    };
+    let received = initiator.plugin.overview(id).unwrap().peers[0]
+        .tunnel
+        .as_ref()
+        .unwrap()
+        .stats
+        .rx_packets;
+    let attempt = initiator.transport.overview(id)[0].attempt;
+    wait_until(
+        "fresh pair delivery checks answered after installation",
+        || async {
+            let view = initiator.plugin.overview(id)?;
+            (view.peers.first()?.tunnel.as_ref()?.stats.rx_packets >= received + 3).then_some(())
+        },
+    )
+    .await;
+    assert_eq!(initiator.transport.overview(id)[0].attempt, attempt);
+    wait_for_peers(&a.agent, id, 1).await;
+    exchange(&a, &b, id).await;
+    exchange(&b, &a, id).await;
+    a.agent.shutdown().await;
+    b.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn runtime_loss_falls_back_to_real_iroh_without_interrupting_control_or_ip() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let a = Node::open_with_fallback(roots[0].path(), &discovery, "fallba", true).await;
+    let b = Node::open_with_fallback(roots[1].path(), &discovery, "fallbb", true).await;
+    let (name, secret) = network("awg-fallback");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    exchange(&a, &b, id).await;
+    let mut events = a.agent.subscribe();
+    a.loss.store(true, Ordering::Relaxed);
+    tsunagi::testing::wait_event(&mut events, |event| match event {
+        tsunagi::Event::DataLinkDown {
+            network, reason, ..
+        } if *network == id && reason.contains("delivery checks") => Some(()),
+        _ => None,
+    })
+    .await;
+    for node in [&a, &b] {
+        wait_until("fallback carrier is wg-quic",||async {
+            let status=node.agent.status().await.ok()?;
+            let network=status.networks.iter().find(|n|n.network_id==id)?;
+            (network.members.iter().any(|m|matches!(&m.data_path,tsunagi::agent::DataPath::Direct{transport} if transport=="wg-quic"))).then_some(())
+        }).await;
+        wait_for_peers(&node.agent, id, 1).await;
+    }
+    exchange(&a, &b, id).await;
+    exchange(&b, &a, id).await;
+    a.agent.shutdown().await;
+    b.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn opaque_pair_carrier_fragments_its_full_advertised_capacity() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let mut nodes = Vec::new();
+    for root in &roots {
+        let identity = tsunagi::storage::StateStore::open(
+            tsunagi::StoragePaths::under(root.path()).state_db(),
+        )
+        .unwrap()
+        .load_or_create_device_identity()
+        .unwrap();
+        let transport = AwgTransport::new(identity.endpoint_id(), Ipv4Addr::LOCALHOST);
+        let agent = Agent::spawn(
+            config_with(root.path(), &discovery)
+                .with_custom_transport(AWG_PROTOCOL, transport.clone()),
+        )
+        .await
+        .unwrap();
+        nodes.push((agent, transport));
+    }
+    let (name, secret) = network("awg-capacity");
+    let id = nodes[0].0.join_network(&name, &secret).await.unwrap();
+    nodes[1].0.join_network(&name, &secret).await.unwrap();
+    for (i, (agent, transport)) in nodes.iter().enumerate() {
+        wait_for_peers(agent, id, 1).await;
+        // This fixture tests only the transport contract: register the member
+        // after the real core has authenticated it, without an IP plugin/TUN.
+        transport.allow_peer(id, nodes[1 - i].0.endpoint_id());
+    }
+    nodes.sort_by_key(|(agent, _)| *agent.endpoint_id().as_bytes());
+    let outgoing = nodes[0]
+        .1
+        .open(id, nodes[1].0.endpoint_id(), AWG_PROTOCOL)
+        .await
+        .unwrap();
+    let incoming = tokio::time::timeout(Duration::from_secs(5), nodes[1].1.next_inbound())
+        .await
+        .unwrap()
+        .unwrap()
+        .link;
+    for (from, to) in [(&outgoing, &incoming), (&incoming, &outgoing)] {
+        let mut bytes = vec![0; from.max_datagram_size()];
+        rand::fill(&mut bytes[..]);
+        let bytes = Bytes::from(bytes);
+        from.send(bytes.clone()).unwrap();
+        assert_eq!(
+            tokio::time::timeout(Duration::from_secs(5), to.recv())
+                .await
+                .unwrap()
+                .unwrap(),
+            bytes
+        );
+        assert!(matches!(
+            from.send(Bytes::from(vec![0; from.max_datagram_size() + 1])),
+            Err(TransportError::TooLarge { .. })
+        ));
+    }
+    outgoing.close();
+    incoming.close();
+    for (agent, transport) in nodes {
+        transport.shutdown();
+        agent.shutdown().await;
+    }
+}
+
+#[tokio::test]
+async fn encrypted_health_checks_fit_a_smaller_peers_configured_interface_mtu() {
+    let discovery = SharedMemoryDiscovery::new();
+    let roots = [TempDir::new().unwrap(), TempDir::new().unwrap()];
+    let a = Node::open_config(roots[0].path(), &discovery, "smalla", false, 1280).await;
+    let b = Node::open_config(roots[1].path(), &discovery, "smallb", false, 576).await;
+    let (name, secret) = network("awg-small-mtu");
+    let id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    a.ready(id, 1).await;
+    b.ready(id, 1).await;
+    let bytes = packet(a.addr(id), b.addr(id), 576);
+    a.tun().push_from_os(bytes.clone());
+    assert_eq!(
+        tokio::time::timeout(Duration::from_secs(5), b.tun().pop_to_os())
+            .await
+            .unwrap()
+            .unwrap(),
+        bytes
+    );
+    wait_until(
+        "encrypted delivery checks answered with a small MTU",
+        || async {
+            let view = a.plugin.overview(id)?;
+            let peer = view.peers.first()?.tunnel.as_ref()?;
+            (peer.stats.rx_packets >= 3).then_some(())
+        },
+    )
+    .await;
+    for node in [&a, &b] {
+        assert_eq!(
+            node.plugin.overview(id).unwrap().peers[0]
+                .tunnel
+                .as_ref()
+                .unwrap()
+                .stats
+                .dropped_oversize,
+            0
+        );
+    }
+    a.agent.shutdown().await;
+    b.agent.shutdown().await;
+}

crates/tsunagi-awg/tests/profiles.rs

@@ -0,0 +1,36 @@
+//! Fresh and bounded masking profiles, and rejection of foreign/truncated wire packets.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+use tsunagi_awg::profile::PairProfile;
+
+#[test]
+fn profiles_are_fresh_bounded_and_redacted() {
+    let profiles: Vec<_> = (0..128).map(|_| PairProfile::generate()).collect();
+    for (i, profile) in profiles.iter().enumerate() {
+        profile.validate().unwrap();
+        assert!(!format!("{profile:?}").contains(&format!("{:?}", profile.header_key())));
+        for other in &profiles[..i] {
+            assert_ne!(profile.header_key(), other.header_key());
+            assert_ne!(profile.id(), other.id());
+        }
+        for packet in profile.signature_packets() {
+            assert!(packet.len() < 1200);
+        }
+    }
+}
+
+#[test]
+fn masked_packets_round_trip_and_reject_other_profiles_and_truncation() {
+    let profile = PairProfile::generate();
+    let other = PairProfile::generate();
+    for (kind, size) in [(1, 148), (2, 92), (3, 64), (4, 800)] {
+        let mut message = vec![0x7b; size];
+        message[..4].copy_from_slice(&(kind as u32).to_le_bytes());
+        let wire = profile.encode(&message).unwrap();
+        assert_ne!(&wire[..4], &message[..4]);
+        assert_eq!(profile.decode(&wire).unwrap(), message);
+        assert!(other.decode(&wire).is_none());
+        for length in 0..44 {
+            assert!(profile.decode(&wire[..length]).is_none());
+        }
+    }
+}

crates/tsunagi-core/Cargo.toml

@@ -0,0 +1,113 @@
+[package]
+name = "tsunagi-core"
+version = "0.1.0"
+description = "Small private mesh networks: persistent agent identity, deterministic network spaces, an iroh control plane, signed state and a local DNS view."
+readme = "../../README.md"
+keywords = ["mesh", "p2p", "iroh", "networking"]
+categories = ["network-programming"]
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[lib]
+name = "tsunagi"
+
+[features]
+default = []
+# A real TUN device, so a plugin can carry actual IP traffic. On Linux it
+# needs CAP_NET_ADMIN; on Windows the process must run elevated and
+# `wintun.dll` must be reachable. Without either the in-memory device serves
+# tests. The platform-specific dependencies are declared per target, so
+# enabling this only pulls what the target it is built for actually uses.
+tun-device = [
+    "dep:tun",
+    "dep:rtnetlink",
+    "dep:caps",
+    "dep:futures-util",
+]
+# Telling the operating system where to send its DNS questions. systemd-resolved
+# on Linux, the Name Resolution Policy Table on Windows; the zone and the
+# server work without it either way.
+dns-publish = ["dep:zbus"]
+# The test harness in `testing`, for this crate's own tests and for a
+# protocol crate's. Not on by default: it is only of use to a test.
+testing = ["dep:tempfile", "dep:tracing-subscriber"]
+
+[dependencies]
+arc-swap = "1.9"
+iroh.workspace = true
+# `net` for the local control interface: a Unix socket on Unix, a named pipe
+# on Windows. Both live behind the same `net` feature.
+tokio = { workspace = true, features = ["net"] }
+tokio-stream = "0.1"
+serde.workspace = true
+postcard.workspace = true
+bytes.workspace = true
+thiserror.workspace = true
+tracing.workspace = true
+hex.workspace = true
+data-encoding.workspace = true
+rusqlite = { version = "0.40", features = ["bundled"] }
+hkdf = "0.13"
+hmac = "0.13"
+sha2 = "0.11"
+subtle = "2.6"
+zeroize = { version = "1.9", features = ["derive"] }
+rand = "0.10"
+# Already in the tree through iroh. A packet codec, not a DNS server: the
+# zone logic is ours and a full server framework would be a large dependency
+# for answering A records from memory.
+simple-dns = "0.12"
+# The IANA top-level domain list, compiled in: one function, no
+# dependencies, no network. Used only to warn that a zone name shadows a
+# real public domain, never to refuse one.
+tld = "2.40"
+fs4 = { version = "1.1", features = ["sync"] }
+directories = "6.0"
+# The real system hostname, without a libc call of our own: this crate is
+# forbidden `unsafe` and will not make one.
+gethostname = "1.1"
+netwatch = "0.19.3"
+netdev = "0.46"
+mainline = "8.0"
+futures-lite = "2.6"
+tun = { version = "0.8", features = ["async"], optional = true }
+# Only for the `testing` harness, which a protocol crate's tests use too.
+tempfile = { version = "3.24", optional = true }
+tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
+
+# Linux-only interface provisioning. `rtnetlink` configures the interface in
+# process, so no `ip` invocation is ever needed; `caps` keeps CAP_NET_ADMIN
+# out of the effective set except during the moments it is used.
+[target.'cfg(target_os = "linux")'.dependencies]
+rtnetlink = { version = "0.23", optional = true }
+# Pure Rust D-Bus, no libdbus to link against. `tokio` rather than the
+# default reactor, because the agent brings its own.
+zbus = { version = "5.19", default-features = false, features = ["tokio"], optional = true }
+caps = { version = "0.5", optional = true }
+futures-util = { version = "0.3", default-features = false, optional = true }
+
+# Windows-only interface provisioning. The Wintun adapter itself is created
+# and driven through the `tun` dependency, whose safe wrapper is the only way
+# in — loading `wintun.dll` is an `unsafe` call this crate forbids. `netdev`
+# (already in the tree through `netwatch`) reads the interface table through a
+# safe wrapper, which is how the provisioner observes an adapter by name
+# before deciding what to change. Creating an adapter needs the process to be
+# elevated and `wintun.dll` to be reachable at run time.
+[target.'cfg(target_os = "windows")'.dependencies]
+windows-sys = { version = "0.61", features = [
+    "Win32_Security",
+    "Win32_Security_Authorization",
+    "Win32_Foundation",
+] }
+
+[dev-dependencies]
+# Its own tests use the harness it publishes.
+tsunagi-core = { path = ".", features = ["testing"] }
+tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros", "process", "test-util"] }
+tempfile.workspace = true
+tracing-subscriber.workspace = true
+
+[lints]
+workspace = true

crates/tsunagi-core/src/agent/events.rs

@@ -129,6 +129,19 @@ pub enum Event {
         /// Why it is not up.
         reason: String,
     },
+    /// The name this agent claimed was claimed first by another member.
+    ///
+    /// Nothing stops working: this agent keeps its address and every
+    /// connection. It just is not found by that name until it is renamed or
+    /// the member that holds the name lets it go.
+    HostnameTaken {
+        /// The network.
+        network: NetworkId,
+        /// The name that is taken.
+        name: String,
+        /// The member that holds it.
+        holder: EndpointId,
+    },
     /// An IP plugin reported an error. Never fatal.
     PluginError {
         /// The network the call was scoped to.

crates/tsunagi-core/src/agent/session.rs

@@ -70,20 +70,24 @@ pub(crate) struct Session {
     pub(crate) conn: Connection,
     /// Already encoded frame payloads, so the runtime can account for the exact
     /// number of control bytes it queues.
-    pub(crate) outbound: mpsc::Sender<Vec<u8>>,
+    pub(crate) outbound: mpsc::Sender<zeroize::Zeroizing<Vec<u8>>>,
     pub(crate) hostname: Option<String>,
     pub(crate) capabilities: Vec<PluginCapability>,
+    pub(crate) broadcast: bool,
+    /// The peer offers itself as an exit node in this network.
+    pub(crate) exit_node: bool,
     pub(crate) messages_sent: u64,
     pub(crate) messages_received: u64,
     pub(crate) bytes_sent: u64,
     pub(crate) bytes_received: u64,
     reader: JoinHandle<()>,
     writer: JoinHandle<()>,
+    path_monitor: JoinHandle<()>,
     shutdown: Shutdown,
 }
 
 impl Session {
-    /// Signals both tasks to stop and waits for them, with a bounded grace
+    /// Signals all tasks to stop and waits for them, with a bounded grace
     /// period.
     ///
     /// A peer that stops reading must not be able to hold up shutdown, so the
@@ -93,6 +97,7 @@ impl Session {
             conn,
             reader,
             writer,
+            path_monitor,
             shutdown,
             ..
         } = self;
@@ -102,23 +107,27 @@ impl Session {
 
         let reader_abort = reader.abort_handle();
         let writer_abort = writer.abort_handle();
+        let path_abort = path_monitor.abort_handle();
         let joined = tokio::time::timeout(STOP_GRACE, async move {
             let _ = reader.await;
             let _ = writer.await;
+            let _ = path_monitor.await;
         })
         .await;
         if joined.is_err() {
             tracing::debug!("session tasks did not wind down in time; aborting them");
             reader_abort.abort();
             writer_abort.abort();
+            path_abort.abort();
         }
     }
 
-    /// Aborts both tasks without waiting. Used on replacement.
+    /// Aborts all tasks without waiting. Used on replacement.
     pub(crate) fn abort(&self) {
         self.shutdown.trigger();
         self.reader.abort();
         self.writer.abort();
+        self.path_monitor.abort();
     }
 }
 
@@ -154,6 +163,52 @@ pub(crate) fn spawn(
         parent_shutdown.clone(),
     ));
 
+    let path_conn = conn.clone();
+    let path_shutdown = shutdown.clone();
+    let path_parent = parent_shutdown.clone();
+    let path_monitor = tokio::spawn(async move {
+        use tokio_stream::StreamExt;
+        let mut events = path_conn.path_events();
+        loop {
+            tokio::select! {
+                biased;
+                _ = path_shutdown.wait() => break,
+                _ = path_parent.wait() => break,
+                ev = events.next() => {
+                    match ev {
+                        Some(iroh::endpoint::PathEvent::Selected { remote_addr, local_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                local = ?local_addr,
+                                "Connection path selected (direct / relay / hole punch switch)"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Opened { remote_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                "Network path opened to peer"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Closed { remote_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                "Network path closed for peer"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Lagged { missed, .. }) => {
+                            tracing::trace!(peer = %peer.fmt_short(), missed, "Path events lagged");
+                        }
+                        Some(_) => {}
+                        None => break,
+                    }
+                }
+            }
+        }
+    });
+
     let reader = tokio::spawn(reader_task(
         id,
         network_id,
@@ -174,19 +229,22 @@ pub(crate) fn spawn(
         outbound: outbound_tx,
         hostname: None,
         capabilities: Vec::new(),
+        broadcast: false,
+        exit_node: false,
         messages_sent: 0,
         messages_received: 0,
         bytes_sent: 0,
         bytes_received: 0,
         reader,
         writer,
+        path_monitor,
         shutdown,
     }
 }
 
 async fn writer_task(
     mut send: SendStream,
-    mut outbound: mpsc::Receiver<Vec<u8>>,
+    mut outbound: mpsc::Receiver<zeroize::Zeroizing<Vec<u8>>>,
     limits: Arc<Limits>,
     shutdown: Shutdown,
     parent: Shutdown,
@@ -273,6 +331,7 @@ async fn reader_task(
             }
         };
 
+        let payload = zeroize::Zeroizing::new(payload);
         let bytes = payload.len();
         let envelope: Envelope = match decode(&payload) {
             Ok(envelope) => envelope,

crates/tsunagi-core/src/agent/shutdown.rs

@@ -51,3 +51,25 @@ impl Shutdown {
         // The sender is gone, which for our purposes means "stop".
     }
 }
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn waiting_resolves_once_triggered_and_stays_resolved() {
+        let shutdown = Shutdown::new();
+        assert!(!shutdown.is_triggered());
+        let waiter = shutdown.clone();
+        let waiting = tokio::spawn(async move { waiter.wait().await });
+        shutdown.trigger();
+        waiting.await.unwrap();
+        assert!(shutdown.is_triggered());
+        // Triggering twice is not an error, and a token cloned afterwards
+        // sees the trigger that already happened.
+        shutdown.trigger();
+        shutdown.clone().wait().await;
+    }
+}

crates/tsunagi-core/src/agent/status.rs

@@ -0,0 +1,309 @@
+//! Status snapshots.
+//!
+//! Metrics are reported at the level they actually belong to. Values that
+//! genuinely cannot be attributed to a single network — everything the iroh
+//! endpoint aggregates, for instance — stay at the endpoint level rather than
+//! being split between networks with invented precision.
+
+use std::time::Duration;
+
+use iroh::{EndpointAddr, EndpointId};
+
+use crate::dataplane::PluginCapability;
+use crate::discovery::CandidateSource;
+use crate::identity::{NetworkDescriptor, NetworkId, NetworkName};
+use crate::net::{ConnectionCounters, PathAddr, PathInfo, TransportKind};
+use crate::proto::handshake::Role;
+use crate::storage::CacheOutcome;
+
+/// Whether a configured network is running locally.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum NetworkState {
+    /// Configured and running.
+    Active,
+    /// Configured but not running.
+    Inactive,
+}
+
+/// An unverified candidate as seen by a network runtime.
+#[derive(Debug, Clone)]
+pub struct CandidateStatus {
+    /// Candidate endpoint id.
+    pub endpoint_id: EndpointId,
+    /// Where it came from.
+    pub source: CandidateSource,
+    /// Consecutive failed dial attempts since the last success.
+    pub consecutive_failures: u32,
+}
+
+/// The overlay interface, as the agent sees it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct OverlayStatus {
+    /// The interface name the operating system gave.
+    pub interface: String,
+    /// Whether that interface exists on the host.
+    ///
+    /// `false` with an in-memory device: the name is real to the agent and
+    /// to nothing else. Reported, because everything that would configure
+    /// the operating system for this interface must not when it is this.
+    pub on_host: bool,
+    /// Its MTU.
+    pub mtu: u32,
+    /// The addresses it should be carrying.
+    pub addresses: Vec<crate::overlay::Cidr>,
+    /// What has happened on it.
+    pub counters: crate::overlay::Counters,
+    /// The route and firewall allowance installed for LAN broadcast, when
+    /// some network has broadcast on and the interface is on the host.
+    /// `None` is "none installed", not "unknown".
+    pub broadcast_rules: Option<crate::overlay::BroadcastRulesReport>,
+}
+
+/// How packets reach a member right now.
+///
+/// Being reachable is one thing and being reachable well is another, and
+/// the second is never a reason to call a member away: a member that can only
+/// be reached through somebody else is online, over a longer path.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum DataPath {
+    /// Nothing carries packets to it yet. The member is connected for
+    /// control only, which is the state of a data path being set up.
+    None,
+    /// A direct link, of the named kind of transport.
+    Direct {
+        /// The kind of transport carrying it, e.g. `wg`.
+        transport: String,
+    },
+    /// Through other members.
+    Relayed {
+        /// Number of links on the way, at least two.
+        hops: u8,
+        /// The member the first of them goes to.
+        via: EndpointId,
+    },
+}
+
+/// A name that a member claimed and somebody else holds.
+///
+/// A name belongs to whoever claimed it first. The member that claimed it
+/// later is still a member and still reachable by its address; it is just
+/// not found by that name, on any member of the network.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct HostnameConflict {
+    /// The name that is taken.
+    pub name: String,
+    /// The member that holds it.
+    pub holder: EndpointId,
+}
+
+/// A member the signed state knows about, connected or not.
+///
+/// This is the durable roster: it comes from signed records, so a member that
+/// went away last month is still here. That is what makes it possible to say
+/// "this peer is offline" rather than only "nobody is connected".
+///
+/// It is not a complete membership list, and cannot be. A member is in signed
+/// state once it has claimed something — an address, a name, or both — so a
+/// member that has joined but never published is visible only while it is
+/// connected.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct MemberStatus {
+    /// Current signed claim version, used to bind cleanup confirmation.
+    pub record_version: u64,
+    /// The member's device identity.
+    pub endpoint_id: EndpointId,
+    /// The IPv4 overlay address it claimed and signed for.
+    pub overlay_address_v4: Option<std::net::Ipv4Addr>,
+    /// The name it claimed, when it holds that name uncontested.
+    ///
+    /// Signed, so it is still known while the member is away — which is what
+    /// lets an absent member be named rather than shown as a bare id.
+    pub hostname: Option<String>,
+    /// The last name this agent saw it announce, whether or not it is here.
+    ///
+    /// Not signed, so it is a convenience for naming a member that is away
+    /// and never a claim: the zone is served from `hostname`.
+    pub last_hostname: Option<String>,
+    /// The name it claimed, when somebody else claimed it first.
+    pub hostname_conflict: Option<HostnameConflict>,
+    /// Whether this agent can reach it now, by any path: a live control
+    /// session, or a data path through other members. The one answer to "is
+    /// it online" for every client, so that none of them has to work it out
+    /// from a different part of the picture. This agent counts as online.
+    pub online: bool,
+    /// How packets reach it. [`DataPath::None`] while it is offline.
+    pub data_path: DataPath,
+}
+
+/// Status of one authenticated session.
+#[derive(Debug, Clone)]
+pub struct PeerStatus {
+    /// This authenticated peer accepts network broadcasts.
+    pub broadcast: bool,
+    /// This authenticated peer offers itself as an exit node.
+    pub exit_node: bool,
+    /// Authenticated endpoint id.
+    pub endpoint_id: EndpointId,
+    /// Which side this agent played in the handshake.
+    pub role: Role,
+    /// Hostname the peer announced, if it has announced one yet.
+    ///
+    /// A mutable binding, not an identity.
+    pub hostname: Option<String>,
+    /// Protocols agreed with this peer: both sides have them at the same
+    /// version.
+    ///
+    /// Empty means no data plane with that peer, which leaves the control
+    /// plane working — messages and state still flow.
+    pub protocols: Vec<String>,
+    /// Capabilities the peer announced. Payloads stay opaque.
+    pub capabilities: Vec<PluginCapability>,
+    /// How long the session has been up.
+    pub connected_for: Duration,
+    /// Verified paths of the underlying connection.
+    pub paths: Vec<PathInfo>,
+    /// How the connection currently reaches the peer.
+    pub transport: TransportKind,
+    /// RTT of the selected path, when iroh reported one.
+    pub rtt: Option<Duration>,
+    /// Per-connection counters.
+    pub connection: ConnectionCounters,
+    /// Control messages sent on this session.
+    pub control_messages_sent: u64,
+    /// Control messages received on this session.
+    pub control_messages_received: u64,
+    /// Control payload bytes queued for this session.
+    pub control_bytes_sent: u64,
+    /// Control payload bytes read from this session.
+    pub control_bytes_received: u64,
+    /// How packets reach this peer right now.
+    pub data_path: DataPath,
+}
+
+/// Counters scoped to one logical network.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct NetworkMetrics {
+    /// Outbound dial attempts started.
+    pub dial_attempts: u64,
+    /// Outbound dials that failed before or during the handshake.
+    pub dial_failures: u64,
+    /// Handshakes rejected, in either direction.
+    pub handshake_failures: u64,
+    /// Sessions that reached the authenticated state.
+    pub sessions_established: u64,
+    /// Sessions that ended.
+    pub disconnects: u64,
+    /// Control messages sent in this network.
+    pub control_messages_sent: u64,
+    /// Control messages received in this network.
+    pub control_messages_received: u64,
+    /// Control bytes sent in this network, payload only.
+    pub control_bytes_sent: u64,
+    /// Control bytes received in this network, payload only.
+    pub control_bytes_received: u64,
+    /// Messages or sessions rejected for protocol violations.
+    pub protocol_violations: u64,
+    /// Errors reported by IP plugins. Never fatal.
+    pub plugin_errors: u64,
+    /// Data plane links that were established.
+    pub data_links_established: u64,
+    /// Attempts to open a data plane link that failed.
+    pub data_link_failures: u64,
+}
+
+/// An active network preventing another network from starting locally.
+#[derive(Debug, Clone)]
+pub struct RangeConflict {
+    /// The blocking network's public identifier.
+    pub network_id: NetworkId,
+    /// Its configured name.
+    pub name: NetworkName,
+    /// Its overlapping overlay range.
+    pub range: crate::state::Ipv4Range,
+}
+
+/// Status of one network.
+#[derive(Debug, Clone)]
+pub struct NetworkStatus {
+    /// Shared service configuration with this client's routing observations.
+    pub vips: Vec<crate::vip::VipStatus>,
+    /// Enabled local data-plane protocols, in preference order.
+    pub protocols: Vec<String>,
+    /// Whether this network is requested to start or resume automatically.
+    pub auto_start: bool,
+    /// Local broadcast participation in this network.
+    pub broadcast: bool,
+    /// This agent offers itself as an exit node in this network.
+    pub exit_offer: bool,
+    /// The member this device sends all its internet traffic through.
+    pub exit_via: Option<EndpointId>,
+    /// Whether that member is connected and still offering to be one.
+    pub exit_via_online: bool,
+    /// What the host rules for the exit settings are doing.
+    pub exit_rules: crate::overlay::ExitRulesReport,
+    /// Immutable deterministic description of the network space.
+    pub descriptor: NetworkDescriptor,
+    /// Network name, for convenience.
+    pub name: NetworkName,
+    /// Public network identifier.
+    pub network_id: NetworkId,
+    /// Whether the network is running locally.
+    pub state: NetworkState,
+    /// Authenticated sessions.
+    pub peers: Vec<PeerStatus>,
+    /// Unverified candidates currently known. Not peers.
+    pub candidates: Vec<CandidateStatus>,
+    /// Members the signed state knows about, whether connected or not.
+    pub members: Vec<MemberStatus>,
+    /// This agent's own name, when another member claimed it first and so
+    /// holds it. This agent keeps working by address; nobody finds it by
+    /// this name until it is renamed or the other member lets it go.
+    pub hostname_conflict: Option<HostnameConflict>,
+    /// The overlay range this network uses, once it has one.
+    pub range: Option<crate::state::Ipv4Range>,
+    /// The first overlapping active range; `range_conflicts` identifies all blockers.
+    pub range_conflict: Option<crate::state::Ipv4Range>,
+    /// Active networks that would be suspended by switching to this network.
+    pub range_conflicts: Vec<RangeConflict>,
+    /// Per-network counters.
+    pub metrics: NetworkMetrics,
+    /// What has gone through a peer in the middle, in both directions.
+    pub relay: crate::dataplane::relay::RelayCounters,
+}
+
+impl NetworkStatus {
+    /// Endpoint ids of peers with an authenticated session.
+    pub fn connected_peers(&self) -> Vec<EndpointId> {
+        self.peers.iter().map(|peer| peer.endpoint_id).collect()
+    }
+}
+
+/// Status of the whole agent.
+#[derive(Debug, Clone)]
+pub struct AgentStatus {
+    /// This device's persistent endpoint id.
+    pub endpoint_id: EndpointId,
+    /// Hostname announced to peers.
+    pub hostname: String,
+    /// Sockets actually bound.
+    pub bound_sockets: Vec<std::net::SocketAddr>,
+    /// Addresses iroh believes this endpoint has. Observed, not verified.
+    pub observed_addrs: Vec<PathAddr>,
+    /// The dialable address of this endpoint, as iroh currently reports it.
+    pub endpoint_addr: EndpointAddr,
+    /// What happened to the disposable cache at startup.
+    pub cache_outcome: CacheOutcome,
+    /// Whether the cache is currently usable.
+    pub cache_healthy: bool,
+    /// Per-network status, including configured but inactive networks.
+    pub networks: Vec<NetworkStatus>,
+}
+
+impl AgentStatus {
+    /// Looks up one network's status.
+    pub fn network(&self, network_id: &NetworkId) -> Option<&NetworkStatus> {
+        self.networks
+            .iter()
+            .find(|status| &status.network_id == network_id)
+    }
+}

crates/tsunagi-core/src/config.rs

@@ -0,0 +1,559 @@
+//! Library configuration.
+//!
+//! Everything the agent needs is passed in explicitly. The library reads no
+//! environment variables, installs no global state and picks no default
+//! directories behind the caller's back — [`StoragePaths::user_default`] exists
+//! but must be called on purpose.
+
+use std::net::SocketAddr;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::Duration;
+
+use crate::dataplane::SharedPlugin;
+use crate::discovery::{MainlineDiscovery, NetworkDiscovery};
+use crate::error::Result;
+
+/// Qualifier/organisation/application triple used for platform directories.
+///
+/// Unused on macOS, where the default paths are fixed system locations rather
+/// than per-user directories derived from this.
+#[cfg_attr(target_os = "macos", allow(dead_code))]
+const APP_NAME: &str = "tsunagi";
+
+/// Rendezvous slots per network; this bounds a discovery sample, not membership.
+pub const DHT_SLOTS: u8 = 16;
+/// BEP44 permits 1000 bencoded bytes; a 996-byte string has a four-byte prefix.
+pub const DHT_MAX_VALUE: usize = 996;
+/// Maximum direct addresses in one rendezvous record.
+pub const DHT_MAX_ADDRS: usize = 8;
+/// Maximum relay URL bytes in a rendezvous record.
+pub const DHT_MAX_RELAY_LEN: usize = 512;
+/// Application freshness, independent of storage nodes' own expiration.
+pub const DHT_RECORD_TTL: Duration = Duration::from_secs(900);
+/// Clock skew tolerated when reading a recently published record.
+pub const DHT_CLOCK_SKEW: Duration = Duration::from_secs(120);
+
+/// Maximum reassembled data-transport payload, independent of the path MTU.
+pub const MAX_DATA_DATAGRAM: usize = 64 * 1024;
+/// Maximum incomplete datagrams held by one authenticated data link.
+pub const MAX_DATA_ASSEMBLIES: usize = 64;
+/// Maximum allocated payload bytes for incomplete datagrams on one data link.
+pub const MAX_DATA_REASSEMBLY_BYTES: usize = 256 * 1024;
+/// Maximum fragments of one logical datagram, including after a path MTU change.
+pub const MAX_DATA_FRAGMENTS: usize = 128;
+/// Time allowed to assemble a datagram; loss never stalls later datagrams.
+pub const DATA_REASSEMBLY_TTL: Duration = Duration::from_secs(5);
+/// Completed or discarded packet IDs retained to ignore late duplicate fragments.
+pub const DATA_RECENT_IDS: usize = 128;
+
+/// Where the two stores live.
+///
+/// The mandatory state and the disposable cache are separate both logically and
+/// physically, so that the cache can be deleted at any time without touching
+/// identity or network configuration.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct StoragePaths {
+    /// Directory holding `state.sqlite` and the ownership lock.
+    pub state_dir: PathBuf,
+    /// Directory holding `cache.sqlite`.
+    pub cache_dir: PathBuf,
+}
+
+impl StoragePaths {
+    /// Uses explicit directories. Tests always use temporary directories.
+    pub fn new(state_dir: impl Into<PathBuf>, cache_dir: impl Into<PathBuf>) -> Self {
+        Self {
+            state_dir: state_dir.into(),
+            cache_dir: cache_dir.into(),
+        }
+    }
+
+    /// Puts both stores under one root, in `state/` and `cache/` subdirectories.
+    pub fn under(root: impl AsRef<Path>) -> Self {
+        let root = root.as_ref();
+        Self {
+            state_dir: root.join("state"),
+            cache_dir: root.join("cache"),
+        }
+    }
+
+    /// The default platform directories.
+    ///
+    /// Per user on Linux and Windows. On macOS the agent runs as a root
+    /// LaunchDaemon (creating a utun interface needs root, and `$HOME` under
+    /// `sudo` points at the wrong place), so the stores live in the system
+    /// locations instead — `/var/db/tsunagi` for the mandatory state and a
+    /// disposable cache beside it. A caller that wants somewhere else passes
+    /// explicit paths to [`StoragePaths::new`] or [`StoragePaths::under`].
+    pub fn user_default() -> Result<Self> {
+        #[cfg(target_os = "macos")]
+        {
+            Ok(Self {
+                state_dir: PathBuf::from("/var/db/tsunagi"),
+                cache_dir: PathBuf::from("/var/db/tsunagi/cache"),
+            })
+        }
+        #[cfg(not(target_os = "macos"))]
+        {
+            let dirs = directories::ProjectDirs::from("", "", APP_NAME).ok_or_else(|| {
+                crate::error::Error::Storage(
+                    "no valid home directory for platform config paths".into(),
+                )
+            })?;
+            Ok(Self {
+                state_dir: dirs.data_dir().to_path_buf(),
+                cache_dir: dirs.cache_dir().to_path_buf(),
+            })
+        }
+    }
+
+    /// Path of the mandatory state database.
+    pub fn state_db(&self) -> PathBuf {
+        self.state_dir.join("state.sqlite")
+    }
+
+    /// Path of the disposable cache database.
+    pub fn cache_db(&self) -> PathBuf {
+        self.cache_dir.join("cache.sqlite")
+    }
+
+    /// Path of the ownership lock file.
+    pub fn lock_file(&self) -> PathBuf {
+        self.state_dir.join("state.lock")
+    }
+}
+
+/// How the iroh endpoint is allowed to reach the outside world.
+///
+/// The default is [`TransportPolicy::LocalOnly`] so that a plain
+/// `AgentConfig::new(...)` never reaches the internet by accident. Callers that
+/// want public connectivity must opt in.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub enum TransportPolicy {
+    /// No relays, no address lookup service, no port mapping.
+    ///
+    /// Suitable for tests and for fully local deployments.
+    #[default]
+    LocalOnly,
+    /// Public address lookup, but no relays.
+    ///
+    /// Enables iroh's DNS/pkarr address lookup against the public service run
+    /// by Number 0 (the company behind iroh) at `dns.iroh.link`. This endpoint
+    /// publishes a signed record of its own addresses there, so peers can dial
+    /// it by endpoint id alone.
+    DirectOnly,
+    /// iroh's standard behaviour: public address lookup plus public relays.
+    ///
+    /// Maps to iroh's own `presets::N0`. As well as the address lookup above,
+    /// it uses Number 0's public relay servers as a fallback when a direct
+    /// path cannot be hole punched. They are fine for development and carry no
+    /// availability guarantee.
+    N0Defaults,
+}
+
+/// Bounds applied to everything that comes off the network.
+///
+/// Each of these is enforced before memory is allocated for the corresponding
+/// object where that is possible (notably [`Limits::max_frame_len`]).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Limits {
+    /// Maximum active VIP identities per network.
+    pub max_vips: usize,
+    /// Maximum permanent VIP deletion tombstones per network.
+    pub max_vip_tombstones: usize,
+    /// Maximum concurrent active TCP port checks across the agent.
+    pub max_vip_checks: usize,
+    /// Maximum signed fields per VIP, including removed fields.
+    pub max_vip_fields: usize,
+    /// Maximum listeners per VIP.
+    pub max_vip_listeners: usize,
+    /// Maximum backend endpoints per listener.
+    pub max_vip_backends: usize,
+    /// Maximum simultaneous local VIP flow bindings.
+    pub max_vip_flows: usize,
+    /// Maximum incomplete IPv4 datagrams per network in VIP reassembly.
+    pub max_vip_fragments: usize,
+    /// Largest accepted control frame payload, in bytes.
+    pub max_frame_len: usize,
+    /// Largest accepted hostname, in bytes.
+    pub max_hostname_len: usize,
+    /// Largest number of plugin capabilities in one announcement.
+    pub max_capabilities: usize,
+    /// Largest opaque plugin payload, in bytes.
+    pub max_capability_data_len: usize,
+    /// Largest private transport negotiation payload (also hard-capped at 4096 bytes).
+    pub max_transport_payload_len: usize,
+    /// Largest accepted echo payload in a ping/pong exchange, in bytes.
+    pub max_echo_payload_len: usize,
+    /// Largest accepted free-text reason string, in bytes.
+    pub max_reason_len: usize,
+    /// Largest number of signed records accepted in one snapshot.
+    pub max_state_records: usize,
+    /// Maximum durable retirement markers retained per network.
+    pub max_retirements: usize,
+    /// Deadline for the whole handshake.
+    pub handshake_timeout: Duration,
+    /// Deadline for one outbound dial attempt.
+    pub dial_timeout: Duration,
+    /// Deadline for writing one control frame.
+    ///
+    /// Liveness of an established session is delegated to QUIC: iroh configures
+    /// keep-alives and an idle timeout, so a dead peer surfaces as a read error
+    /// rather than needing a protocol-level heartbeat here.
+    pub write_timeout: Duration,
+    /// Maximum simultaneous outbound dials per network.
+    pub max_concurrent_dials: usize,
+    /// Maximum unverified discovery candidates retained per network.
+    pub max_discovery_candidates: usize,
+    /// Maximum simultaneous authenticated sessions per network.
+    pub max_sessions_per_network: usize,
+    /// Maximum simultaneous inbound connections being handshaken.
+    pub max_inbound_handshakes: usize,
+    /// Capacity of a session's outbound queue, providing backpressure.
+    pub session_send_queue: usize,
+    /// Capacity of the event broadcast channel.
+    pub event_buffer: usize,
+    /// Maximum address hints kept per peer in the cache.
+    pub max_hints_per_peer: usize,
+}
+
+impl Default for Limits {
+    fn default() -> Self {
+        Self {
+            max_vips: 64,
+            max_vip_tombstones: 1024,
+            max_vip_checks: 32,
+            max_vip_fields: 192,
+            max_vip_listeners: 16,
+            max_vip_backends: 32,
+            max_vip_flows: 4096,
+            max_vip_fragments: 64,
+            max_frame_len: 64 * 1024,
+            max_hostname_len: 255,
+            max_capabilities: 16,
+            max_transport_payload_len: 4096,
+            max_capability_data_len: 4 * 1024,
+            max_echo_payload_len: 4 * 1024,
+            max_reason_len: 256,
+            max_state_records: crate::state::MAX_RECORDS_PER_MESSAGE,
+            max_retirements: 1024,
+            handshake_timeout: Duration::from_secs(10),
+            dial_timeout: Duration::from_secs(10),
+            write_timeout: Duration::from_secs(30),
+            max_concurrent_dials: 8,
+            max_discovery_candidates: 16,
+            max_sessions_per_network: 64,
+            max_inbound_handshakes: 32,
+            session_send_queue: 64,
+            event_buffer: 512,
+            max_hints_per_peer: 8,
+        }
+    }
+}
+
+/// Bounded exponential backoff with jitter for reconnect attempts.
+#[derive(Debug, Clone, PartialEq)]
+pub struct ReconnectPolicy {
+    /// Delay before the first retry.
+    pub initial_delay: Duration,
+    /// Upper bound on the delay.
+    pub max_delay: Duration,
+    /// Multiplier applied after each failed attempt.
+    pub factor: f64,
+    /// Fraction of the delay applied as random jitter, in `0.0..=1.0`.
+    pub jitter: f64,
+    /// Give up on a peer after this many consecutive failures until it is seen
+    /// again by discovery. `None` means never give up while the network is up.
+    pub max_consecutive_failures: Option<u32>,
+}
+
+impl Default for ReconnectPolicy {
+    fn default() -> Self {
+        Self {
+            initial_delay: Duration::from_millis(250),
+            max_delay: Duration::from_secs(30),
+            factor: 2.0,
+            jitter: 0.3,
+            max_consecutive_failures: None,
+        }
+    }
+}
+
+impl ReconnectPolicy {
+    /// Delay to wait before retry number `attempt` (1-based), with jitter.
+    pub(crate) fn delay_for(&self, attempt: u32) -> Duration {
+        let exp = self.factor.powi(attempt.saturating_sub(1).min(32) as i32);
+        let base = self.initial_delay.as_secs_f64() * exp;
+        let capped = base.min(self.max_delay.as_secs_f64());
+        let jitter = self.jitter.clamp(0.0, 1.0);
+        let factor = 1.0 - jitter + jitter * 2.0 * rand::random::<f64>();
+        Duration::from_secs_f64((capped * factor).max(0.0))
+    }
+}
+
+/// Scheduling of candidate lookup and independent self-publication.
+#[derive(Debug, Clone)]
+pub struct DiscoveryPolicy {
+    /// Mean interval between successful self-publications (20% jitter).
+    pub publish_interval: Duration,
+    /// Initial delay between unsuccessful bootstrap lookups.
+    pub lookup_interval: Duration,
+    /// Maximum delay between bootstrap lookups.
+    pub max_lookup_interval: Duration,
+    /// Continuous isolation before a previously connected network searches again.
+    pub reconnect_delay: Duration,
+    /// Deadline for one backend operation.
+    pub request_timeout: Duration,
+}
+
+impl Default for DiscoveryPolicy {
+    fn default() -> Self {
+        Self {
+            publish_interval: Duration::from_secs(300),
+            lookup_interval: Duration::from_secs(5),
+            max_lookup_interval: Duration::from_secs(30),
+            reconnect_delay: Duration::from_secs(60),
+            request_timeout: Duration::from_secs(60),
+        }
+    }
+}
+
+/// Everything needed to start an [`crate::Agent`].
+#[derive(Clone)]
+pub struct AgentConfig {
+    /// Where the mandatory state and the disposable cache live.
+    pub paths: StoragePaths,
+    /// Explicit local bind addresses. Empty means iroh's defaults.
+    ///
+    /// Tests bind to `127.0.0.1:0` so each agent gets a dynamic port.
+    pub bind_addrs: Vec<SocketAddr>,
+    /// How much external connectivity machinery the endpoint may use.
+    pub transport: TransportPolicy,
+    /// Real QUIC transport settings for constrained-path integration tests.
+    #[cfg(feature = "testing")]
+    pub test_quic_transport: Option<iroh::endpoint::QuicTransportConfig>,
+    /// Peers with no direct data path, for tests. See
+    /// [`AgentConfig::with_unreachable_data_peers`].
+    #[cfg(feature = "testing")]
+    pub unreachable_data_peers: Arc<std::sync::Mutex<std::collections::HashSet<iroh::EndpointId>>>,
+    /// Hostname announced to peers. `None` keeps whatever the state store holds,
+    /// falling back to the OS hostname and finally to a short endpoint id.
+    pub hostname: Option<String>,
+    /// Additional discovery backend, composed with Mainline when it is enabled.
+    pub discovery: Option<Arc<dyn NetworkDiscovery>>,
+    /// Optional Mainline client, shared by this agent's networks. The CLI enables
+    /// it by default except in local-only mode; library callers opt in explicitly.
+    pub dht: Option<MainlineDiscovery>,
+    /// Lookup, publication and recovery scheduling.
+    pub discovery_policy: DiscoveryPolicy,
+    /// How often each active network maintains known peers and observes address changes.
+    pub discovery_interval: Duration,
+    /// Bounds applied to network input.
+    pub limits: Limits,
+    /// Reconnect backoff policy.
+    pub reconnect: ReconnectPolicy,
+    /// IP plugins whose capabilities are announced and dispatched.
+    pub plugins: Vec<SharedPlugin>,
+    /// Initial ordered selection for networks without a saved override.
+    /// `None` enables all registered plugins; an empty list is control-only.
+    pub default_protocols: Option<Vec<String>>,
+    /// Custom packet transports registered per protocol.
+    pub custom_transports:
+        std::collections::HashMap<String, Arc<dyn crate::dataplane::PacketTransport>>,
+    /// Where the overlay interface comes from, when the agent should have one.
+    ///
+    /// One agent, one interface: it belongs here rather than to a protocol,
+    /// because every protocol carries traffic for the same addresses on it.
+    /// `None` means no interface — the protocols still run and their packets
+    /// are discarded.
+    pub tun_factory: Option<Arc<dyn crate::overlay::TunFactory>>,
+    /// The name to ask the operating system for.
+    pub interface_name: String,
+    /// The interface MTU.
+    pub interface_mtu: u32,
+    /// Whether new networks derive their IPv4 range from their network id.
+    ///
+    /// Enabled by default. Explicit ranges and IPv4 opt-out disable this.
+    pub overlay_ipv4_auto: bool,
+    /// The IPv4 overlay range this agent proposes when automatic selection is off.
+    ///
+    /// Addresses are allocated from it and recorded in signed state, so a
+    /// participant keeps the same one across restarts. A network that has
+    /// already settled on another range wins: a joining agent adopts what it
+    /// finds rather than imposing this.
+    pub overlay_ipv4_range: Option<crate::state::Ipv4Range>,
+}
+
+impl AgentConfig {
+    /// Creates a configuration with local-only transport and default limits.
+    pub fn new(paths: StoragePaths) -> Self {
+        Self {
+            paths,
+            bind_addrs: Vec::new(),
+            transport: TransportPolicy::default(),
+            #[cfg(feature = "testing")]
+            test_quic_transport: None,
+            #[cfg(feature = "testing")]
+            unreachable_data_peers: Arc::new(std::sync::Mutex::new(
+                std::collections::HashSet::new(),
+            )),
+            hostname: None,
+            discovery: None,
+            dht: None,
+            discovery_policy: DiscoveryPolicy::default(),
+            discovery_interval: Duration::from_secs(5),
+            limits: Limits::default(),
+            reconnect: ReconnectPolicy::default(),
+            plugins: Vec::new(),
+            default_protocols: None,
+            custom_transports: std::collections::HashMap::new(),
+            tun_factory: None,
+            interface_name: "tsun0".to_string(),
+            interface_mtu: 1280,
+            overlay_ipv4_auto: true,
+            overlay_ipv4_range: None,
+        }
+    }
+
+    /// Binds to loopback with a dynamic port. Used by the test suite.
+    pub fn with_loopback_bind(mut self) -> Self {
+        self.bind_addrs = vec![
+            SocketAddr::from(([127, 0, 0, 1], 0)),
+            SocketAddr::from(([0, 0, 0, 0, 0, 0, 0, 1], 0)),
+        ];
+        self
+    }
+
+    /// Sets an explicit list of bind addresses.
+    pub fn with_bind_addrs(mut self, addrs: impl IntoIterator<Item = SocketAddr>) -> Self {
+        self.bind_addrs = addrs.into_iter().collect();
+        self
+    }
+
+    /// Peers this agent will not open or accept a **direct** data link with.
+    ///
+    /// Test-only, and behind the `testing` feature so a release build has
+    /// no such switch. The one arrangement that cannot be made on a single
+    /// host is two agents that both reach a third and not each other,
+    /// which is exactly the case a relay exists for.
+    #[cfg(feature = "testing")]
+    pub fn with_unreachable_data_peers(
+        mut self,
+        peers: Arc<std::sync::Mutex<std::collections::HashSet<iroh::EndpointId>>>,
+    ) -> Self {
+        self.unreachable_data_peers = peers;
+        self
+    }
+
+    /// Sets the transport policy.
+    pub fn with_transport(mut self, transport: TransportPolicy) -> Self {
+        self.transport = transport;
+        self
+    }
+
+    /// Sets the discovery backend.
+    pub fn with_discovery(mut self, discovery: Arc<dyn NetworkDiscovery>) -> Self {
+        self.discovery = Some(discovery);
+        self
+    }
+
+    /// Adds Mainline discovery alongside the configured candidate backend.
+    pub fn with_dht(mut self, dht: MainlineDiscovery) -> Self {
+        self.dht = Some(dht);
+        self
+    }
+
+    /// Sets lookup, publication and isolation recovery timing.
+    pub fn with_discovery_policy(mut self, policy: DiscoveryPolicy) -> Self {
+        self.discovery_policy = policy;
+        self
+    }
+
+    /// Sets the known-peer maintenance and endpoint-address observation interval.
+    pub fn with_discovery_interval(mut self, interval: Duration) -> Self {
+        self.discovery_interval = interval;
+        self
+    }
+
+    /// Sets the announced hostname.
+    pub fn with_hostname(mut self, hostname: impl Into<String>) -> Self {
+        self.hostname = Some(hostname.into());
+        self
+    }
+
+    /// Sets the IPv4 overlay range this agent proposes, or disables IPv4.
+    pub fn with_overlay_ipv4_range(mut self, range: Option<crate::state::Ipv4Range>) -> Self {
+        self.overlay_ipv4_auto = false;
+        self.overlay_ipv4_range = range;
+        self
+    }
+
+    /// Registers an IP plugin.
+    pub fn with_plugin(mut self, plugin: SharedPlugin) -> Self {
+        self.plugins.push(plugin);
+        self
+    }
+
+    /// Registers a custom packet transport for a specific protocol.
+    pub fn with_custom_transport(
+        mut self,
+        protocol: impl Into<String>,
+        transport: Arc<dyn crate::dataplane::PacketTransport>,
+    ) -> Self {
+        self.custom_transports.insert(protocol.into(), transport);
+        self
+    }
+
+    /// Gives the agent an overlay interface, from this factory.
+    pub fn with_interface(
+        mut self,
+        factory: Arc<dyn crate::overlay::TunFactory>,
+        name: impl Into<String>,
+        mtu: u32,
+    ) -> Self {
+        self.tun_factory = Some(factory);
+        self.interface_name = name.into();
+        self.interface_mtu = mtu;
+        self
+    }
+
+    /// Replaces the limits.
+    pub fn with_limits(mut self, limits: Limits) -> Self {
+        self.limits = limits;
+        self
+    }
+
+    /// Replaces the reconnect policy.
+    pub fn with_reconnect(mut self, reconnect: ReconnectPolicy) -> Self {
+        self.reconnect = reconnect;
+        self
+    }
+}
+
+impl std::fmt::Debug for AgentConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AgentConfig")
+            .field("paths", &self.paths)
+            .field("bind_addrs", &self.bind_addrs)
+            .field("transport", &self.transport)
+            .field("hostname", &self.hostname)
+            .field("discovery", &self.discovery.as_ref().map(|d| d.name()))
+            .field("dht", &self.dht.is_some())
+            .field("discovery_policy", &self.discovery_policy)
+            .field("discovery_interval", &self.discovery_interval)
+            .field("limits", &self.limits)
+            .field("reconnect", &self.reconnect)
+            .field(
+                "plugins",
+                &self
+                    .plugins
+                    .iter()
+                    .map(|p| p.protocol_id().to_string())
+                    .collect::<Vec<_>>(),
+            )
+            .finish()
+    }
+}
+/// Maximum number of transport links a routed datagram may traverse.
+pub const ROUTING_HOP_LIMIT: u8 = 16;

crates/tsunagi-core/src/dataplane/mod.rs

@@ -19,8 +19,10 @@
 //! A data plane failure never stops the daemon: errors returned here are
 //! recorded and surfaced, the control plane keeps running.
 
+pub mod relay;
+pub mod routing;
 pub mod transport;
-pub mod wireguard;
+pub mod tunnel;
 
 use std::sync::Arc;
 
@@ -30,11 +32,41 @@ use tokio::sync::mpsc;
 use crate::BoxFuture;
 use crate::identity::NetworkId;
 
-pub use transport::{PacketLink, PacketTransport, SharedLink, TransportError};
+pub use transport::{
+    MultiPacketTransport, PacketLink, PacketTransport, SharedLink, TransportError,
+};
+pub use tunnel::{GenericTunnelPlugin, TunnelCodec};
 
 /// Maximum length of a plugin protocol identifier.
 pub const MAX_PROTOCOL_ID_LEN: usize = 32;
 
+/// Validates bounded local protocol preferences before storing or applying them.
+pub(crate) fn validate_protocols(protocols: &[String], max: usize) -> crate::error::Result<()> {
+    if protocols.len() > max {
+        return Err(crate::Error::InvalidProtocols(format!(
+            "at most {max} protocols are allowed"
+        )));
+    }
+    for (index, name) in protocols.iter().enumerate() {
+        if name.is_empty()
+            || name.len() > MAX_PROTOCOL_ID_LEN
+            || !name
+                .bytes()
+                .all(|c| c.is_ascii_alphanumeric() || c == b'-' || c == b'_')
+        {
+            return Err(crate::Error::InvalidProtocols(
+                "names must be 1–32 ASCII letters, digits, hyphens or underscores".into(),
+            ));
+        }
+        if protocols[..index].contains(name) {
+            return Err(crate::Error::InvalidProtocols(format!(
+                "`{name}` is listed more than once"
+            )));
+        }
+    }
+    Ok(())
+}
+
 /// An announcement of one IP plugin's capability.
 ///
 /// `data` is opaque to the core. Nothing in it may be interpreted as a shell
@@ -42,7 +74,7 @@ pub const MAX_PROTOCOL_ID_LEN: usize = 32;
 /// chooses to do so must validate it itself.
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub struct PluginCapability {
-    /// Protocol identifier, e.g. `wireguard`. Bounded by [`MAX_PROTOCOL_ID_LEN`].
+    /// Protocol identifier, e.g. `wg-quic`. Bounded by [`MAX_PROTOCOL_ID_LEN`].
     pub protocol: String,
     /// Version of the plugin's announcement format.
     pub version: u16,
@@ -52,6 +84,51 @@ pub struct PluginCapability {
     pub data: Vec<u8>,
 }
 
+/// Where a protocol hands the packets it has decrypted.
+///
+/// A protocol proves *who* sent a packet; it does not know what that member
+/// is entitled to say, because entitlement is an address claim the system
+/// level holds. So a decrypted packet goes here rather than straight to an
+/// interface, and is checked on the way.
+pub trait PacketSink: Send + Sync + 'static {
+    /// Hands over one packet, attributed to the peer whose tunnel decrypted
+    /// it.
+    fn deliver<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        packet: bytes::Bytes,
+    ) -> BoxFuture<'a, ()>;
+}
+
+/// A sink that drops everything, for a protocol running without an interface.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct DiscardPackets;
+
+impl PacketSink for DiscardPackets {
+    fn deliver<'a>(
+        &'a self,
+        _network: NetworkId,
+        _peer: EndpointId,
+        _packet: bytes::Bytes,
+    ) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// One setting a protocol accepts.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ProtocolOption {
+    /// The key, as written in `key=value`.
+    pub key: &'static str,
+    /// A word for the value, for the help line: `BYTES`, `NAME`.
+    pub value: &'static str,
+    /// What it does.
+    pub help: &'static str,
+    /// What happens when it is not given.
+    pub default: Option<&'static str>,
+}
+
 /// Errors a plugin may return. They are recorded, never fatal for the agent.
 #[derive(Debug, thiserror::Error)]
 #[non_exhaustive]
@@ -69,7 +146,7 @@ pub enum PluginError {
 
 /// A request a plugin makes of the agent that owns it.
 #[derive(Debug)]
-pub(crate) enum PluginRequest {
+pub enum PluginRequest {
     /// Re-send this agent's announcement to every peer of a network.
     Reannounce(NetworkId),
     /// Surface a plugin error on the agent's event stream.
@@ -81,6 +158,17 @@ pub(crate) enum PluginRequest {
         /// Human readable reason, free of secrets.
         reason: String,
     },
+    /// A plugin reported that a data link or tunnel for a peer failed.
+    ProtocolFailure {
+        /// Network the failure occurred in.
+        network: NetworkId,
+        /// Peer whose protocol failed.
+        peer: EndpointId,
+        /// Protocol that failed.
+        protocol: String,
+        /// Human readable reason.
+        reason: String,
+    },
 }
 
 /// The agent-side handle a plugin is given when it is attached.
@@ -94,18 +182,73 @@ pub(crate) enum PluginRequest {
 #[derive(Clone)]
 pub struct PluginContext {
     sender: Option<mpsc::Sender<PluginRequest>>,
+    local: Option<EndpointId>,
+    /// Where decrypted packets go. Absent when nothing is carrying traffic,
+    /// in which case a protocol still runs and its packets are discarded.
+    sink: Option<Arc<dyn PacketSink>>,
 }
 
 impl PluginContext {
-    pub(crate) fn new(sender: mpsc::Sender<PluginRequest>) -> Self {
+    pub(crate) fn new(
+        sender: mpsc::Sender<PluginRequest>,
+        local: EndpointId,
+        sink: Option<Arc<dyn PacketSink>>,
+    ) -> Self {
         Self {
             sender: Some(sender),
+            local: Some(local),
+            sink,
+        }
+    }
+
+    /// Where to hand a decrypted packet.
+    ///
+    /// Always answers: with no interface to write to, the packets are
+    /// discarded, which is what `--no-tun` means and is not an error.
+    pub fn packet_sink(&self) -> Arc<dyn PacketSink> {
+        match &self.sink {
+            Some(sink) => Arc::clone(sink),
+            None => Arc::new(DiscardPackets),
         }
     }
 
     /// A context that discards everything, for plugins used outside an agent.
     pub fn detached() -> Self {
-        Self { sender: None }
+        Self {
+            sender: None,
+            local: None,
+            sink: None,
+        }
+    }
+
+    /// A context configured with a custom packet sink.
+    pub fn with_sink(sink: Arc<dyn PacketSink>) -> Self {
+        Self {
+            sender: None,
+            local: None,
+            sink: Some(sink),
+        }
+    }
+
+    /// A context configured with a channel for plugin requests and custom sink, for tests.
+    #[cfg(feature = "testing")]
+    pub fn for_test(
+        sender: mpsc::Sender<PluginRequest>,
+        local: EndpointId,
+        sink: Option<Arc<dyn PacketSink>>,
+    ) -> Self {
+        Self {
+            sender: Some(sender),
+            local: Some(local),
+            sink,
+        }
+    }
+
+    /// This agent's own endpoint id, when the context is attached.
+    ///
+    /// A plugin needs it to find itself in the agreed allocation.
+    pub fn local_endpoint_id(&self) -> Option<EndpointId> {
+        self.local
     }
 
     fn send(&self, request: PluginRequest) {
@@ -142,12 +285,29 @@ impl PluginContext {
             reason: reason.into(),
         });
     }
+
+    /// Reports that data plane connectivity or a tunnel for a peer has failed.
+    pub fn report_peer_protocol_failure(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: impl Into<String>,
+        reason: impl Into<String>,
+    ) {
+        self.send(PluginRequest::ProtocolFailure {
+            network,
+            peer,
+            protocol: protocol.into(),
+            reason: reason.into(),
+        });
+    }
 }
 
 impl std::fmt::Debug for PluginContext {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("PluginContext")
             .field("attached", &self.sender.is_some())
+            .field("local", &self.local.map(|id| id.fmt_short().to_string()))
             .finish()
     }
 }
@@ -157,11 +317,34 @@ impl std::fmt::Debug for PluginContext {
 /// Implementations must be cheap and non-blocking: the agent calls them from
 /// its runtime tasks. Anything slow belongs in the plugin's own tasks.
 pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
-    /// Stable protocol identifier, e.g. `wireguard`.
+    /// Stable protocol identifier, e.g. `wg-quic`.
     ///
     /// Must be non-empty and at most [`MAX_PROTOCOL_ID_LEN`] bytes.
     fn protocol_id(&self) -> &str;
 
+    /// The wire version of this protocol.
+    ///
+    /// Two peers carry traffic for each other only when they have the same
+    /// protocol at the same version. There is no negotiating a middle
+    /// ground: a protocol either speaks the same words at both ends or it
+    /// does not, and pretending otherwise produces a link that fails later
+    /// and less clearly.
+    ///
+    /// **Not the software version.** A plugin crate has its own version and
+    /// it is nobody else's business: two peers on different builds work
+    /// together for as long as the wire between them has not changed. This
+    /// number moves only when the bytes do, so it must never be derived from
+    /// `CARGO_PKG_VERSION` or anything else that moves with a release.
+    fn protocol_version(&self) -> u16;
+
+    /// The settings this protocol accepts, and what they mean.
+    ///
+    /// Declared rather than documented elsewhere, so the agent can list them
+    /// without knowing anything about the protocol.
+    fn options(&self) -> &'static [ProtocolOption] {
+        &[]
+    }
+
     /// Called once, when the agent starts, before any network is activated.
     ///
     /// The plugin keeps the context to ask for re-announcements and to report
@@ -197,6 +380,34 @@ pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
         capability: &PluginCapability,
     ) -> std::result::Result<(), PluginError>;
 
+    /// The overlay addresses the network has agreed on.
+    ///
+    /// Allocated rather than derived, and backed by the signed records in
+    /// [`crate::state`], so a participant keeps its address across restarts
+    /// and long absences. Called whenever the agreed picture changes.
+    fn on_address_allocation(
+        &self,
+        network: NetworkId,
+        range: crate::state::Ipv4Range,
+        allocations: &[(EndpointId, std::net::Ipv4Addr)],
+    ) {
+        let _ = (network, range, allocations);
+    }
+
+    /// Carries one packet to a peer, encrypting it however this protocol
+    /// does.
+    ///
+    /// `false` when it cannot right now — no link, no tunnel, not this
+    /// protocol's peer — which the caller reports rather than treats as an
+    /// error. The packet came off the one interface the agent owns, and
+    /// which protocol takes it is settled by asking.
+    ///
+    /// The default carries nothing, which is right for a plugin that only
+    /// announces something.
+    fn carry(&self, _network: NetworkId, _peer: EndpointId, _packet: bytes::Bytes) -> bool {
+        false
+    }
+
     /// A data plane link to a peer is available for this plugin's protocol.
     ///
     /// The plugin moves its packets over this link and never learns how the
@@ -216,6 +427,19 @@ pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
     /// The plugin is expected to remove whatever it created for that network.
     fn on_network_deactivated(&self, network: NetworkId);
 
+    /// Called when this agent has left a network for good.
+    ///
+    /// Deactivation is temporary and keeps everything ready for next time;
+    /// this is the other one. Whatever the plugin holds *durably* for that
+    /// network — a key of its own, a file, a record — goes now, because the
+    /// agent is no longer a member and keeping it is keeping a secret for a
+    /// network it cannot rejoin without being told the secret again.
+    ///
+    /// Always preceded by [`IpPlugin::on_network_deactivated`], so this is
+    /// only about what outlives a session. Defaulted to nothing, for a
+    /// plugin that stores nothing.
+    fn on_network_forgotten(&self, _network: NetworkId) {}
+
     /// Called once when the agent shuts down.
     ///
     /// The plugin removes the system objects it created and stops its tasks.
@@ -263,6 +487,10 @@ impl IpPlugin for TestCapabilityPlugin {
         &self.protocol
     }
 
+    fn protocol_version(&self) -> u16 {
+        1
+    }
+
     fn local_capability(
         &self,
         _network: NetworkId,

crates/tsunagi-core/src/dataplane/relay.rs

@@ -0,0 +1,646 @@
+//! Userspace forwarding. Topology updates publish immutable tables; the packet
+//! path takes no routing mutex, walks no graph and never inspects plugin bytes.
+//! Each transport reader forwards transit immediately, independently of the
+//! destination protocol's receive task. Only local delivery enters its inbox.
+
+use std::collections::{HashMap, HashSet};
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+
+use arc_swap::ArcSwap;
+use bytes::Bytes;
+use iroh::EndpointId;
+use tokio::sync::{Notify, mpsc, watch};
+use tokio::task::JoinHandle;
+
+use super::routing::{FlowId, PeerId, RoutingTable, envelope, mix64};
+use super::transport::{PacketLink, SharedLink, TransportError};
+use crate::BoxFuture;
+use crate::config::{MAX_DATA_DATAGRAM, ROUTING_HOP_LIMIT};
+use crate::identity::NetworkId;
+
+/// Fixed routing header, subtracted from the logical transport MTU.
+pub const RELAY_OVERHEAD: usize = envelope::HEADER;
+const INBOX: usize = 256;
+
+/// Forwarding counters, shared by all protocols of one network.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct RelayCounters {
+    /// Datagrams originated here and sent through another member.
+    pub sent_via: u64,
+    /// Transit datagrams handed to the next transport link.
+    pub forwarded: u64,
+    /// Local datagrams received through another member.
+    pub received_via: u64,
+    /// Missing or failed next hop.
+    pub dropped_no_link: u64,
+    /// Malformed envelope or unknown member.
+    pub dropped_unknown: u64,
+    /// Exhausted hop limit, including loops during topology convergence.
+    pub dropped_hop_limit: u64,
+    /// Full local protocol inbox. Transit does not use this queue.
+    pub dropped_congested: u64,
+}
+
+#[derive(Debug, Default)]
+struct Tally {
+    sent_via: AtomicU64,
+    forwarded: AtomicU64,
+    received_via: AtomicU64,
+    dropped_no_link: AtomicU64,
+    dropped_unknown: AtomicU64,
+    dropped_hop_limit: AtomicU64,
+    dropped_congested: AtomicU64,
+}
+
+impl Tally {
+    fn snapshot(&self) -> RelayCounters {
+        RelayCounters {
+            sent_via: self.sent_via.load(Ordering::Relaxed),
+            forwarded: self.forwarded.load(Ordering::Relaxed),
+            received_via: self.received_via.load(Ordering::Relaxed),
+            dropped_no_link: self.dropped_no_link.load(Ordering::Relaxed),
+            dropped_unknown: self.dropped_unknown.load(Ordering::Relaxed),
+            dropped_hop_limit: self.dropped_hop_limit.load(Ordering::Relaxed),
+            dropped_congested: self.dropped_congested.load(Ordering::Relaxed),
+        }
+    }
+}
+
+/// The links to one neighbour, best first. Which transport carries a frame
+/// is decided here and nowhere else: the end-to-end protocol is in the
+/// envelope and never depends on it.
+#[derive(Debug)]
+struct NextHop {
+    peer: PeerId,
+    links: Box<[SharedLink]>,
+}
+
+impl NextHop {
+    fn usable(&self) -> bool {
+        self.links.iter().any(|link| !link.is_closed())
+    }
+
+    fn first_open(&self) -> Option<&SharedLink> {
+        self.links.iter().find(|link| !link.is_closed())
+    }
+
+    /// Hands a frame to the best open link that accepts it. A link that
+    /// cannot carry the frame (too large for its path, closed a moment ago)
+    /// leaves it to the next one; the frame is only copied when there is
+    /// another to try.
+    fn send(&self, frame: Bytes) -> Result<(), TransportError> {
+        let mut open = self.links.iter().filter(|link| !link.is_closed());
+        let Some(mut current) = open.next() else {
+            return Err(TransportError::Closed);
+        };
+        for next in open {
+            if current.send(frame.clone()).is_ok() {
+                return Ok(());
+            }
+            current = next;
+        }
+        current.send(frame)
+    }
+}
+
+#[derive(Debug)]
+struct ForwardRoute {
+    hops: u8,
+    next: Box<[NextHop]>,
+}
+
+impl ForwardRoute {
+    fn select(&self, source: &PeerId, flow: FlowId) -> Option<&NextHop> {
+        if self.next.len() == 1 {
+            return self.next.first().filter(|hop| hop.usable());
+        }
+        let seed = u64::from_le_bytes(source[..8].try_into().ok()?);
+        let start = mix64(flow ^ seed) as usize % self.next.len();
+        // Stable fallback if a link closed just before its table was replaced.
+        (0..self.next.len())
+            .map(|i| &self.next[(start + i) % self.next.len()])
+            .find(|hop| hop.usable())
+    }
+}
+
+#[derive(Debug, Default)]
+struct ForwardingTable {
+    routes: HashMap<PeerId, ForwardRoute>,
+    /// Local delivery, by who sent it and which end-to-end protocol it is.
+    inboxes: HashMap<(PeerId, u64), mpsc::Sender<Bytes>>,
+    members: HashSet<PeerId>,
+}
+
+#[derive(Debug)]
+struct Plane {
+    local: PeerId,
+    table: ArcSwap<ForwardingTable>,
+    tally: Arc<Tally>,
+}
+
+impl Plane {
+    /// Synchronous hot path: one snapshot, bounded header decode, table lookup,
+    /// and transport send. An exclusively owned buffer is modified in place.
+    fn receive(&self, incoming: PeerId, frame: Bytes) {
+        let Some(header) = envelope::decode(&frame) else {
+            self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+            return;
+        };
+        let table = self.table.load();
+        if !table.members.contains(&header.source) || header.source == self.local {
+            self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+            return;
+        }
+        if header.destination == self.local {
+            let Some(inbox) = table.inboxes.get(&(header.source, header.protocol)) else {
+                self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+                return;
+            };
+            if inbox.try_send(frame.slice(RELAY_OVERHEAD..)).is_err() {
+                self.tally.dropped_congested.fetch_add(1, Ordering::Relaxed);
+            } else if incoming != header.source {
+                self.tally.received_via.fetch_add(1, Ordering::Relaxed);
+            }
+            return;
+        }
+        if header.remaining <= 1 {
+            self.tally.dropped_hop_limit.fetch_add(1, Ordering::Relaxed);
+            return;
+        }
+        let hop = table
+            .routes
+            .get(&header.destination)
+            .and_then(|route| route.select(&header.source, header.flow));
+        let Some(hop) = hop else {
+            self.tally.dropped_no_link.fetch_add(1, Ordering::Relaxed);
+            return;
+        };
+        if hop.send(envelope::decrement(frame)).is_ok() {
+            self.tally.forwarded.fetch_add(1, Ordering::Relaxed);
+        } else {
+            self.tally.dropped_no_link.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+}
+
+/// One physical link to a neighbour, of whatever kind of transport.
+#[derive(Debug)]
+struct RawLink {
+    kind: String,
+    link: SharedLink,
+    reader: JoinHandle<()>,
+}
+
+/// How a destination is reached right now.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct RouteSummary {
+    /// Number of transport links to it. One is a direct link.
+    pub hops: u8,
+    /// The neighbour the first link goes to. For a direct link that is the
+    /// destination itself.
+    pub via: EndpointId,
+}
+
+#[derive(Debug)]
+struct State {
+    plane: Arc<Plane>,
+    /// Transport kinds, best first. A neighbour with several links is
+    /// reached over the first of these it has.
+    priority: Vec<String>,
+    graph: HashMap<PeerId, Vec<PeerId>>,
+    members: HashSet<PeerId>,
+    raw: HashMap<PeerId, Vec<RawLink>>,
+    links: HashMap<(PeerId, String), Arc<PeerLink>>,
+}
+
+impl State {
+    fn rank(&self, kind: &str) -> usize {
+        self.priority
+            .iter()
+            .position(|known| known == kind)
+            .unwrap_or(usize::MAX)
+    }
+
+    /// A neighbour's links, best kind first.
+    fn ordered(&self, peer: &PeerId) -> Vec<&RawLink> {
+        let mut links: Vec<&RawLink> = self
+            .raw
+            .get(peer)
+            .map(|links| links.iter().collect())
+            .unwrap_or_default();
+        links.sort_by_key(|raw| (self.rank(&raw.kind), raw.kind.clone()));
+        links
+    }
+
+    fn publish(&self) {
+        let mut graph = self.graph.clone();
+        let mut neighbours: Vec<PeerId> = self
+            .raw
+            .iter()
+            .filter(|(_, links)| links.iter().any(|raw| !raw.link.is_closed()))
+            .map(|(peer, _)| *peer)
+            .collect();
+        neighbours.sort_unstable();
+        graph.insert(self.plane.local, neighbours);
+        let routing = RoutingTable::build(self.plane.local, &graph, ROUTING_HOP_LIMIT);
+        let routes = routing
+            .iter()
+            .filter_map(|(peer, route)| {
+                let next: Box<[_]> = route
+                    .next_hops
+                    .iter()
+                    .filter_map(|id| {
+                        let links: Box<[_]> = self
+                            .ordered(id)
+                            .into_iter()
+                            .map(|raw| raw.link.clone())
+                            .collect();
+                        (!links.is_empty()).then_some(NextHop { peer: *id, links })
+                    })
+                    .collect();
+                (!next.is_empty()).then_some((
+                    *peer,
+                    ForwardRoute {
+                        hops: route.hops,
+                        next,
+                    },
+                ))
+            })
+            .collect();
+        let mut members = self.members.clone();
+        members.extend(self.raw.keys().copied());
+        self.plane.table.store(Arc::new(ForwardingTable {
+            routes,
+            members,
+            inboxes: self
+                .links
+                .iter()
+                .map(|((peer, _), link)| ((*peer, link.tag), link.inbox_tx.clone()))
+                .collect(),
+        }));
+    }
+
+    fn stop_readers(&mut self) {
+        for raw in self.raw.values().flatten() {
+            raw.reader.abort();
+        }
+    }
+}
+
+impl Drop for State {
+    fn drop(&mut self) {
+        self.stop_readers();
+        for link in self.links.values() {
+            link.close();
+        }
+        self.plane.table.store(Arc::default());
+    }
+}
+
+/// One network's transport links and its published forwarding table.
+///
+/// There is one table for the whole network, whatever protocols run over it.
+/// A link is a way to reach a neighbour; a protocol is an end-to-end channel
+/// to a peer. The two meet only in the envelope, which names the protocol, so
+/// a tunnel keeps working over any mix of transports and a middle member
+/// needs no protocol in common with either end.
+#[derive(Debug)]
+pub struct RelayHub {
+    network: NetworkId,
+    state: Mutex<State>,
+    tally: Arc<Tally>,
+    changed: Arc<Notify>,
+}
+
+impl RelayHub {
+    /// Creates the router at this network's local endpoint.
+    pub fn new(network: NetworkId, local: EndpointId) -> Arc<Self> {
+        let tally = Arc::new(Tally::default());
+        Arc::new(Self {
+            network,
+            state: Mutex::new(State {
+                plane: Arc::new(Plane {
+                    local: *local.as_bytes(),
+                    table: ArcSwap::from_pointee(ForwardingTable::default()),
+                    tally: tally.clone(),
+                }),
+                priority: Vec::new(),
+                graph: HashMap::new(),
+                members: HashSet::new(),
+                raw: HashMap::new(),
+                links: HashMap::new(),
+            }),
+            tally,
+            changed: Arc::default(),
+        })
+    }
+
+    fn state(&self) -> std::sync::MutexGuard<'_, State> {
+        self.state.lock().unwrap_or_else(|error| error.into_inner())
+    }
+
+    /// Sets which kinds of transport are preferred, best first, when a
+    /// neighbour can be reached over more than one.
+    pub fn set_priority(&self, kinds: Vec<String>) {
+        let mut state = self.state();
+        if state.priority != kinds {
+            state.priority = kinds;
+            state.publish();
+        }
+    }
+
+    /// Wakes the control loop promptly when a transport reader ends.
+    pub async fn changed(&self) {
+        self.changed.notified().await;
+    }
+
+    /// Current counters.
+    pub fn counters(&self) -> RelayCounters {
+        self.tally.snapshot()
+    }
+
+    /// Publishes authenticated topology. Refreshing identical state is cheap.
+    ///
+    /// Every row of `graph` is one member's own account of who it has a link
+    /// to; what kind of link it is does not matter here.
+    pub fn set_topology(&self, graph: HashMap<PeerId, Vec<PeerId>>, members: HashSet<PeerId>) {
+        let mut state = self.state();
+        if state.graph != graph || state.members != members {
+            state.graph = graph;
+            state.members = members;
+            state.publish();
+        }
+    }
+
+    /// Whether the current table has a path to this destination.
+    pub fn reachable(&self, peer: EndpointId) -> bool {
+        self.state()
+            .plane
+            .table
+            .load()
+            .routes
+            .contains_key(peer.as_bytes())
+    }
+
+    /// How a destination is reached right now, if it is.
+    pub fn route(&self, peer: EndpointId) -> Option<RouteSummary> {
+        let table = self.state().plane.table.load_full();
+        let route = table.routes.get(peer.as_bytes())?;
+        let via = EndpointId::from_bytes(&route.next.first()?.peer).ok()?;
+        Some(RouteSummary {
+            hops: route.hops,
+            via,
+        })
+    }
+
+    /// The kinds of transport with an open direct link to a peer, best
+    /// first.
+    pub fn direct_kinds(&self, peer: EndpointId) -> Vec<String> {
+        let state = self.state();
+        state
+            .ordered(peer.as_bytes())
+            .into_iter()
+            .filter(|raw| !raw.link.is_closed())
+            .map(|raw| raw.kind.clone())
+            .collect()
+    }
+
+    /// Stable end-to-end channel for a protocol, preserved across topology
+    /// and transport changes.
+    pub fn link(&self, peer: EndpointId, protocol: &str) -> Arc<PeerLink> {
+        let mut state = self.state();
+        let key = (*peer.as_bytes(), protocol.to_owned());
+        if let Some(link) = state.links.get(&key) {
+            return link.clone();
+        }
+        let (inbox_tx, inbox_rx) = mpsc::channel(INBOX);
+        let (closed, _) = watch::channel(false);
+        let link = Arc::new(PeerLink {
+            network: self.network,
+            peer,
+            tag: envelope::protocol_tag(protocol),
+            plane: state.plane.clone(),
+            inbox_tx,
+            inbox_rx: tokio::sync::Mutex::new(inbox_rx),
+            closed,
+        });
+        state.links.insert(key, link.clone());
+        state.publish();
+        link
+    }
+
+    /// Whether the protocol already owns its end-to-end channel to a peer.
+    pub fn has_link(&self, peer: EndpointId, protocol: &str) -> bool {
+        self.state()
+            .links
+            .contains_key(&(*peer.as_bytes(), protocol.to_owned()))
+    }
+
+    /// Installs a raw transport link and starts its independent ingress reader.
+    ///
+    /// A neighbour may have one link per kind of transport; installing a
+    /// second of the same kind replaces the first.
+    pub fn set_direct(&self, peer: EndpointId, kind: &str, raw: SharedLink) {
+        if raw.network() != self.network || raw.peer() != peer {
+            return;
+        }
+        let mut state = self.state();
+        let id = *peer.as_bytes();
+        let plane = state.plane.clone();
+        let changed = self.changed.clone();
+        let reader_link = raw.clone();
+        let reader = tokio::spawn(async move {
+            let mut batch = 0;
+            while let Some(frame) = reader_link.recv().await {
+                plane.receive(id, frame);
+                batch += 1;
+                if batch == 64 {
+                    tokio::task::yield_now().await;
+                    batch = 0;
+                }
+            }
+            changed.notify_one();
+        });
+        let links = state.raw.entry(id).or_default();
+        if let Some(position) = links.iter().position(|known| known.kind == kind) {
+            links.remove(position).reader.abort();
+        }
+        links.push(RawLink {
+            kind: kind.to_owned(),
+            link: raw,
+            reader,
+        });
+        state.publish();
+    }
+
+    /// Removes one physical link without tearing down end-to-end protocol state.
+    pub fn clear_direct(&self, peer: EndpointId, kind: &str) {
+        let mut state = self.state();
+        let mut changed = false;
+        if let Some(links) = state.raw.get_mut(peer.as_bytes()) {
+            if let Some(position) = links.iter().position(|known| known.kind == kind) {
+                links.remove(position).reader.abort();
+                changed = true;
+            }
+            if links.is_empty() {
+                state.raw.remove(peer.as_bytes());
+            }
+        }
+        if changed {
+            state.publish();
+        }
+    }
+
+    /// Retires data channels without removing an authenticated peer's topology.
+    pub fn clear_peer_links(&self, peer: EndpointId) {
+        let mut state = self.state();
+        if let Some(links) = state.raw.remove(peer.as_bytes()) {
+            for raw in links {
+                raw.link.close();
+                raw.reader.abort();
+            }
+        }
+        let gone: Vec<_> = state
+            .links
+            .keys()
+            .filter(|(id, _)| id == peer.as_bytes())
+            .cloned()
+            .collect();
+        for key in gone {
+            if let Some(link) = state.links.remove(&key) {
+                link.close();
+            }
+        }
+        state.publish();
+    }
+
+    /// Revokes all links and topology involving a departed authenticated peer.
+    pub fn remove_peer(&self, peer: EndpointId) {
+        self.clear_peer_links(peer);
+        let mut state = self.state();
+        state.members.remove(peer.as_bytes());
+        state.graph.remove(peer.as_bytes());
+        for neighbors in state.graph.values_mut() {
+            neighbors.retain(|id| id != peer.as_bytes());
+        }
+        state.publish();
+    }
+
+    /// Stops every reader and releases all transport handles.
+    pub fn close(&self) {
+        let mut state = self.state();
+        state.stop_readers();
+        state.raw.clear();
+        for link in state.links.values() {
+            link.close();
+        }
+        state.links.clear();
+        state.graph.clear();
+        state.members.clear();
+        state.plane.table.store(Arc::default());
+    }
+}
+
+/// A protocol's end-to-end channel, independent of its current next hop.
+#[derive(Debug)]
+pub struct PeerLink {
+    network: NetworkId,
+    peer: EndpointId,
+    tag: u64,
+    plane: Arc<Plane>,
+    inbox_tx: mpsc::Sender<Bytes>,
+    inbox_rx: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    closed: watch::Sender<bool>,
+}
+
+impl PeerLink {
+    fn close(&self) {
+        self.closed.send_replace(true);
+    }
+}
+
+impl PacketLink for PeerLink {
+    fn network(&self) -> NetworkId {
+        self.network
+    }
+    fn peer(&self) -> EndpointId {
+        self.peer
+    }
+    fn max_datagram_size(&self) -> usize {
+        MAX_DATA_DATAGRAM - RELAY_OVERHEAD
+    }
+    fn send(&self, payload: Bytes) -> Result<(), TransportError> {
+        self.send_flow(payload, 0)
+    }
+    fn send_flow(&self, payload: Bytes, flow: FlowId) -> Result<(), TransportError> {
+        if self.is_closed() {
+            return Err(TransportError::Closed);
+        }
+        if payload.len() > self.max_datagram_size() {
+            return Err(TransportError::TooLarge {
+                size: payload.len(),
+                limit: self.max_datagram_size(),
+            });
+        }
+        let table = self.plane.table.load();
+        let route = table
+            .routes
+            .get(self.peer.as_bytes())
+            .ok_or(TransportError::Closed)?;
+        let hop = route
+            .select(&self.plane.local, flow)
+            .ok_or(TransportError::Closed)?;
+        hop.send(envelope::encode(
+            self.plane.local,
+            *self.peer.as_bytes(),
+            flow,
+            self.tag,
+            &payload,
+        ))?;
+        if route.hops > 1 {
+            self.plane.tally.sent_via.fetch_add(1, Ordering::Relaxed);
+        }
+        Ok(())
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut closed = self.closed.subscribe();
+            if *closed.borrow_and_update() {
+                return None;
+            }
+            let mut inbox = self.inbox_rx.lock().await;
+            tokio::select! { biased; _ = closed.changed() => None, frame = inbox.recv() => frame }
+        })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(async move {
+            let mut closed = self.closed.subscribe();
+            if !*closed.borrow_and_update() {
+                let _ = closed.changed().await;
+            }
+        })
+    }
+    fn is_closed(&self) -> bool {
+        *self.closed.borrow()
+    }
+    fn path_description(&self) -> String {
+        let table = self.plane.table.load();
+        match table.routes.get(self.peer.as_bytes()) {
+            Some(route) if route.hops == 1 => route.next[0]
+                .first_open()
+                .map_or_else(|| "unreachable".into(), |link| link.path_description()),
+            Some(route) => format!(
+                "relay {} hops via {} ({} equal paths)",
+                route.hops,
+                hex::encode(route.next[0].peer),
+                route.next.len()
+            ),
+            None => "unreachable".into(),
+        }
+    }
+}
+
+#[cfg(test)]
+#[path = "relay_tests.rs"]
+mod tests;

crates/tsunagi-core/src/dataplane/relay_tests.rs

@@ -0,0 +1,418 @@
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+use super::*;
+use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+use std::sync::atomic::AtomicBool;
+
+fn network() -> NetworkId {
+    NetworkKeys::derive(
+        &NetworkName::new("routing-tests").unwrap(),
+        &NetworkSecret::from_bytes([3; 32]).unwrap(),
+    )
+    .network_id()
+}
+fn peer(seed: u8) -> EndpointId {
+    iroh::SecretKey::from_bytes(&[seed; 32]).public()
+}
+fn id(seed: u8) -> PeerId {
+    *peer(seed).as_bytes()
+}
+fn tag() -> u64 {
+    envelope::protocol_tag("ip")
+}
+
+#[derive(Debug)]
+struct Wire {
+    peer: EndpointId,
+    sent: Mutex<Vec<Bytes>>,
+    feed: mpsc::Sender<Bytes>,
+    inbound: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    dead: AtomicBool,
+    count: AtomicU64,
+    record: bool,
+}
+impl Wire {
+    fn new(seed: u8, record: bool) -> Arc<Self> {
+        let (feed, inbound) = mpsc::channel(32);
+        Arc::new(Self {
+            peer: peer(seed),
+            sent: Mutex::default(),
+            feed,
+            inbound: tokio::sync::Mutex::new(inbound),
+            dead: AtomicBool::new(false),
+            count: AtomicU64::new(0),
+            record,
+        })
+    }
+    fn take(&self) -> Vec<Bytes> {
+        std::mem::take(&mut self.sent.lock().unwrap())
+    }
+}
+impl PacketLink for Wire {
+    fn network(&self) -> NetworkId {
+        network()
+    }
+    fn peer(&self) -> EndpointId {
+        self.peer
+    }
+    fn max_datagram_size(&self) -> usize {
+        MAX_DATA_DATAGRAM
+    }
+    fn send(&self, frame: Bytes) -> Result<(), TransportError> {
+        if self.is_closed() {
+            return Err(TransportError::Closed);
+        }
+        self.count.fetch_add(1, Ordering::Relaxed);
+        if self.record {
+            self.sent.lock().unwrap().push(frame);
+        }
+        Ok(())
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move { self.inbound.lock().await.recv().await })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(std::future::pending())
+    }
+    fn is_closed(&self) -> bool {
+        self.dead.load(Ordering::Relaxed)
+    }
+    fn path_description(&self) -> String {
+        "test-direct".into()
+    }
+}
+
+fn topology(hub: &RelayHub, edges: &[(u8, &[u8])], members: &[u8]) {
+    hub.set_topology(
+        edges
+            .iter()
+            .map(|(a, bs)| (id(*a), bs.iter().map(|b| id(*b)).collect()))
+            .collect(),
+        members.iter().map(|p| id(*p)).collect(),
+    );
+}
+
+#[tokio::test]
+async fn ecmp_is_per_flow_direct_wins_and_existing_tunnel_survives_changes() {
+    let hub = RelayHub::new(network(), peer(1));
+    let b = Wire::new(2, true);
+    let c = Wire::new(3, true);
+    hub.set_direct(peer(2), "ip", b.clone());
+    hub.set_direct(peer(3), "ip", c.clone());
+    topology(&hub, &[(2, &[4]), (3, &[4])], &[1, 2, 3, 4]);
+    let link = hub.link(peer(4), "ip");
+    let mut chosen = HashMap::new();
+    for _ in 0..8 {
+        for flow in 0..100 {
+            link.send_flow(Bytes::from_static(b"ciphertext"), flow)
+                .unwrap();
+        }
+        for (hop, wire) in [(2, &b), (3, &c)] {
+            for frame in wire.take() {
+                let header = envelope::decode(&frame).unwrap();
+                assert_eq!(header.source, id(1));
+                assert_eq!(header.destination, id(4));
+                if let Some(old) = chosen.insert(header.flow, hop) {
+                    assert_eq!(old, hop);
+                }
+            }
+        }
+    }
+    assert_eq!(
+        chosen.values().copied().collect::<HashSet<_>>(),
+        HashSet::from([2, 3])
+    );
+    let direct = Wire::new(4, true);
+    hub.set_direct(peer(4), "ip", direct.clone());
+    link.send_flow(Bytes::from_static(b"direct"), 8).unwrap();
+    assert_eq!(direct.take().len(), 1);
+    assert!(b.take().is_empty() && c.take().is_empty());
+    hub.clear_direct(peer(4), "ip");
+    b.dead.store(true, Ordering::Relaxed);
+    for flow in 0..100 {
+        link.send_flow(Bytes::from_static(b"fallback"), flow)
+            .unwrap();
+    }
+    assert_eq!(c.take().len(), 100);
+    hub.clear_direct(peer(2), "ip");
+    topology(&hub, &[(3, &[])], &[1, 2, 3, 4]);
+    assert!(link.send(Bytes::new()).is_err());
+    assert!(!link.is_closed());
+    assert!(Arc::ptr_eq(&link, &hub.link(peer(4), "ip")));
+    hub.remove_peer(peer(4));
+    assert!(link.is_closed());
+    tokio::time::timeout(std::time::Duration::from_secs(1), link.closed())
+        .await
+        .unwrap();
+    assert!(link.recv().await.is_none());
+    hub.close();
+}
+
+#[tokio::test]
+async fn transport_reader_forwards_without_any_plugin_reader_and_isolates_protocols() {
+    let hub = RelayHub::new(network(), peer(2));
+    let a = Wire::new(1, true);
+    let c = Wire::new(3, true);
+    hub.set_direct(peer(1), "ip", a.clone());
+    hub.set_direct(peer(3), "ip", c.clone());
+    topology(&hub, &[(3, &[4])], &[1, 2, 3, 4]);
+    let frame = envelope::encode(id(1), id(4), 91, tag(), b"end-to-end encrypted");
+    let pointer = frame.as_ptr();
+    a.feed.send(frame).await.unwrap();
+    tokio::time::timeout(std::time::Duration::from_secs(1), async {
+        while hub.counters().forwarded == 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .unwrap();
+    let sent = c.take();
+    assert_eq!(sent[0].as_ptr(), pointer);
+    assert_eq!(
+        envelope::decode(&sent[0]).unwrap().remaining,
+        ROUTING_HOP_LIMIT - 1
+    );
+    assert_eq!(&sent[0][RELAY_OVERHEAD..], b"end-to-end encrypted");
+    assert_eq!(hub.counters().forwarded, 1);
+    hub.close();
+}
+
+#[tokio::test]
+async fn a_frame_is_delivered_only_to_the_protocol_it_names() {
+    let hub = RelayHub::new(network(), peer(2));
+    hub.set_direct(peer(1), "ip", Wire::new(1, true));
+    let ip = hub.link(peer(1), "ip");
+    let other = hub.link(peer(1), "other-ip");
+    let plane = ip.plane.clone();
+    plane.receive(id(1), envelope::encode(id(1), id(2), 0, tag(), b"for ip"));
+    plane.receive(
+        id(1),
+        envelope::encode(
+            id(1),
+            id(2),
+            0,
+            envelope::protocol_tag("other-ip"),
+            b"for other",
+        ),
+    );
+    plane.receive(
+        id(1),
+        envelope::encode(
+            id(1),
+            id(2),
+            0,
+            envelope::protocol_tag("nobody-runs-this"),
+            b"for nobody",
+        ),
+    );
+    assert_eq!(ip.recv().await.unwrap(), Bytes::from_static(b"for ip"));
+    assert_eq!(
+        other.recv().await.unwrap(),
+        Bytes::from_static(b"for other")
+    );
+    assert_eq!(hub.counters().dropped_unknown, 1);
+    hub.close();
+}
+
+/// A middle member joins two neighbours over different transports, and the
+/// two ends share a protocol the middle has never heard of.
+#[tokio::test]
+async fn a_tunnel_crosses_a_middle_member_whose_links_are_of_different_kinds() {
+    let a = RelayHub::new(network(), peer(1));
+    let c = RelayHub::new(network(), peer(3));
+    let b = RelayHub::new(network(), peer(2));
+    let a_to_c = Wire::new(3, true);
+    let c_to_a = Wire::new(1, true);
+    let c_to_b = Wire::new(2, true);
+    let b_to_c = Wire::new(3, true);
+    a.set_direct(peer(3), "wg", a_to_c.clone());
+    c.set_direct(peer(1), "wg", c_to_a);
+    c.set_direct(peer(2), "wg-quic", c_to_b.clone());
+    b.set_direct(peer(3), "wg-quic", b_to_c.clone());
+    topology(&a, &[(3, &[1, 2])], &[1, 2, 3]);
+    topology(&b, &[(3, &[1, 2])], &[1, 2, 3]);
+    topology(&c, &[(1, &[3]), (2, &[3])], &[1, 2, 3]);
+
+    let outgoing = a.link(peer(2), "tunnel");
+    let incoming = b.link(peer(1), "tunnel");
+    outgoing
+        .send_flow(Bytes::from_static(b"sealed"), 5)
+        .unwrap();
+    let frame = a_to_c.take().remove(0);
+
+    // The middle forwards it onto its other kind of link without looking
+    // inside.
+    c.state().plane.receive(id(1), frame);
+    let frame = c_to_b.take().remove(0);
+    assert_eq!(c.counters().forwarded, 1);
+
+    b.state().plane.receive(id(3), frame);
+    assert_eq!(
+        incoming.recv().await.unwrap(),
+        Bytes::from_static(b"sealed")
+    );
+    assert_eq!(b.counters().received_via, 1);
+    assert_eq!(
+        b.route(peer(1)),
+        Some(RouteSummary {
+            hops: 2,
+            via: peer(3)
+        })
+    );
+    a.close();
+    b.close();
+    c.close();
+}
+
+#[tokio::test]
+async fn a_neighbour_is_reached_over_the_preferred_kind_and_falls_back_when_it_is_lost() {
+    let hub = RelayHub::new(network(), peer(1));
+    hub.set_priority(vec!["wg".into(), "tcp-tls".into(), "wg-quic".into()]);
+    let slow = Wire::new(2, true);
+    let fast = Wire::new(2, true);
+    hub.set_direct(peer(2), "wg-quic", slow.clone());
+    let link = hub.link(peer(2), "tunnel");
+    link.send(Bytes::from_static(b"one")).unwrap();
+    assert_eq!(slow.take().len(), 1);
+    assert_eq!(hub.direct_kinds(peer(2)), ["wg-quic"]);
+
+    // A better kind appears: traffic moves to it with nothing to redo.
+    hub.set_direct(peer(2), "wg", fast.clone());
+    assert_eq!(hub.direct_kinds(peer(2)), ["wg", "wg-quic"]);
+    link.send(Bytes::from_static(b"two")).unwrap();
+    assert_eq!(fast.take().len(), 1);
+    assert!(slow.take().is_empty());
+
+    // It dies: the worse one carries on, and the logical link is untouched.
+    fast.dead.store(true, Ordering::Relaxed);
+    link.send(Bytes::from_static(b"three")).unwrap();
+    assert_eq!(slow.take().len(), 1);
+    assert!(!link.is_closed());
+    assert_eq!(
+        hub.route(peer(2)),
+        Some(RouteSummary {
+            hops: 1,
+            via: peer(2)
+        })
+    );
+
+    hub.clear_direct(peer(2), "wg-quic");
+    hub.clear_direct(peer(2), "wg");
+    assert!(!hub.reachable(peer(2)));
+    assert!(Arc::ptr_eq(&link, &hub.link(peer(2), "tunnel")));
+    hub.close();
+}
+
+#[tokio::test]
+async fn local_delivery_is_bounded_and_malformed_or_unknown_sources_are_dropped() {
+    let hub = RelayHub::new(network(), peer(2));
+    let a = Wire::new(1, true);
+    hub.set_direct(peer(1), "ip", a);
+    let link = hub.link(peer(1), "ip");
+    let plane = link.plane.clone();
+    for _ in 0..INBOX + 1 {
+        plane.receive(id(1), envelope::encode(id(1), id(2), 0, tag(), b"ip"));
+    }
+    assert_eq!(hub.counters().dropped_congested, 1);
+    for _ in 0..INBOX {
+        assert_eq!(link.recv().await.unwrap(), Bytes::from_static(b"ip"));
+    }
+    plane.receive(id(1), Bytes::from_static(b"broken"));
+    plane.receive(
+        id(1),
+        envelope::encode(id(9), id(2), 0, tag(), b"unknown member"),
+    );
+    assert_eq!(hub.counters().dropped_unknown, 2);
+    plane.receive(
+        id(1),
+        envelope::encode(id(1), id(9), 0, tag(), b"unknown destination"),
+    );
+    assert_eq!(hub.counters().dropped_no_link, 1);
+    hub.close();
+}
+
+#[tokio::test]
+async fn inconsistent_tables_cannot_loop_beyond_the_hop_limit() {
+    let a = RelayHub::new(network(), peer(1));
+    let b = RelayHub::new(network(), peer(2));
+    let ab = Wire::new(2, true);
+    let ba = Wire::new(1, true);
+    a.set_direct(peer(2), "ip", ab.clone());
+    b.set_direct(peer(1), "ip", ba.clone());
+    // Deliberately inconsistent snapshots while updates are in flight.
+    topology(&a, &[(2, &[4])], &[1, 2, 3, 4]);
+    topology(&b, &[(1, &[4])], &[1, 2, 3, 4]);
+    let ap = a.link(peer(4), "ip").plane.clone();
+    let bp = b.link(peer(4), "ip").plane.clone();
+    let mut frame = envelope::encode(id(3), id(4), 7, tag(), b"transit");
+    for turn in 0..ROUTING_HOP_LIMIT {
+        let (plane, incoming, wire) = if turn % 2 == 0 {
+            (&ap, id(2), &ab)
+        } else {
+            (&bp, id(1), &ba)
+        };
+        plane.receive(incoming, frame);
+        let sent = wire.take();
+        if turn == ROUTING_HOP_LIMIT - 1 {
+            assert!(sent.is_empty());
+            break;
+        }
+        assert_eq!(sent.len(), 1);
+        frame = sent.into_iter().next().unwrap();
+    }
+    assert_eq!(
+        a.counters().forwarded + b.counters().forwarded,
+        u64::from(ROUTING_HOP_LIMIT - 1)
+    );
+    assert_eq!(
+        a.counters().dropped_hop_limit + b.counters().dropped_hop_limit,
+        1
+    );
+    a.close();
+    b.close();
+}
+
+/// Measures the actual synchronous transit routine, including header checks,
+/// snapshot lookup, ECMP, in-place TTL update and mock transport submission.
+/// Packet construction, encryption and socket I/O are outside this measurement.
+#[tokio::test]
+#[ignore = "manual release-mode forwarding microbenchmark"]
+async fn forwarding_benchmark() {
+    let hub = RelayHub::new(network(), peer(2));
+    let wire = Wire::new(3, false);
+    hub.set_direct(peer(3), "ip", wire.clone());
+    let alternate = Wire::new(5, false);
+    hub.set_direct(peer(5), "ip", alternate.clone());
+    topology(&hub, &[(3, &[4]), (5, &[4])], &[1, 2, 3, 4, 5]);
+    let plane = hub.link(peer(4), "ip").plane.clone();
+    let source = id(1);
+    let destination = id(4);
+    let incoming = id(1);
+    for size in [64, 1280] {
+        let mut elapsed = std::time::Duration::ZERO;
+        const COUNT: usize = 10000;
+        const BATCHES: usize = 100;
+        for _ in 0..BATCHES {
+            let frames: Vec<_> = (0..COUNT)
+                .map(|flow| {
+                    envelope::encode(source, destination, flow as u64, tag(), &vec![42; size])
+                })
+                .collect();
+            let start = std::time::Instant::now();
+            for frame in frames {
+                plane.receive(incoming, std::hint::black_box(frame));
+            }
+            elapsed += start.elapsed();
+        }
+        let count = (COUNT * BATCHES) as f64;
+        println!(
+            "transit {size} B: {:.1} ns/packet, {:.2} Mpps (1M packets, two equal paths)",
+            elapsed.as_nanos() as f64 / count,
+            count / elapsed.as_secs_f64() / 1e6
+        );
+    }
+    assert_eq!(
+        wire.count.load(Ordering::Relaxed) + alternate.count.load(Ordering::Relaxed),
+        2000000
+    );
+    hub.close();
+}

crates/tsunagi-core/src/dataplane/routing/envelope.rs

@@ -0,0 +1,104 @@
+//! Fixed header. Transit changes one byte in an owned receive buffer.
+use super::{FlowId, PeerId};
+use crate::config::{MAX_DATA_DATAGRAM, ROUTING_HOP_LIMIT};
+use bytes::{BufMut, Bytes, BytesMut};
+
+pub(crate) const HEADER: usize = 2 + 32 + 32 + 8 + 8;
+const VERSION: u8 = 2;
+
+#[derive(Debug, PartialEq, Eq)]
+pub(crate) struct Header {
+    pub source: PeerId,
+    pub destination: PeerId,
+    pub remaining: u8,
+    pub flow: FlowId,
+    /// Which end-to-end protocol the payload belongs to, whatever links carry it.
+    pub protocol: u64,
+}
+
+/// A stable 64-bit tag for a protocol id (FNV-1a), the same on every member.
+pub(crate) fn protocol_tag(name: &str) -> u64 {
+    name.bytes().fold(0xcbf2_9ce4_8422_2325, |hash, byte| {
+        (hash ^ u64::from(byte)).wrapping_mul(0x0000_0100_0000_01b3)
+    })
+}
+
+pub(crate) fn decode(bytes: &[u8]) -> Option<Header> {
+    if bytes.len() < HEADER
+        || bytes.len() > MAX_DATA_DATAGRAM
+        || bytes[0] != VERSION
+        || bytes[1] == 0
+        || bytes[1] > ROUTING_HOP_LIMIT
+    {
+        return None;
+    }
+    Some(Header {
+        source: bytes[2..34].try_into().ok()?,
+        destination: bytes[34..66].try_into().ok()?,
+        remaining: bytes[1],
+        flow: u64::from_be_bytes(bytes[66..74].try_into().ok()?),
+        protocol: u64::from_be_bytes(bytes[74..82].try_into().ok()?),
+    })
+}
+
+pub(crate) fn encode(
+    source: PeerId,
+    destination: PeerId,
+    flow: FlowId,
+    protocol: u64,
+    payload: &[u8],
+) -> Bytes {
+    let mut bytes = BytesMut::with_capacity(HEADER + payload.len());
+    bytes.put_u8(VERSION);
+    bytes.put_u8(ROUTING_HOP_LIMIT);
+    bytes.extend_from_slice(&source);
+    bytes.extend_from_slice(&destination);
+    bytes.put_u64(flow);
+    bytes.put_u64(protocol);
+    bytes.extend_from_slice(payload);
+    bytes.freeze()
+}
+
+pub(crate) fn decrement(bytes: Bytes) -> Bytes {
+    let mut bytes = match bytes.try_into_mut() {
+        Ok(bytes) => bytes,
+        Err(bytes) => BytesMut::from(bytes.as_ref()),
+    };
+    // Called only after decode and after excluding remaining <= 1.
+    bytes[1] -= 1;
+    bytes.freeze()
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[test]
+    fn transit_reuses_the_buffer_and_rejects_invalid_headers() {
+        let frame = encode([1; 32], [2; 32], 17, protocol_tag("wg"), &[42; 1280]);
+        let address = frame.as_ptr();
+        let frame = decrement(frame);
+        assert_eq!(
+            frame.as_ptr(),
+            address,
+            "no payload copy for an owned buffer"
+        );
+        let header = decode(&frame).unwrap();
+        assert_eq!(header.remaining, ROUTING_HOP_LIMIT - 1);
+        assert_eq!(header.protocol, protocol_tag("wg"));
+        assert_ne!(protocol_tag("wg"), protocol_tag("wg-quic"));
+        assert_eq!(&frame[HEADER..], &[42; 1280]);
+        for length in 0..HEADER {
+            assert!(decode(&frame[..length]).is_none());
+        }
+        for ttl in [0, ROUTING_HOP_LIMIT + 1] {
+            let mut bad = frame.to_vec();
+            bad[1] = ttl;
+            assert!(decode(&bad).is_none());
+        }
+        let mut bad = frame.to_vec();
+        bad[0] = 99;
+        assert!(decode(&bad).is_none());
+        assert!(decode(&vec![1; MAX_DATA_DATAGRAM + 1]).is_none());
+    }
+}

crates/tsunagi-core/src/dataplane/routing/flow.rs

@@ -0,0 +1,125 @@
+//! Optional IP flow classifier for IP plugins, called before encryption.
+//! The router itself treats this identifier and the payload as opaque.
+
+use super::FlowId;
+
+fn hash(parts: &[&[u8]]) -> FlowId {
+    let mut value = 0xcbf29ce484222325u64;
+    for part in parts {
+        for byte in *part {
+            value = (value ^ u64::from(*byte)).wrapping_mul(0x100000001b3);
+        }
+    }
+    value
+}
+
+/// Stable directional IP flow hash, independent of TCP sequence numbers,
+/// checksums and packet contents. Fragmented traffic uses addresses/protocol
+/// instead of ports: even non-initial fragments and subsequent fragmented
+/// datagrams stay together. This deliberately coalesces fragmented flows.
+/// Malformed packets get a coarse hash; parsing is bounded and never allocates.
+pub fn ip_flow(packet: &[u8]) -> FlowId {
+    match packet.first().map(|b| b >> 4) {
+        Some(4) if packet.len() >= 20 => {
+            let addresses = &packet[12..20];
+            let protocol = &packet[9..10];
+            let offset = usize::from(packet[0] & 15) * 4;
+            if u16::from_be_bytes([packet[6], packet[7]]) & 0x3fff != 0 {
+                return hash(&[addresses, protocol]);
+            }
+            if offset >= 20
+                && matches!(packet[9], 6 | 17)
+                && let Some(ports) = packet.get(offset..offset + 4)
+            {
+                return hash(&[addresses, protocol, ports]);
+            }
+            hash(&[addresses, protocol])
+        }
+        Some(6) if packet.len() >= 40 => {
+            let addresses = &packet[8..40];
+            let mut protocol = packet[6];
+            let mut offset = 40;
+            for _ in 0..8 {
+                match protocol {
+                    44 => {
+                        if let Some(fragment) = packet.get(offset..offset + 8) {
+                            return hash(&[addresses, &fragment[..1]]);
+                        }
+                        break;
+                    }
+                    0 | 43 | 60 | 51 => {
+                        let Some(header) = packet.get(offset..offset + 2) else {
+                            break;
+                        };
+                        let size = if protocol == 51 {
+                            (usize::from(header[1]) + 2) * 4
+                        } else {
+                            (usize::from(header[1]) + 1) * 8
+                        };
+                        protocol = header[0];
+                        offset += size;
+                    }
+                    6 | 17 => {
+                        if let Some(ports) = packet.get(offset..offset + 4) {
+                            return hash(&[addresses, &[protocol], ports]);
+                        }
+                        break;
+                    }
+                    _ => break,
+                }
+            }
+            hash(&[addresses, &[protocol]])
+        }
+        _ => 0,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn tcp_flow_ignores_packet_contents_but_distinguishes_ports() {
+        let mut packet = [0u8; 80];
+        packet[0] = 0x45;
+        packet[9] = 6;
+        packet[12..24].copy_from_slice(&[10, 0, 0, 1, 10, 0, 0, 2, 0, 22, 1, 2]);
+        let flow = ip_flow(&packet);
+        packet[4] = 123; // IP id changes for each unfragmented packet.
+        packet[24..].fill(47);
+        assert_eq!(flow, ip_flow(&packet));
+        packet[23] = 3;
+        assert_ne!(flow, ip_flow(&packet));
+        packet[6] = 0x20;
+        let fragment_flow = ip_flow(&packet);
+        packet[6] = 0;
+        packet[7] = 1;
+        packet[20..].fill(99);
+        assert_eq!(fragment_flow, ip_flow(&packet));
+        packet[4] = 97;
+        assert_eq!(fragment_flow, ip_flow(&packet));
+    }
+
+    #[test]
+    fn ipv6_extensions_and_fragments_are_bounded_and_stable() {
+        let mut packet = [0u8; 80];
+        packet[0] = 0x60;
+        packet[6] = 0;
+        packet[40] = 17;
+        packet[48..52].copy_from_slice(&[1, 2, 3, 4]);
+        let flow = ip_flow(&packet);
+        packet[52..].fill(5);
+        assert_eq!(flow, ip_flow(&packet));
+        packet[50] = 9;
+        assert_ne!(flow, ip_flow(&packet));
+        packet[6] = 44;
+        packet[40] = 6;
+        let flow = ip_flow(&packet);
+        packet[42] = 32;
+        packet[48..].fill(3);
+        assert_eq!(flow, ip_flow(&packet));
+        for len in 0..packet.len() {
+            let _ = ip_flow(&packet[..len]);
+        }
+    }
+}

crates/tsunagi-core/src/dataplane/routing/mod.rs

@@ -0,0 +1,131 @@
+//! Transport-independent routing over opaque identities. Tables are built on
+//! topology changes, never on the packet path. No WireGuard or iroh APIs here.
+pub(crate) mod envelope;
+pub mod flow;
+
+use std::collections::{BTreeSet, HashMap, VecDeque};
+
+/// Opaque authenticated identity; the transport adapter validates its keys.
+pub type PeerId = [u8; 32];
+/// Stable flow identifier, carried unchanged between transit routers.
+pub type FlowId = u64;
+
+/// A shortest route with all equal-cost first hops in stable order.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Route {
+    /// Number of transport links to the destination.
+    pub hops: u8,
+    /// Equal-cost first hops. Direct destinations have one entry.
+    pub next_hops: Box<[PeerId]>,
+}
+
+/// Immutable shortest-path table, replaced when connectivity changes.
+#[derive(Debug, Clone, Default)]
+pub struct RoutingTable {
+    routes: HashMap<PeerId, Route>,
+}
+
+impl RoutingTable {
+    /// Computes shortest directed paths bounded by `hop_limit`.
+    /// The caller supplies only authenticated, fresh, compatible edges.
+    pub fn build(local: PeerId, graph: &HashMap<PeerId, Vec<PeerId>>, hop_limit: u8) -> Self {
+        let mut distances = HashMap::from([(local, 0u8)]);
+        let mut first_hops: HashMap<PeerId, BTreeSet<PeerId>> = HashMap::new();
+        let mut pending = VecDeque::from([local]);
+        while let Some(node) = pending.pop_front() {
+            let distance = distances[&node];
+            if distance >= hop_limit {
+                continue;
+            }
+            let inherited = first_hops.get(&node).cloned().unwrap_or_default();
+            for &next in graph.get(&node).into_iter().flatten() {
+                let next_distance = distance + 1;
+                match distances.get(&next) {
+                    Some(&known) if known < next_distance => continue,
+                    None => {
+                        distances.insert(next, next_distance);
+                        pending.push_back(next);
+                    }
+                    _ => {}
+                }
+                let hops = first_hops.entry(next).or_default();
+                if node == local {
+                    hops.insert(next);
+                } else {
+                    hops.extend(inherited.iter().copied());
+                }
+            }
+        }
+        Self {
+            routes: first_hops
+                .into_iter()
+                .map(|(peer, hops)| {
+                    (
+                        peer,
+                        Route {
+                            hops: distances[&peer],
+                            next_hops: hops.into_iter().collect(),
+                        },
+                    )
+                })
+                .collect(),
+        }
+    }
+    /// Precomputed route to a destination.
+    pub fn get(&self, destination: &PeerId) -> Option<&Route> {
+        self.routes.get(destination)
+    }
+    /// All reachable destinations, excluding ourselves.
+    pub fn iter(&self) -> impl Iterator<Item = (&PeerId, &Route)> {
+        self.routes.iter()
+    }
+}
+
+/// Stable ECMP mixing, without random state or hashing the payload per hop.
+pub(crate) fn mix64(mut value: u64) -> u64 {
+    value ^= value >> 30;
+    value = value.wrapping_mul(0xbf58476d1ce4e5b9);
+    value ^= value >> 27;
+    value = value.wrapping_mul(0x94d049bb133111eb);
+    value ^ (value >> 31)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    fn p(n: u8) -> PeerId {
+        [n; 32]
+    }
+    #[test]
+    fn shortest_paths_keep_equal_cost_hops_and_prefer_direct_links() {
+        let mut graph = HashMap::from([
+            (p(1), vec![p(3), p(2)]),
+            (p(2), vec![p(4)]),
+            (p(3), vec![p(4)]),
+            (p(4), vec![p(5)]),
+            (p(5), vec![p(1)]),
+        ]);
+        let routes = RoutingTable::build(p(1), &graph, 16);
+        assert_eq!(routes.get(&p(5)).unwrap().hops, 3);
+        assert_eq!(&*routes.get(&p(5)).unwrap().next_hops, &[p(2), p(3)]);
+        assert!(routes.get(&p(1)).is_none());
+        assert!(routes.get(&p(9)).is_none());
+        graph.get_mut(&p(1)).unwrap().push(p(5));
+        assert_eq!(
+            RoutingTable::build(p(1), &graph, 16).get(&p(5)).unwrap(),
+            &Route {
+                hops: 1,
+                next_hops: Box::new([p(5)])
+            }
+        );
+    }
+    #[test]
+    fn topology_removal_and_hop_budget_remove_impossible_routes() {
+        let mut graph = HashMap::from([(p(1), vec![p(2)]), (p(2), vec![p(3)]), (p(3), vec![p(4)])]);
+        assert!(RoutingTable::build(p(1), &graph, 2).get(&p(4)).is_none());
+        assert!(RoutingTable::build(p(1), &graph, 3).get(&p(4)).is_some());
+        graph.remove(&p(2));
+        assert!(RoutingTable::build(p(1), &graph, 16).get(&p(4)).is_none());
+    }
+}

crates/tsunagi-core/src/dataplane/transport/fragments.rs

@@ -0,0 +1,234 @@
+//! Bounded fragmentation of opaque transport payloads, below peer relaying.
+//!
+//! Data ALPN v3: u64 packet ID, u32 total length, u32 byte offset (big endian),
+//! followed by payload. Each QUIC connection owns its IDs and reassembly state.
+//! There is no retransmission: one missing fragment loses one datagram only.
+
+use std::collections::{HashMap, VecDeque};
+
+use bytes::{BufMut, Bytes, BytesMut};
+use tokio::time::Instant;
+
+use crate::config::{
+    DATA_REASSEMBLY_TTL, DATA_RECENT_IDS, MAX_DATA_ASSEMBLIES, MAX_DATA_DATAGRAM,
+    MAX_DATA_FRAGMENTS, MAX_DATA_REASSEMBLY_BYTES,
+};
+
+pub(super) const HEADER: usize = 16;
+
+pub(super) fn encode(id: u64, total: usize, offset: usize, payload: &[u8]) -> Bytes {
+    let mut frame = BytesMut::with_capacity(HEADER + payload.len());
+    frame.put_u64(id);
+    frame.put_u32(total as u32);
+    frame.put_u32(offset as u32);
+    frame.extend_from_slice(payload);
+    frame.freeze()
+}
+
+#[derive(Debug)]
+struct Assembly {
+    started: Instant,
+    bytes: Vec<u8>,
+    ranges: Vec<(usize, usize)>,
+    received: usize,
+}
+
+#[derive(Debug, Default)]
+pub(super) struct Reassembler {
+    packets: HashMap<u64, Assembly>,
+    buffered: usize,
+    recent: VecDeque<u64>,
+}
+
+impl Reassembler {
+    fn remember(&mut self, id: u64) {
+        if self.recent.len() == DATA_RECENT_IDS {
+            self.recent.pop_front();
+        }
+        self.recent.push_back(id);
+    }
+
+    fn remove(&mut self, id: u64) -> Option<Assembly> {
+        let packet = self.packets.remove(&id)?;
+        self.buffered -= packet.bytes.len();
+        self.remember(id);
+        Some(packet)
+    }
+
+    pub(super) fn expire(&mut self, now: Instant) {
+        let expired: Vec<_> = self
+            .packets
+            .iter()
+            .filter(|(_, p)| now.duration_since(p.started) >= DATA_REASSEMBLY_TTL)
+            .map(|(id, _)| *id)
+            .collect();
+        for id in expired {
+            self.remove(id);
+        }
+    }
+
+    pub(super) fn push(&mut self, frame: Bytes, now: Instant) -> Option<Bytes> {
+        self.expire(now);
+        let id = u64::from_be_bytes(frame.get(..8)?.try_into().ok()?);
+        let total = u32::from_be_bytes(frame.get(8..12)?.try_into().ok()?) as usize;
+        let offset = u32::from_be_bytes(frame.get(12..HEADER)?.try_into().ok()?) as usize;
+        let payload = frame.get(HEADER..)?;
+        let end = offset.checked_add(payload.len())?;
+        if total > MAX_DATA_DATAGRAM
+            || end > total
+            || (payload.is_empty() && total != 0)
+            || self.recent.contains(&id)
+        {
+            return None;
+        }
+        if !self.packets.contains_key(&id) {
+            if offset == 0 && end == total {
+                self.remember(id);
+                return Some(frame.slice(HEADER..));
+            }
+            // The length is validated before allocating. Drop the oldest
+            // incomplete packet when the per-link byte or entry budget is full.
+            while self.packets.len() >= MAX_DATA_ASSEMBLIES
+                || self.buffered + total > MAX_DATA_REASSEMBLY_BYTES
+            {
+                let oldest = self
+                    .packets
+                    .iter()
+                    .min_by_key(|(_, p)| p.started)
+                    .map(|(id, _)| *id)?;
+                self.remove(oldest);
+            }
+            self.packets.insert(
+                id,
+                Assembly {
+                    started: now,
+                    bytes: vec![0; total],
+                    ranges: Vec::new(),
+                    received: 0,
+                },
+            );
+            self.buffered += total;
+        }
+        let packet = self.packets.get_mut(&id)?;
+        if packet.bytes.len() != total {
+            self.remove(id);
+            return None;
+        }
+        for &(start, stop) in &packet.ranges {
+            if start == offset && stop == end && packet.bytes[offset..end] == *payload {
+                return None; // a duplicate does not reset the expiry time
+            }
+            if offset < stop && end > start {
+                self.remove(id); // conflicting or overlapping fragments
+                return None;
+            }
+        }
+        if packet.ranges.len() == MAX_DATA_FRAGMENTS {
+            self.remove(id);
+            return None;
+        }
+        packet.bytes[offset..end].copy_from_slice(payload);
+        packet.ranges.push((offset, end));
+        packet.received += payload.len();
+        if packet.received == total {
+            return self.remove(id).map(|p| Bytes::from(p.bytes));
+        }
+        None
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+
+    #[test]
+    fn reordered_duplicates_and_changing_fragment_sizes_reassemble_once() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        let packet: Vec<u8> = (0..9000).map(|i| (i % 251) as u8).collect();
+        let mut parts = Vec::new();
+        let mut offset = 0;
+        for size in [1100, 700, 1100, 1100, 700, 1100, 1100, 1100, 1000] {
+            parts.push(encode(
+                7,
+                packet.len(),
+                offset,
+                &packet[offset..offset + size],
+            ));
+            offset += size;
+        }
+        assert_eq!(offset, packet.len());
+        for part in parts[1..].iter().rev() {
+            assert!(rx.push(part.clone(), now).is_none());
+            assert!(rx.push(part.clone(), now).is_none());
+        }
+        assert_eq!(rx.push(parts[0].clone(), now).unwrap().as_ref(), packet);
+        assert!(rx.push(parts[0].clone(), now).is_none());
+        assert_eq!(rx.buffered, 0);
+        assert!(rx.packets.is_empty());
+    }
+
+    #[test]
+    fn a_missing_fragment_never_blocks_the_next_packet_and_expires() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        assert!(rx.push(encode(1, 1280, 0, &[1; 800]), now).is_none());
+        assert_eq!(
+            rx.push(encode(2, 4, 0, b"next"), now).unwrap(),
+            &b"next"[..]
+        );
+        rx.expire(now + DATA_REASSEMBLY_TTL);
+        assert!(rx.packets.is_empty());
+        assert_eq!(rx.buffered, 0);
+        assert!(
+            rx.push(encode(1, 1280, 800, &[1; 480]), now + DATA_REASSEMBLY_TTL)
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn malformed_conflicting_and_oversized_fragments_are_bounded() {
+        let now = Instant::now();
+        let mut rx = Reassembler::default();
+        for len in 0..HEADER {
+            assert!(rx.push(Bytes::from(vec![0xff; len]), now).is_none());
+        }
+        assert!(
+            rx.push(encode(1, MAX_DATA_DATAGRAM + 1, 0, b"x"), now)
+                .is_none()
+        );
+        assert!(rx.push(encode(2, 100, 100, b"x"), now).is_none());
+        assert!(rx.push(encode(3, 100, 0, b""), now).is_none());
+        assert_eq!(rx.buffered, 0);
+        for (id, bad) in [
+            (4, encode(4, 100, 0, b"different")),
+            (5, encode(5, 100, 3, b"overlap")),
+            (6, encode(6, 101, 9, b"changed total")),
+        ] {
+            assert!(rx.push(encode(id, 100, 0, b"123456789"), now).is_none());
+            assert!(rx.push(bad, now).is_none());
+            assert_eq!(rx.buffered, 0);
+        }
+        for id in 100..400 {
+            assert!(
+                rx.push(encode(id, MAX_DATA_DATAGRAM, 0, b"x"), now)
+                    .is_none()
+            );
+            assert!(rx.buffered <= MAX_DATA_REASSEMBLY_BYTES);
+            assert!(rx.packets.len() <= MAX_DATA_ASSEMBLIES);
+            assert!(rx.recent.len() <= DATA_RECENT_IDS);
+        }
+    }
+
+    #[test]
+    fn excessive_fragment_counts_are_discarded() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        for offset in 0..=MAX_DATA_FRAGMENTS {
+            assert!(rx.push(encode(1, 1000, offset, b"x"), now).is_none());
+        }
+        assert!(rx.packets.is_empty());
+        assert_eq!(rx.buffered, 0);
+    }
+}

crates/tsunagi-core/src/dataplane/transport/health.rs

@@ -0,0 +1,350 @@
+//! Authenticated probes of the actual data link, independent of control RTT.
+//! Keys are negotiated privately for each physical link. No fixed wire magic.
+use super::{PacketLink, SharedLink, TransportError};
+use crate::dataplane::SharedPlugin;
+use crate::overlay::RoutingTable;
+use crate::{BoxFuture, NetworkId};
+use bytes::Bytes;
+use hmac::{Hmac, KeyInit, Mac};
+use iroh::EndpointId;
+use serde::{Deserialize, Serialize};
+use sha2::Sha256;
+use std::collections::VecDeque;
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+use tokio::sync::{mpsc, watch};
+
+#[derive(Clone, Serialize, Deserialize, zeroize::ZeroizeOnDrop)]
+pub(crate) struct Settings {
+    marker: [u8; 16],
+    key: [u8; 32],
+}
+impl Settings {
+    pub(crate) fn fresh() -> Self {
+        Self {
+            marker: rand::random(),
+            key: rand::random(),
+        }
+    }
+}
+impl std::fmt::Debug for Settings {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("HealthSettings([REDACTED])")
+    }
+}
+struct Samples {
+    pending: VecDeque<(u64, Instant)>,
+    recent: VecDeque<bool>,
+}
+impl Samples {
+    fn new() -> Self {
+        Self {
+            pending: VecDeque::new(),
+            recent: VecDeque::new(),
+        }
+    }
+    fn sent(&mut self, seq: u64, now: Instant) {
+        self.pending.push_back((seq, now));
+    }
+    fn acknowledge(&mut self, seq: u64) {
+        if let Some(i) = self.pending.iter().position(|(n, _)| *n == seq) {
+            self.pending.remove(i);
+            self.recent.push_back(true);
+        }
+    }
+    fn expire(&mut self, now: Instant) {
+        while self
+            .pending
+            .front()
+            .is_some_and(|(_, at)| now.duration_since(*at) >= Duration::from_secs(3))
+        {
+            self.pending.pop_front();
+            self.recent.push_back(false);
+        }
+        while self.recent.len() > 20 {
+            self.recent.pop_front();
+        }
+    }
+    fn degraded(&self) -> bool {
+        self.recent.len() >= 15
+            && self.recent.iter().filter(|ok| !**ok).count() * 100 >= self.recent.len() * 60
+    }
+}
+// This prefix is inside end-to-end encryption. It never appears on the UDP wire.
+const IP_MAGIC: &[u8; 16] = b"tsunagi-health01";
+pub(crate) fn ip_payload(packet: &[u8]) -> Option<&[u8]> {
+    (packet.len() >= 36 && packet[0] == 0x45 && packet[9] == 253 && &packet[20..36] == IP_MAGIC)
+        .then(|| &packet[36..])
+}
+/// Inject probes through the same encrypted tunnel as application IP packets.
+/// This matters for pure WG: raw UDP probes could pass DPI while WG is blocked.
+pub(crate) struct ProbeRoute {
+    pub(crate) plugin: SharedPlugin,
+    pub(crate) routes: Arc<RoutingTable>,
+    pub(crate) network: NetworkId,
+    pub(crate) peer: EndpointId,
+    pub(crate) max_probe_size: usize,
+}
+impl ProbeRoute {
+    fn send(&self, payload: Bytes) -> bool {
+        let Some((source, destination)) = self.routes.probe_addresses(self.network, self.peer)
+        else {
+            return false;
+        };
+        let mut packet = vec![0; 36];
+        packet[0] = 0x45;
+        packet[8] = 64;
+        packet[9] = 253;
+        packet[12..16].copy_from_slice(&source.octets());
+        packet[16..20].copy_from_slice(&destination.octets());
+        packet[20..36].copy_from_slice(IP_MAGIC);
+        let length = (packet.len() + payload.len()) as u16;
+        packet[2..4].copy_from_slice(&length.to_be_bytes());
+        let sum = packet[..20]
+            .as_chunks::<2>()
+            .0
+            .iter()
+            .map(|b| u32::from(u16::from_be_bytes(*b)))
+            .sum::<u32>();
+        let sum = (sum & 0xffff) + (sum >> 16);
+        packet[10..12].copy_from_slice(&(!(sum as u16)).to_be_bytes());
+        packet.extend_from_slice(&payload);
+        self.plugin
+            .carry(self.network, self.peer, Bytes::from(packet))
+    }
+}
+/// A monitored physical link. It forwards application bytes unchanged.
+pub(crate) struct HealthLink {
+    probes: mpsc::Sender<Bytes>,
+    raw: SharedLink,
+    settings: Mutex<Option<Settings>>,
+    rx: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    stop: watch::Sender<bool>,
+    reason: Mutex<Option<String>>,
+}
+impl std::fmt::Debug for HealthLink {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("HealthLink")
+            .field("raw", &self.raw)
+            .finish_non_exhaustive()
+    }
+}
+fn lock<T>(m: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match m.lock() {
+        Ok(g) => g,
+        Err(p) => p.into_inner(),
+    }
+}
+impl HealthLink {
+    pub(crate) fn new(
+        raw: SharedLink,
+        local: EndpointId,
+        settings: Option<Settings>,
+        route: Option<ProbeRoute>,
+    ) -> Arc<Self> {
+        let (tx, rx) = mpsc::channel(64);
+        let (probe_tx, mut probe_rx) = mpsc::channel(64);
+        let (stop, _) = watch::channel(false);
+        let link = Arc::new(Self {
+            probes: probe_tx,
+            raw: raw.clone(),
+            settings: Mutex::new(settings),
+            rx: tokio::sync::Mutex::new(rx),
+            stop,
+            reason: Mutex::new(None),
+        });
+        let weak = Arc::downgrade(&link);
+        let mut stopped = link.stop.subscribe();
+        tokio::spawn(async move {
+            let mut interval = tokio::time::interval(Duration::from_secs(1));
+            interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+            let mut samples = Samples::new();
+            let mut sequence = rand::random::<u64>();
+            let mut since: Option<Instant> = None;
+            let mut marker = None;
+            loop {
+                tokio::select! {
+                    biased;
+                    _=stopped.changed()=>break,
+                    _=interval.tick()=>{
+                        let Some(link)=weak.upgrade()else{break;};let Some(settings)=lock(&link.settings).clone()else{continue;};
+                        if marker != Some(settings.marker) { marker = Some(settings.marker); samples = Samples::new(); since = None; }
+                        let start=*since.get_or_insert_with(Instant::now);let now=Instant::now();samples.expire(now);
+                        if now.duration_since(start)>=Duration::from_secs(20) && samples.degraded() {
+                            *lock(&link.reason)=Some("data delivery checks: at least 60% loss over the last 20 probes".into());break;
+                        }
+                        sequence=sequence.wrapping_add(1);
+                        let size=route.as_ref().map_or(raw.max_datagram_size(),|route|route.max_probe_size);
+                        let packet=probe(&settings,local,1,sequence,size.min(1000));
+                        if let Some(route)=&route { if route.send(packet) {samples.sent(sequence,now);} }
+                        else { let _=raw.send(packet); samples.sent(sequence,now); }
+                    }
+                    packet=raw.recv()=>{
+                        let Some(packet)=packet else{break;};let Some(link)=weak.upgrade()else{break;};
+                        let settings=lock(&link.settings).clone();
+                        if route.is_none() && let Some(settings)=settings && packet.starts_with(&settings.marker) {
+                            if let Some((kind,seq))=verify(&settings,raw.peer(),&packet) {
+                                if kind==1 {let _=raw.send(probe(&settings,local,2,seq,packet.len()));}
+                                else if kind==2 {samples.acknowledge(seq);}
+                            }
+                        } else {let _=tx.try_send(packet);}
+                    }
+                    packet=probe_rx.recv(),if route.is_some()=>{
+                        let Some(packet)=packet else {break;};let Some(link)=weak.upgrade()else{break;};
+                        let Some(settings)=lock(&link.settings).clone()else{continue;};
+                        if let Some((kind,seq))=verify(&settings,raw.peer(),&packet) {
+                            if kind==1 { if let Some(route)=&route { route.send(probe(&settings,local,2,seq,packet.len().min(route.max_probe_size))); } }
+                            else if kind==2 {samples.acknowledge(seq);}
+                        }
+                    }
+
+                }
+            }
+            raw.close();
+            if let Some(link) = weak.upgrade() {
+                link.stop.send_replace(true);
+            }
+        });
+        link
+    }
+    pub(crate) fn receive_probe(&self, packet: Bytes) {
+        let _ = self.probes.try_send(packet);
+    }
+    pub(crate) fn configure(&self, settings: Settings) {
+        *lock(&self.settings) = Some(settings);
+    }
+}
+fn probe(settings: &Settings, sender: EndpointId, kind: u8, seq: u64, size: usize) -> Bytes {
+    let mut bytes = vec![0; size.max(57)];
+    bytes[..16].copy_from_slice(&settings.marker);
+    bytes[16] = kind;
+    bytes[17..25].copy_from_slice(&seq.to_be_bytes());
+    rand::fill(&mut bytes[25..size.max(57) - 32]);
+    let tail = bytes.len() - 32;
+    if let Ok(mut mac) = <Hmac<Sha256> as KeyInit>::new_from_slice(&settings.key) {
+        mac.update(sender.as_bytes());
+        mac.update(&bytes[..tail]);
+        bytes[tail..].copy_from_slice(&mac.finalize().into_bytes());
+    }
+    Bytes::from(bytes)
+}
+fn verify(settings: &Settings, sender: EndpointId, bytes: &[u8]) -> Option<(u8, u64)> {
+    if bytes.len() < 57 || !bytes.starts_with(&settings.marker) {
+        return None;
+    }
+    let mut mac = <Hmac<Sha256> as KeyInit>::new_from_slice(&settings.key).ok()?;
+    let tail = bytes.len() - 32;
+    mac.update(sender.as_bytes());
+    mac.update(&bytes[..tail]);
+    mac.verify_slice(&bytes[tail..]).ok()?;
+    Some((
+        bytes[16],
+        u64::from_be_bytes(bytes.get(17..25)?.try_into().ok()?),
+    ))
+}
+impl PacketLink for HealthLink {
+    fn network(&self) -> NetworkId {
+        self.raw.network()
+    }
+    fn peer(&self) -> EndpointId {
+        self.raw.peer()
+    }
+    fn max_datagram_size(&self) -> usize {
+        self.raw.max_datagram_size()
+    }
+    fn send(&self, payload: Bytes) -> Result<(), TransportError> {
+        if self.is_closed() {
+            Err(TransportError::Closed)
+        } else {
+            self.raw.send(payload)
+        }
+    }
+    fn send_flow(
+        &self,
+        payload: Bytes,
+        flow: super::super::routing::FlowId,
+    ) -> Result<(), TransportError> {
+        if self.is_closed() {
+            Err(TransportError::Closed)
+        } else {
+            self.raw.send_flow(payload, flow)
+        }
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut stop = self.stop.subscribe();
+            if *stop.borrow() {
+                return None;
+            }
+            let mut rx = self.rx.lock().await;
+            tokio::select! {biased;_=stop.changed()=>None,packet=rx.recv()=>packet}
+        })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(async move {
+            let mut stop = self.stop.subscribe();
+            if !*stop.borrow() {
+                let _ = stop.changed().await;
+            }
+        })
+    }
+    fn is_closed(&self) -> bool {
+        *self.stop.borrow() || self.raw.is_closed()
+    }
+    fn close(&self) {
+        self.stop.send_replace(true);
+        self.raw.close();
+    }
+    fn failure_reason(&self) -> Option<String> {
+        lock(&self.reason)
+            .clone()
+            .or_else(|| self.raw.failure_reason())
+    }
+    fn path_description(&self) -> String {
+        self.raw.path_description()
+    }
+}
+impl Drop for HealthLink {
+    fn drop(&mut self) {
+        self.stop.send_replace(true);
+        self.raw.close();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[test]
+    fn no_false_failure_before_window_but_persistent_loss_is_detected() {
+        let mut samples = Samples::new();
+        let now = Instant::now();
+        for seq in 0..20 {
+            samples.sent(seq, now);
+            if seq % 5 == 0 {
+                samples.acknowledge(seq);
+            }
+        }
+        assert!(!samples.degraded());
+        samples.expire(now + Duration::from_secs(4));
+        assert!(samples.degraded());
+        for seq in 20..40 {
+            samples.sent(seq, now);
+            samples.acknowledge(seq);
+        }
+        samples.expire(now + Duration::from_secs(4));
+        assert!(!samples.degraded());
+    }
+    #[test]
+    fn a_forged_or_reflected_probe_cannot_claim_delivery() {
+        let a = crate::identity::DeviceIdentity::generate().endpoint_id();
+        let b = crate::identity::DeviceIdentity::generate().endpoint_id();
+        let settings = Settings::fresh();
+        let packet = probe(&settings, a, 1, 42, 1100);
+        assert_eq!(verify(&settings, a, &packet), Some((1, 42)));
+        assert!(verify(&settings, b, &packet).is_none());
+        let mut forged = packet.to_vec();
+        forged[16] = 2;
+        assert!(verify(&settings, a, &forged).is_none());
+    }
+}

crates/tsunagi-core/src/dataplane/transport/iroh_link.rs

@@ -20,13 +20,14 @@
 
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
 
 use bytes::Bytes;
 use iroh::EndpointId;
 use iroh::endpoint::{Connection, RecvStream, SendStream};
 
 use crate::BoxFuture;
-use crate::config::Limits;
+use crate::config::{DATA_REASSEMBLY_TTL, Limits, MAX_DATA_DATAGRAM, MAX_DATA_FRAGMENTS};
 use crate::error::ProtocolError;
 use crate::identity::{NetworkId, NetworkKeys};
 use crate::net::EndpointAdapter;
@@ -36,6 +37,7 @@ use crate::proto::message::{
 };
 use crate::proto::{read_frame, write_frame};
 
+use super::fragments::{self, Reassembler};
 use super::{InboundLink, PacketLink, PacketTransport, SharedLink, TransportError};
 
 /// What the iroh transport needs from the agent.
@@ -61,6 +63,8 @@ pub struct IrohLink {
     peer: EndpointId,
     conn: Connection,
     max_datagram: usize,
+    next_packet: AtomicU64,
+    reassembly: tokio::sync::Mutex<Reassembler>,
     // Kept alive so the peer sees the channel as open; the connection closes
     // when the link is dropped.
     _send: tokio::sync::Mutex<SendStream>,
@@ -76,14 +80,16 @@ impl IrohLink {
         send: SendStream,
         recv: RecvStream,
     ) -> Self {
-        let local_limit = conn.max_datagram_size().unwrap_or(0);
-        // Both ends must agree, so the smaller limit wins.
-        let max_datagram = local_limit.min(peer_limit);
+        // The negotiated limit is for reassembled payloads, never a snapshot
+        // of the initial path (which may still be a relay or another VPN).
+        let max_datagram = MAX_DATA_DATAGRAM.min(peer_limit);
         Self {
             network,
             peer,
             conn,
             max_datagram,
+            next_packet: AtomicU64::new(0),
+            reassembly: tokio::sync::Mutex::new(Reassembler::default()),
             _send: tokio::sync::Mutex::new(send),
             _recv: tokio::sync::Mutex::new(recv),
         }
@@ -91,6 +97,9 @@ impl IrohLink {
 }
 
 impl PacketLink for IrohLink {
+    fn close(&self) {
+        self.conn.close(0u32.into(), b"data link retired");
+    }
     fn network(&self) -> NetworkId {
         self.network
     }
@@ -110,17 +119,65 @@ impl PacketLink for IrohLink {
                 limit: self.max_datagram,
             });
         }
-        self.conn.send_datagram(payload).map_err(|err| {
-            use iroh::endpoint::SendDatagramError;
-            match err {
-                SendDatagramError::ConnectionLost(_) => TransportError::Closed,
-                other => TransportError::Other(other.to_string()),
+        let id = self.next_packet.fetch_add(1, Ordering::Relaxed);
+        let mut offset: usize = 0;
+        let mut attempts = 0;
+        // Re-read the current path's capacity for every fragment. A migration
+        // can shrink it even between two send_datagram calls.
+        for _ in 0..MAX_DATA_FRAGMENTS {
+            loop {
+                let capacity = self
+                    .conn
+                    .max_datagram_size()
+                    .unwrap_or(0)
+                    .checked_sub(fragments::HEADER)
+                    .filter(|n| *n > 0)
+                    .ok_or_else(|| {
+                        TransportError::Other("QUIC path cannot carry data fragments".into())
+                    })?;
+                let end = offset.saturating_add(capacity).min(payload.len());
+                let frame = fragments::encode(id, payload.len(), offset, &payload[offset..end]);
+                match self.conn.send_datagram(frame) {
+                    Ok(()) => {
+                        offset = end;
+                        break;
+                    }
+                    Err(iroh::endpoint::SendDatagramError::TooLarge) if attempts < 3 => {
+                        attempts += 1;
+                    }
+                    Err(iroh::endpoint::SendDatagramError::ConnectionLost(_)) => {
+                        return Err(TransportError::Closed);
+                    }
+                    Err(err) => return Err(TransportError::Other(err.to_string())),
+                }
             }
-        })
+            if offset == payload.len() {
+                return Ok(());
+            }
+        }
+        Err(TransportError::Other(
+            "QUIC path needs too many fragments".into(),
+        ))
     }
 
     fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
-        Box::pin(async move { self.conn.read_datagram().await.ok() })
+        Box::pin(async move {
+            let mut reassembly = self.reassembly.lock().await;
+            loop {
+                match tokio::time::timeout(DATA_REASSEMBLY_TTL, self.conn.read_datagram()).await {
+                    Ok(Ok(frame)) => {
+                        if let Some(packet) = reassembly.push(frame, tokio::time::Instant::now()) {
+                            return Some(packet);
+                        }
+                    }
+                    Ok(Err(_)) => {
+                        *reassembly = Reassembler::default();
+                        return None;
+                    }
+                    Err(_) => reassembly.expire(tokio::time::Instant::now()),
+                }
+            }
+        })
     }
 
     fn closed(&self) -> BoxFuture<'_, ()> {
@@ -137,8 +194,8 @@ impl PacketLink for IrohLink {
         // Report what iroh actually knows, never a guess.
         let snapshot = crate::net::snapshot_connection(&self.conn);
         match snapshot.paths.iter().find(|path| path.is_selected) {
-            Some(path) => format!("{:?} via {:?}", snapshot.transport, path.remote),
-            None => format!("{:?}, no selected path yet", snapshot.transport),
+            Some(path) => format!("{} via {}", snapshot.transport, path.remote),
+            None => format!("{}, no selected path yet", snapshot.transport),
         }
     }
 }
@@ -208,10 +265,14 @@ impl IrohTransport {
         }
 
         let serves = self.lookup.serves(outcome.network_id, &open.protocol).await;
-        let max_datagram = conn.max_datagram_size().unwrap_or(0);
+        if open.max_datagram == 0 || conn.max_datagram_size().is_none() {
+            return Err(TransportError::Other(
+                "data channel has no datagram support".into(),
+            ));
+        }
         let ack = DataOpenAck {
             accepted: serves,
-            max_datagram: max_datagram as u32,
+            max_datagram: MAX_DATA_DATAGRAM as u32,
         };
         write_frame(
             &mut send,
@@ -226,7 +287,14 @@ impl IrohTransport {
             return Err(TransportError::Declined(open.protocol));
         }
 
-        let link = IrohLink::new(outcome.network_id, peer, conn, usize::MAX, send, recv);
+        let link = IrohLink::new(
+            outcome.network_id,
+            peer,
+            conn,
+            open.max_datagram as usize,
+            send,
+            recv,
+        );
         Ok(InboundLink {
             network: outcome.network_id,
             peer,
@@ -289,6 +357,7 @@ impl PacketTransport for IrohTransport {
 
             let open = DataOpen {
                 protocol: protocol.to_string(),
+                max_datagram: MAX_DATA_DATAGRAM as u32,
             };
             write_frame(
                 &mut send,
@@ -310,6 +379,12 @@ impl PacketTransport for IrohTransport {
                 return Err(TransportError::Declined(protocol.to_string()));
             }
 
+            if ack.max_datagram == 0 || conn.max_datagram_size().is_none() {
+                return Err(TransportError::Other(
+                    "data channel has no datagram support".into(),
+                ));
+            }
+
             let link = IrohLink::new(network, peer, conn, ack.max_datagram as usize, send, recv);
             Ok(Arc::new(link) as SharedLink)
         })

crates/tsunagi-core/src/dataplane/transport/mod.rs

@@ -0,0 +1,317 @@
+//! The data plane transport boundary.
+//!
+//! This is the seam that keeps the control protocol and the data plane
+//! independent. An [`IpPlugin`] never learns how its packets are moved: it is
+//! handed a [`PacketLink`] to a peer and writes datagrams into it. Whether
+//! that link runs over iroh today, a raw UDP socket, or something else
+//! entirely tomorrow is the transport's business alone.
+//!
+//! [`IpPlugin`]: crate::dataplane::IpPlugin
+//!
+//! # Why the transport may use iroh
+//!
+//! The separation between control and data is **logical**, not a ban on
+//! sharing technology. Refusing to use iroh for data would throw away exactly
+//! what iroh is good at — hole punching a direct path between two peers behind
+//! NAT, with a relay as fallback — and force the data plane to reimplement it.
+//! So the default transport is [`iroh_link::IrohTransport`], which gives every
+//! plugin that connectivity for free.
+//!
+//! What the separation does buy is that the control protocol in
+//! [`crate::proto`] knows nothing about packets, and this module knows nothing
+//! about WireGuard. Either side can be replaced on its own.
+//!
+//! # Semantics
+//!
+//! A link is an **unreliable, unordered datagram** channel, because that is
+//! what a tunnelled UDP protocol needs: no retransmission, no head-of-line
+//! blocking, loss is normal rather than an error. It is authenticated and
+//! encrypted by the transport, and scoped to exactly one network, one peer and
+//! one plugin protocol.
+
+mod fragments;
+pub(crate) mod health;
+pub mod iroh_link;
+
+use std::sync::Arc;
+
+use bytes::Bytes;
+use iroh::EndpointId;
+use serde::{Deserialize, Serialize};
+
+use crate::BoxFuture;
+use crate::identity::NetworkId;
+
+/// Why a data plane link failed.
+///
+/// None of these ever stop the control plane.
+#[derive(Debug, thiserror::Error)]
+#[non_exhaustive]
+pub enum TransportError {
+    /// The peer is not reachable for the data plane right now.
+    #[error("peer is unreachable: {0}")]
+    Unreachable(String),
+    /// The peer declined to open a channel for this protocol.
+    #[error("peer declined a data channel for protocol `{0}`")]
+    Declined(String),
+    /// The link is closed.
+    #[error("data link is closed")]
+    Closed,
+    /// A datagram was larger than the link can carry.
+    #[error("datagram of {size} bytes exceeds the {limit} byte link limit")]
+    TooLarge {
+        /// Size that was attempted.
+        size: usize,
+        /// Largest datagram this link accepts.
+        limit: usize,
+    },
+    /// Anything else.
+    #[error("data transport error: {0}")]
+    Other(String),
+}
+
+/// An authenticated datagram channel to one peer, for one plugin protocol.
+///
+/// Dropping the link closes it.
+pub trait PacketLink: Send + Sync + std::fmt::Debug + 'static {
+    /// The network this link belongs to.
+    fn network(&self) -> NetworkId;
+
+    /// The authenticated peer on the other end.
+    fn peer(&self) -> EndpointId;
+
+    /// The largest datagram this link can carry, in bytes.
+    ///
+    /// This is the logical payload limit. A transport can fragment underneath
+    /// it so path MTU changes do not force a plugin to resize the host interface.
+    fn max_datagram_size(&self) -> usize;
+
+    /// Sends one datagram.
+    ///
+    /// Delivery is not guaranteed. Returning `Ok` means the datagram was
+    /// handed to the transport, nothing more.
+    fn send(&self, payload: Bytes) -> Result<(), TransportError>;
+
+    /// Sends a datagram with an opaque stable flow identifier. Plugins derive
+    /// this before encryption; transports that do not route can ignore it.
+    fn send_flow(
+        &self,
+        payload: Bytes,
+        _flow: super::routing::FlowId,
+    ) -> Result<(), TransportError> {
+        self.send(payload)
+    }
+
+    /// Receives the next datagram, or `None` once the link is finished.
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>>;
+
+    /// Resolves once the link is closed, for whatever reason.
+    fn closed(&self) -> BoxFuture<'_, ()>;
+
+    /// Whether the link is already closed.
+    ///
+    /// Lets the owner notice a dead link and ask for a new one without
+    /// keeping a task parked on [`PacketLink::closed`].
+    fn is_closed(&self) -> bool;
+
+    /// Explicitly retires a physical link. Implementations should wake readers.
+    fn close(&self) {}
+
+    /// Observed reason for retiring this link, if available.
+    fn failure_reason(&self) -> Option<String> {
+        None
+    }
+
+    /// A short description of the path in use, for diagnostics.
+    ///
+    /// Reports what the transport actually knows. It must not invent a value.
+    fn path_description(&self) -> String;
+}
+
+/// A shared handle to a link.
+pub type SharedLink = std::sync::Arc<dyn PacketLink>;
+
+/// Maximum opaque pair-negotiation payload.
+pub const MAX_TRANSPORT_PAYLOAD: usize = 4096;
+
+/// Private transport bytes, never included in Debug output.
+#[derive(Clone, PartialEq, Eq, Serialize, zeroize::ZeroizeOnDrop)]
+pub struct TransportPayload(Vec<u8>);
+
+impl<'de> Deserialize<'de> for TransportPayload {
+    fn deserialize<D: serde::Deserializer<'de>>(
+        deserializer: D,
+    ) -> std::result::Result<Self, D::Error> {
+        struct Visitor;
+        impl<'de> serde::de::Visitor<'de> for Visitor {
+            type Value = TransportPayload;
+            fn expecting(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+                write!(f, "at most {MAX_TRANSPORT_PAYLOAD} private transport bytes")
+            }
+            fn visit_seq<A: serde::de::SeqAccess<'de>>(
+                self,
+                mut sequence: A,
+            ) -> std::result::Result<Self::Value, A::Error> {
+                if sequence
+                    .size_hint()
+                    .is_some_and(|size| size > MAX_TRANSPORT_PAYLOAD)
+                {
+                    return Err(serde::de::Error::custom("transport payload too large"));
+                }
+                let mut payload = TransportPayload(Vec::with_capacity(
+                    sequence.size_hint().unwrap_or(0).min(MAX_TRANSPORT_PAYLOAD),
+                ));
+                while let Some(byte) = sequence.next_element::<u8>()? {
+                    if payload.0.len() == MAX_TRANSPORT_PAYLOAD {
+                        return Err(serde::de::Error::custom("transport payload too large"));
+                    }
+                    payload.0.push(byte);
+                }
+                Ok(payload)
+            }
+        }
+        deserializer.deserialize_seq(Visitor)
+    }
+}
+impl TransportPayload {
+    /// Wraps bytes for one authenticated control session.
+    pub fn new(bytes: Vec<u8>) -> Self {
+        Self(bytes)
+    }
+    /// Borrows private bytes; callers must not log them.
+    pub fn as_bytes(&self) -> &[u8] {
+        &self.0
+    }
+}
+impl std::fmt::Debug for TransportPayload {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("TransportPayload([REDACTED])")
+    }
+}
+
+/// Private control-plane sender for transports. Contains no network secret.
+pub trait ControlSender: Send + Sync + std::fmt::Debug + 'static {
+    /// Queues a bounded message to one authenticated member of one network.
+    fn send<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &'a str,
+        payload: TransportPayload,
+    ) -> BoxFuture<'a, Result<(), TransportError>>;
+}
+
+/// An inbound link a peer opened towards us.
+#[derive(Debug)]
+pub struct InboundLink {
+    /// The network it belongs to.
+    pub network: NetworkId,
+    /// The peer that opened it.
+    pub peer: EndpointId,
+    /// The plugin protocol it carries.
+    pub protocol: String,
+    /// The link itself.
+    pub link: SharedLink,
+}
+
+/// Opens and accepts data plane links.
+///
+/// The agent owns one of these and hands links to plugins; plugins never call
+/// it directly.
+pub trait PacketTransport: Send + Sync + std::fmt::Debug + 'static {
+    /// Supplies a private sender before network activation.
+    fn attach_control(&self, _control: Arc<dyn ControlSender>) {}
+    /// Validates and queues private negotiation received from an authenticated member.
+    fn on_control(
+        &self,
+        _network: NetworkId,
+        _peer: EndpointId,
+        _protocol: &str,
+        _payload: &[u8],
+    ) -> Result<(), TransportError> {
+        Err(TransportError::Declined(self.name().into()))
+    }
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Lets the agent recover the concrete transport to drive its accept side.
+    ///
+    /// Accepting is inherently transport-specific — it starts from whatever
+    /// the transport's own listener produced — so it is not part of this
+    /// trait's uniform interface.
+    fn as_any(&self) -> &dyn std::any::Any;
+
+    /// Opens a link to `peer` in `network` for `protocol`.
+    fn open<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &'a str,
+    ) -> BoxFuture<'a, Result<SharedLink, TransportError>>;
+}
+
+/// A composite packet transport that dispatches requests by protocol id.
+#[derive(Debug)]
+pub struct MultiPacketTransport {
+    default: Arc<dyn PacketTransport>,
+    by_protocol: std::collections::HashMap<String, Arc<dyn PacketTransport>>,
+}
+
+impl MultiPacketTransport {
+    /// Creates a new composite transport with a default fallback transport.
+    pub fn new(
+        default: Arc<dyn PacketTransport>,
+        by_protocol: std::collections::HashMap<String, Arc<dyn PacketTransport>>,
+    ) -> Self {
+        Self {
+            default,
+            by_protocol,
+        }
+    }
+
+    /// Access the default underlying transport.
+    pub fn default_transport(&self) -> &Arc<dyn PacketTransport> {
+        &self.default
+    }
+}
+
+impl PacketTransport for MultiPacketTransport {
+    fn attach_control(&self, control: Arc<dyn ControlSender>) {
+        self.default.attach_control(control.clone());
+        for transport in self.by_protocol.values() {
+            transport.attach_control(control.clone());
+        }
+    }
+    fn on_control(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &str,
+        payload: &[u8],
+    ) -> Result<(), TransportError> {
+        match self.by_protocol.get(protocol) {
+            Some(transport) => transport.on_control(network, peer, protocol, payload),
+            None => Err(TransportError::Declined(protocol.into())),
+        }
+    }
+    fn name(&self) -> &str {
+        "multi"
+    }
+
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+
+    fn open<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &'a str,
+    ) -> BoxFuture<'a, Result<SharedLink, TransportError>> {
+        if let Some(transport) = self.by_protocol.get(protocol) {
+            transport.open(network, peer, protocol)
+        } else {
+            self.default.open(network, peer, protocol)
+        }
+    }
+}

crates/tsunagi-core/src/dataplane/tunnel.rs

@@ -0,0 +1,380 @@
+//! Generic harness for IP tunnel plugins.
+//!
+//! Implementing [`super::IpPlugin`] directly requires managing per-peer background
+//! tasks, synchronization, packet queues, and graceful shutdown.
+//!
+//! [`TunnelCodec`] provides a simplified interface where a plugin author only
+//! defines cryptographic transformation and handshake payloads.
+//! [`GenericTunnelPlugin`] adapts any [`TunnelCodec`] into a full [`super::IpPlugin`].
+
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex, RwLock};
+
+use bytes::Bytes;
+use iroh::EndpointId;
+use tokio::task::JoinHandle;
+
+use super::{IpPlugin, PluginCapability, PluginContext, PluginError, ProtocolOption, SharedLink};
+use crate::BoxFuture;
+use crate::identity::NetworkId;
+
+/// The simplified contract for an IP tunnel protocol implementation.
+///
+/// Implementors only handle key announcements, link lifecycle events,
+/// and packet encrypt/decrypt. Concurrency and task management are handled
+/// by [`GenericTunnelPlugin`].
+pub trait TunnelCodec: Send + Sync + 'static {
+    /// Protocol identifier, e.g. `tcp-tls`.
+    fn protocol_id(&self) -> &str;
+
+    /// Protocol wire version.
+    fn protocol_version(&self) -> u16;
+
+    /// Options accepted by the protocol.
+    fn options(&self) -> &'static [ProtocolOption] {
+        &[]
+    }
+
+    /// Generates local capability announcement data for a network.
+    fn local_capability_data(&self, network: NetworkId) -> Result<Vec<u8>, PluginError>;
+
+    /// Handles a peer's announced capability payload.
+    fn on_peer_capability_data(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        data: &[u8],
+    ) -> Result<(), PluginError>;
+
+    /// Called when an authenticated link to a peer is established.
+    fn on_peer_link_up(&self, network: NetworkId, peer: EndpointId, link: &SharedLink) {
+        let _ = (network, peer, link);
+    }
+
+    /// Called when a peer goes away.
+    fn on_peer_down(&self, network: NetworkId, peer: EndpointId) {
+        let _ = (network, peer);
+    }
+
+    /// Encrypts an outbound IP packet into a payload to send over the link.
+    fn encrypt(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        packet: &[u8],
+    ) -> Result<Bytes, PluginError>;
+
+    /// Decrypts an inbound payload from the link into an IP packet.
+    fn decrypt(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        payload: &[u8],
+    ) -> Result<Bytes, PluginError>;
+
+    /// Called during shutdown.
+    fn shutdown<'a>(&'a self) -> BoxFuture<'a, ()> {
+        Box::pin(async {})
+    }
+}
+
+struct TunnelPeerState {
+    link: SharedLink,
+    tx_packets: AtomicU64,
+    rx_packets: AtomicU64,
+    tx_bytes: AtomicU64,
+    rx_bytes: AtomicU64,
+    dropped: AtomicU64,
+    established_at: std::time::Instant,
+}
+
+/// Statistics and status for one active peer in [`GenericTunnelPlugin`].
+#[derive(Debug, Clone)]
+pub struct TunnelPeerReport {
+    /// Peer endpoint identity.
+    pub peer: EndpointId,
+    /// Protocol identifier.
+    pub protocol: String,
+    /// Human-readable path description.
+    pub path: String,
+    /// Packets successfully encrypted and sent.
+    pub tx_packets: u64,
+    /// Packets successfully received and decrypted.
+    pub rx_packets: u64,
+    /// Plaintext bytes successfully encrypted and sent.
+    pub tx_bytes: u64,
+    /// Plaintext bytes successfully received and decrypted.
+    pub rx_bytes: u64,
+    /// Packets dropped due to encryption/decryption/send failures.
+    pub dropped: u64,
+    /// Uptime in seconds since the data link was established.
+    pub uptime_secs: u64,
+}
+
+/// Generic wrapper that turns a [`TunnelCodec`] into a full [`IpPlugin`].
+pub struct GenericTunnelPlugin<C: TunnelCodec> {
+    codec: Arc<C>,
+    context: Mutex<Option<PluginContext>>,
+    links: RwLock<HashMap<(NetworkId, EndpointId), Arc<TunnelPeerState>>>,
+    tasks: Mutex<HashMap<(NetworkId, EndpointId), JoinHandle<()>>>,
+}
+
+impl<C: TunnelCodec> GenericTunnelPlugin<C> {
+    /// Wraps `codec` into an [`IpPlugin`].
+    pub fn new(codec: C) -> Self {
+        Self {
+            codec: Arc::new(codec),
+            context: Mutex::new(None),
+            links: RwLock::new(HashMap::new()),
+            tasks: Mutex::new(HashMap::new()),
+        }
+    }
+
+    /// Returns a reference to the underlying codec.
+    pub fn codec(&self) -> &C {
+        &self.codec
+    }
+
+    /// Returns the active peers with established data links for a network.
+    pub fn active_peers(&self, network: NetworkId) -> Vec<TunnelPeerReport> {
+        let guard = match self.links.read() {
+            Ok(g) => g,
+            Err(p) => p.into_inner(),
+        };
+        guard
+            .iter()
+            .filter(|((net, _), _)| *net == network)
+            .map(|((_, peer), state)| TunnelPeerReport {
+                peer: *peer,
+                protocol: self.codec.protocol_id().to_string(),
+                path: state.link.path_description(),
+                tx_packets: state.tx_packets.load(Ordering::Relaxed),
+                rx_packets: state.rx_packets.load(Ordering::Relaxed),
+                tx_bytes: state.tx_bytes.load(Ordering::Relaxed),
+                rx_bytes: state.rx_bytes.load(Ordering::Relaxed),
+                dropped: state.dropped.load(Ordering::Relaxed),
+                uptime_secs: state.established_at.elapsed().as_secs(),
+            })
+            .collect()
+    }
+}
+
+impl<C: TunnelCodec> std::fmt::Debug for GenericTunnelPlugin<C> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("GenericTunnelPlugin")
+            .field("protocol_id", &self.codec.protocol_id())
+            .field("protocol_version", &self.codec.protocol_version())
+            .finish()
+    }
+}
+
+impl<C: TunnelCodec> IpPlugin for GenericTunnelPlugin<C> {
+    fn protocol_id(&self) -> &str {
+        self.codec.protocol_id()
+    }
+
+    fn protocol_version(&self) -> u16 {
+        self.codec.protocol_version()
+    }
+
+    fn options(&self) -> &'static [ProtocolOption] {
+        self.codec.options()
+    }
+
+    fn attach(&self, context: PluginContext) {
+        if let Ok(mut guard) = self.context.lock() {
+            *guard = Some(context);
+        }
+    }
+
+    fn local_capability(
+        &self,
+        network: NetworkId,
+    ) -> Result<Option<PluginCapability>, PluginError> {
+        let data = self.codec.local_capability_data(network)?;
+        Ok(Some(PluginCapability {
+            protocol: self.codec.protocol_id().to_string(),
+            version: self.codec.protocol_version(),
+            enabled: true,
+            data,
+        }))
+    }
+
+    fn on_peer_capability(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        capability: &PluginCapability,
+    ) -> Result<(), PluginError> {
+        if capability.protocol != self.codec.protocol_id() {
+            return Ok(());
+        }
+        self.codec
+            .on_peer_capability_data(network, peer, &capability.data)
+    }
+
+    fn carry(&self, network: NetworkId, peer: EndpointId, packet: bytes::Bytes) -> bool {
+        let state = {
+            let guard = match self.links.read() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.get(&(network, peer)).cloned()
+        };
+
+        let Some(state) = state else {
+            return false;
+        };
+
+        match self.codec.encrypt(network, peer, &packet) {
+            Ok(ciphertext) => {
+                if state.link.send(ciphertext).is_ok() {
+                    state.tx_packets.fetch_add(1, Ordering::Relaxed);
+                    state
+                        .tx_bytes
+                        .fetch_add(packet.len() as u64, Ordering::Relaxed);
+                    true
+                } else {
+                    state.dropped.fetch_add(1, Ordering::Relaxed);
+                    false
+                }
+            }
+            Err(err) => {
+                state.dropped.fetch_add(1, Ordering::Relaxed);
+                tracing::debug!(%peer, %err, "tunnel encrypt failed");
+                false
+            }
+        }
+    }
+
+    fn on_peer_link(&self, network: NetworkId, peer: EndpointId, link: SharedLink) {
+        let state = Arc::new(TunnelPeerState {
+            link: Arc::clone(&link),
+            tx_packets: AtomicU64::new(0),
+            rx_packets: AtomicU64::new(0),
+            tx_bytes: AtomicU64::new(0),
+            rx_bytes: AtomicU64::new(0),
+            dropped: AtomicU64::new(0),
+            established_at: std::time::Instant::now(),
+        });
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.insert((network, peer), Arc::clone(&state));
+        }
+
+        self.codec.on_peer_link_up(network, peer, &link);
+
+        let codec = Arc::clone(&self.codec);
+        let sink = match self.context.lock() {
+            Ok(g) => g.as_ref().map(|ctx| ctx.packet_sink()),
+            Err(p) => p.into_inner().as_ref().map(|ctx| ctx.packet_sink()),
+        };
+
+        let rx_link = Arc::clone(&link);
+        let rx_state = Arc::clone(&state);
+        let handle = tokio::spawn(async move {
+            while let Some(payload) = rx_link.recv().await {
+                match codec.decrypt(network, peer, &payload) {
+                    Ok(packet) => {
+                        rx_state.rx_packets.fetch_add(1, Ordering::Relaxed);
+                        rx_state
+                            .rx_bytes
+                            .fetch_add(packet.len() as u64, Ordering::Relaxed);
+                        if let Some(ref sink) = sink {
+                            sink.deliver(network, peer, packet).await;
+                        }
+                    }
+                    Err(err) => {
+                        rx_state.dropped.fetch_add(1, Ordering::Relaxed);
+                        tracing::debug!(%peer, %err, "tunnel decrypt failed");
+                    }
+                }
+            }
+        });
+
+        let mut guard = match self.tasks.lock() {
+            Ok(g) => g,
+            Err(p) => p.into_inner(),
+        };
+        if let Some(old) = guard.insert((network, peer), handle) {
+            old.abort();
+        }
+    }
+
+    fn on_peer_gone(&self, network: NetworkId, peer: EndpointId) {
+        {
+            let mut guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            if let Some(task) = guard.remove(&(network, peer)) {
+                task.abort();
+            }
+        }
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.remove(&(network, peer));
+        }
+        self.codec.on_peer_down(network, peer);
+    }
+
+    fn on_network_deactivated(&self, network: NetworkId) {
+        let keys_to_remove: Vec<(NetworkId, EndpointId)> = {
+            let guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard
+                .keys()
+                .filter(|(net, _)| *net == network)
+                .copied()
+                .collect()
+        };
+
+        {
+            let mut guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            for key in &keys_to_remove {
+                if let Some(task) = guard.remove(key) {
+                    task.abort();
+                }
+            }
+        }
+
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.retain(|(net, _), _| *net != network);
+        }
+    }
+
+    fn shutdown<'a>(&'a self) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let mut tasks = Vec::new();
+            {
+                let mut guard = match self.tasks.lock() {
+                    Ok(g) => g,
+                    Err(p) => p.into_inner(),
+                };
+                for (_, task) in guard.drain() {
+                    tasks.push(task);
+                }
+            }
+            for task in tasks {
+                task.abort();
+            }
+            self.codec.shutdown().await;
+        })
+    }
+}

crates/tsunagi-core/src/discovery.rs

@@ -12,8 +12,7 @@
 //!
 //! * *Finding members of a network* — [`NetworkDiscovery::resolve`], keyed by
 //!   the secret-derived [`DiscoveryKey`]. That is what lives here, and today
-//!   it is [`StaticBootstrap`] plus a test backend; a DHT backend is future
-//!   work.
+//!   it includes [`StaticBootstrap`], [`MainlineDiscovery`] and a test backend.
 //! * *Resolving the address of one iroh endpoint* — **iroh's job, not ours**.
 //!   With [`crate::config::TransportPolicy::N0Defaults`] or `DirectOnly`, iroh
 //!   publishes and resolves endpoint addresses through Number 0's public
@@ -23,14 +22,18 @@
 //! No empty result ever proves a network is empty. It only means "nobody found
 //! yet".
 //!
-//! Mainline DHT discovery is future work and is not implemented here.
+//! Mainline rendezvous is opt-in for library callers; the command line enables it.
+
+mod mainline;
+pub(crate) mod worker;
+pub use mainline::MainlineDiscovery;
 
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex};
 
 use iroh::{EndpointAddr, EndpointId};
 
-use crate::error::Result;
+use crate::error::{Error, Result};
 use crate::identity::DiscoveryKey;
 
 pub use crate::BoxFuture;
@@ -44,6 +47,18 @@ pub enum CandidateSource {
     Discovery,
     /// An address hint restored from the disposable cache.
     Cache,
+    /// An author of a signed record: somebody who belongs to this network,
+    /// learned from state that reached us through anybody.
+    ///
+    /// It carries no address of its own — the endpoint's own discovery has
+    /// to resolve it — but knowing that a member exists is what turns a
+    /// star around whoever was named on the command line into a mesh.
+    Member,
+    /// Passed on by a member we are talking to.
+    ///
+    /// A candidate like any other: an introduction is not a vouching, and
+    /// membership is still decided by the handshake.
+    Introduced,
 }
 
 /// An unverified candidate peer.
@@ -91,6 +106,23 @@ pub trait NetworkDiscovery: Send + Sync + std::fmt::Debug + 'static {
 
     /// Returns the candidates currently known for `key`.
     fn resolve<'a>(&'a self, key: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>>;
+
+    /// Delivers candidates as they arrive. The default adapts a batch backend;
+    /// remote backends can override this so the first dial need not wait for all lookups.
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: tokio::sync::mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            for candidate in self.resolve(key).await? {
+                if candidates.send(candidate).await.is_err() {
+                    break;
+                }
+            }
+            Ok(())
+        })
+    }
 }
 
 /// A statically configured list of bootstrap candidates.
@@ -261,13 +293,25 @@ impl NetworkDiscovery for CompositeDiscovery {
 
     fn publish<'a>(&'a self, key: DiscoveryKey, addr: EndpointAddr) -> BoxFuture<'a, Result<()>> {
         Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
             for backend in &self.backends {
-                // One failing backend must not stop the others.
-                if let Err(err) = backend.publish(key, addr.clone()).await {
-                    tracing::debug!(backend = backend.name(), %err, "publish failed");
+                let backend = backend.clone();
+                let addr = addr.clone();
+                tasks.spawn(async move { backend.publish(key, addr).await });
+            }
+            let mut failed = false;
+            while let Some(result) = tasks.join_next().await {
+                if !matches!(result, Ok(Ok(()))) {
+                    failed = true;
                 }
             }
-            Ok(())
+            if failed {
+                Err(Error::Discovery(
+                    "a discovery publication failed; will retry".into(),
+                ))
+            } else {
+                Ok(())
+            }
         })
     }
 
@@ -277,11 +321,12 @@ impl NetworkDiscovery for CompositeDiscovery {
         endpoint: EndpointId,
     ) -> BoxFuture<'a, Result<()>> {
         Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
             for backend in &self.backends {
-                if let Err(err) = backend.unpublish(key, endpoint).await {
-                    tracing::debug!(backend = backend.name(), %err, "unpublish failed");
-                }
+                let backend = backend.clone();
+                tasks.spawn(async move { backend.unpublish(key, endpoint).await });
             }
+            while tasks.join_next().await.is_some() {}
             Ok(())
         })
     }
@@ -300,4 +345,25 @@ impl NetworkDiscovery for CompositeDiscovery {
             Ok(out)
         })
     }
+
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: tokio::sync::mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
+            for backend in &self.backends {
+                let backend = backend.clone();
+                let candidates = candidates.clone();
+                tasks.spawn(async move {
+                    if let Err(err) = backend.resolve_into(key, candidates).await {
+                        tracing::debug!(backend = backend.name(), %err, "resolve failed");
+                    }
+                });
+            }
+            while tasks.join_next().await.is_some() {}
+            Ok(())
+        })
+    }
 }

crates/tsunagi-core/src/discovery/mainline.rs

@@ -0,0 +1,516 @@
+//! Mainline BEP44 rendezvous. Slots are a changing sample, not a membership list.
+
+use ::mainline::{Dht, MutableItem, SigningKey, async_dht::AsyncDht};
+use futures_lite::StreamExt;
+use iroh::{EndpointAddr, EndpointId};
+use sha2::{Digest, Sha256};
+use std::collections::{HashSet, VecDeque};
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+use tokio::sync::{Mutex as AsyncMutex, mpsc};
+use tokio::task::JoinSet;
+use tokio::time::timeout;
+use zeroize::Zeroizing;
+
+use super::{BoxFuture, Candidate, CandidateSource, NetworkDiscovery};
+use crate::config::{
+    DHT_CLOCK_SKEW, DHT_MAX_ADDRS, DHT_MAX_RELAY_LEN, DHT_MAX_VALUE, DHT_RECORD_TTL, DHT_SLOTS,
+};
+use crate::error::{Error, Result};
+use crate::identity::{DiscoveryKey, NetworkKeys};
+
+const QUERY_TIMEOUT: Duration = Duration::from_secs(8);
+const MAGIC: &[u8; 5] = b"TSND\x01";
+
+/// One lazily started Mainline client, shared by all networks on one agent.
+/// No socket is opened until an active network publishes or looks up candidates.
+/// Agent shutdown closes this client and all its clones; a new agent needs a new client.
+#[derive(Debug, Clone, Default)]
+pub struct MainlineDiscovery {
+    client: Arc<AsyncMutex<ClientState>>,
+    allow_loopback: bool,
+}
+
+#[derive(Debug, Default)]
+enum ClientState {
+    #[default]
+    Pending,
+    Ready(AsyncDht),
+    Stopped,
+}
+
+impl MainlineDiscovery {
+    /// Uses a caller-supplied Mainline node instead of public bootstrap defaults.
+    pub fn from_dht(dht: Dht) -> Self {
+        Self {
+            client: Arc::new(AsyncMutex::new(ClientState::Ready(dht.as_async()))),
+            allow_loopback: false,
+        }
+    }
+
+    /// Creates a loopback-only client for a local Mainline Testnet.
+    #[cfg(feature = "testing")]
+    pub fn local_testnet(bootstrap: &[String]) -> Result<Self> {
+        // Reject anything that could resolve or send outside loopback.
+        if bootstrap.is_empty()
+            || bootstrap.iter().any(|s| {
+                s.parse::<SocketAddr>()
+                    .map_or(true, |a| !a.ip().is_loopback())
+            })
+        {
+            return Err(Error::Discovery(
+                "test bootstrap must contain loopback sockets".into(),
+            ));
+        }
+        let dht = Dht::builder()
+            .bootstrap(bootstrap)
+            .bind_address(Ipv4Addr::LOCALHOST)
+            .port(0)
+            .request_timeout(Duration::from_millis(200))
+            .build()
+            .map_err(|_| Error::Discovery("cannot bind local DHT".into()))?;
+        Ok(Self {
+            allow_loopback: true,
+            ..Self::from_dht(dht)
+        })
+    }
+
+    /// Binds this client's rendezvous backend to one network's independent keys.
+    pub fn for_network(&self, keys: &NetworkKeys) -> Arc<dyn NetworkDiscovery> {
+        Arc::new(NetworkDht {
+            client: self.clone(),
+            key: keys.discovery_key(),
+            seed: Zeroizing::new(*keys.dht_write_key()),
+            observed: Mutex::new(VecDeque::new()),
+        })
+    }
+
+    async fn client(&self) -> Result<AsyncDht> {
+        let mut client = self.client.lock().await;
+        match &*client {
+            ClientState::Ready(dht) => return Ok(dht.clone()),
+            ClientState::Stopped => return Err(Error::Discovery("DHT client is stopped".into())),
+            ClientState::Pending => {}
+        }
+        // Construction may resolve bootstrap hostnames. Keep it off Tokio's
+        // executor; a construction failure is retried on the next operation.
+        let dht = tokio::task::spawn_blocking(|| Dht::builder().port(0).build())
+            .await
+            .map_err(|_| Error::Discovery("DHT startup task failed".into()))?
+            .map(Dht::as_async)
+            .map_err(|_| Error::Discovery("cannot start DHT client".into()))?;
+        *client = ClientState::Ready(dht.clone());
+        Ok(dht)
+    }
+
+    pub(crate) async fn shutdown(&self) {
+        *self.client.lock().await = ClientState::Stopped;
+    }
+}
+
+struct NetworkDht {
+    client: MainlineDiscovery,
+    key: DiscoveryKey,
+    seed: Zeroizing<[u8; 32]>,
+    // Records encountered while publishing. Only an explicit lookup consumes
+    // these hints: publication never causes dials on a connected network.
+    observed: Mutex<VecDeque<Vec<u8>>>,
+}
+
+impl std::fmt::Debug for NetworkDht {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("NetworkDht(<redacted>)")
+    }
+}
+
+fn salt(slot: u8) -> Vec<u8> {
+    let mut salt = b"tsunagi-rendezvous-v1".to_vec();
+    salt.push(slot);
+    salt
+}
+
+fn slots_for(id: EndpointId) -> [u8; 2] {
+    let hash = Sha256::digest(id.as_bytes());
+    let first = hash[0] % DHT_SLOTS;
+    [first, (first + 1 + hash[1] % (DHT_SLOTS - 1)) % DHT_SLOTS]
+}
+
+fn now() -> Result<u64> {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .map_err(|_| Error::Discovery("system clock predates Unix epoch".into()))
+}
+
+impl NetworkDht {
+    async fn publish_slot(&self, dht: &AsyncDht, slot: u8, value: &[u8]) -> Result<()> {
+        let salt = salt(slot);
+        let signer = SigningKey::from_bytes(&self.seed);
+        let public = signer.verifying_key().to_bytes();
+        for attempt in 0..3 {
+            let recent = timeout(
+                QUERY_TIMEOUT,
+                dht.get_mutable_most_recent(&public, Some(&salt)),
+            )
+            .await
+            .map_err(|_| Error::Discovery("DHT read timed out".into()))?;
+            if let Some(item) = &recent
+                && item.value().len() <= DHT_MAX_VALUE
+                && let Ok(mut observed) = self.observed.lock()
+            {
+                if observed.len() == DHT_SLOTS as usize {
+                    observed.pop_front();
+                }
+                observed.push_back(item.value().to_vec());
+            }
+            let seq = recent
+                .as_ref()
+                .map_or(0, MutableItem::seq)
+                .checked_add(1)
+                .ok_or_else(|| Error::Discovery("DHT sequence exhausted".into()))?;
+            let item = MutableItem::new(signer.clone(), value, seq, Some(&salt));
+            match timeout(
+                QUERY_TIMEOUT,
+                dht.put_mutable(item, recent.as_ref().map(MutableItem::seq)),
+            )
+            .await
+            {
+                Ok(Ok(outcome)) if outcome.stored_at > 0 => return Ok(()),
+                _ if attempt < 2 => {
+                    // CAS is per storage node, not a global lock. Read again on
+                    // conflicts; another writer's entry is an acceptable sample.
+                    tokio::time::sleep(Duration::from_millis(rand::random_range(30..150))).await;
+                }
+                _ => break,
+            }
+        }
+        Err(Error::Discovery(
+            "DHT publication was not acknowledged".into(),
+        ))
+    }
+}
+
+impl NetworkDiscovery for NetworkDht {
+    fn name(&self) -> &str {
+        "mainline"
+    }
+
+    fn publish<'a>(&'a self, key: DiscoveryKey, addr: EndpointAddr) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            if key != self.key {
+                return Err(Error::Discovery("wrong rendezvous network".into()));
+            }
+            let value = encode_record(&addr, now()?, self.client.allow_loopback)?;
+            let dht = self.client.client().await?;
+            let [first, second] = slots_for(addr.id);
+            let (a, b) = tokio::join!(
+                self.publish_slot(&dht, first, &value),
+                self.publish_slot(&dht, second, &value)
+            );
+            a.and(b)
+        })
+    }
+
+    fn unpublish<'a>(&'a self, _: DiscoveryKey, _: EndpointId) -> BoxFuture<'a, Result<()>> {
+        // Stopping publication lets freshness expire. Deleting a shared slot
+        // could erase a different writer, and BEP44 has no delete operation.
+        Box::pin(async { Ok(()) })
+    }
+
+    fn resolve<'a>(&'a self, key: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+        Box::pin(async move {
+            let (tx, mut rx) = mpsc::channel(DHT_SLOTS as usize);
+            let collect = async move {
+                let mut found = Vec::new();
+                while let Some(candidate) = rx.recv().await {
+                    found.push(candidate);
+                }
+                found
+            };
+            let (result, found) = tokio::join!(self.resolve_into(key, tx), collect);
+            result?;
+            Ok(found)
+        })
+    }
+
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            if key != self.key {
+                return Err(Error::Discovery("wrong rendezvous network".into()));
+            }
+            let dht = self.client.client().await?;
+            let public = SigningKey::from_bytes(&self.seed)
+                .verifying_key()
+                .to_bytes();
+            let observed = self
+                .observed
+                .lock()
+                .map(|mut values| std::mem::take(&mut *values))
+                .unwrap_or_default();
+            let mut seen = HashSet::new();
+            for value in observed {
+                if let Some(addr) = decode_record(&value, now()?, self.client.allow_loopback)
+                    && seen.insert(addr.id)
+                    && candidates
+                        .send(Candidate::new(addr, CandidateSource::Discovery))
+                        .await
+                        .is_err()
+                {
+                    return Ok(());
+                }
+            }
+            let mut slots: Vec<_> = (0..DHT_SLOTS).collect();
+            for i in (1..slots.len()).rev() {
+                slots.swap(i, rand::random_range(0..=i));
+            }
+            let mut queries = JoinSet::new();
+            loop {
+                if seen.len() >= DHT_SLOTS as usize {
+                    break;
+                }
+                while queries.len() < 4 {
+                    let Some(slot) = slots.pop() else {
+                        break;
+                    };
+                    let dht = dht.clone();
+                    let allow_loopback = self.client.allow_loopback;
+                    queries.spawn(async move {
+                        timeout(QUERY_TIMEOUT, async move {
+                            let salt = salt(slot);
+                            let target = MutableItem::target_from_key(&public, Some(&salt));
+                            let mut items = dht.get_mutable(&public, Some(&salt), None);
+                            while let Some(item) = items.next().await {
+                                // Mainline verifies the signature; also bind the
+                                // returned item to the key and slot we requested.
+                                if item.key() != &public || item.target() != &target {
+                                    continue;
+                                }
+                                if let Some(addr) =
+                                    decode_record(item.value(), now().ok()?, allow_loopback)
+                                {
+                                    return Some(addr);
+                                }
+                            }
+                            None
+                        })
+                        .await
+                        .ok()
+                        .flatten()
+                    });
+                }
+                let Some(result) = queries.join_next().await else {
+                    break;
+                };
+                if let Ok(Some(addr)) = result
+                    && seen.insert(addr.id)
+                    && candidates
+                        .send(Candidate::new(addr, CandidateSource::Discovery))
+                        .await
+                        .is_err()
+                {
+                    break;
+                }
+            }
+            Ok(())
+        })
+    }
+}
+
+fn usable(addr: &SocketAddr, allow_loopback: bool) -> bool {
+    addr.port() != 0
+        && !addr.ip().is_unspecified()
+        && !addr.ip().is_multicast()
+        && (allow_loopback || !addr.ip().is_loopback())
+        && match addr.ip() {
+            IpAddr::V4(ip) => !ip.is_broadcast() && !ip.is_link_local(),
+            IpAddr::V6(ip) => !ip.is_unicast_link_local(),
+        }
+}
+
+fn encode_record(addr: &EndpointAddr, timestamp: u64, allow_loopback: bool) -> Result<Vec<u8>> {
+    let ips: Vec<_> = addr
+        .ip_addrs()
+        .filter(|ip| usable(ip, allow_loopback))
+        .take(DHT_MAX_ADDRS)
+        .collect();
+    let relay = addr
+        .relay_urls()
+        .next()
+        .map(ToString::to_string)
+        .unwrap_or_default();
+    if relay.len() > DHT_MAX_RELAY_LEN || (ips.is_empty() && relay.is_empty()) {
+        return Err(Error::Discovery("no encodable endpoint address yet".into()));
+    }
+    let mut bytes = Vec::with_capacity(DHT_MAX_VALUE);
+    bytes.extend_from_slice(MAGIC);
+    bytes.extend_from_slice(&timestamp.to_be_bytes());
+    bytes.extend_from_slice(addr.id.as_bytes());
+    bytes.extend_from_slice(&(relay.len() as u16).to_be_bytes());
+    bytes.extend_from_slice(relay.as_bytes());
+    bytes.push(ips.len() as u8);
+    for ip in ips {
+        match ip.ip() {
+            IpAddr::V4(v) => {
+                bytes.push(4);
+                bytes.extend_from_slice(&v.octets());
+            }
+            IpAddr::V6(v) => {
+                bytes.push(6);
+                bytes.extend_from_slice(&v.octets());
+            }
+        }
+        bytes.extend_from_slice(&ip.port().to_be_bytes());
+    }
+    if bytes.len() > DHT_MAX_VALUE {
+        return Err(Error::Discovery("DHT record too large".into()));
+    }
+    Ok(bytes)
+}
+
+fn take<'a>(input: &mut &'a [u8], n: usize) -> Option<&'a [u8]> {
+    let (head, rest) = input.split_at_checked(n)?;
+    *input = rest;
+    Some(head)
+}
+
+fn decode_record(mut bytes: &[u8], now: u64, allow_loopback: bool) -> Option<EndpointAddr> {
+    if bytes.len() > DHT_MAX_VALUE || take(&mut bytes, MAGIC.len())? != MAGIC {
+        return None;
+    }
+    let timestamp = u64::from_be_bytes(take(&mut bytes, 8)?.try_into().ok()?);
+    if timestamp > now.saturating_add(DHT_CLOCK_SKEW.as_secs())
+        || now.saturating_sub(timestamp) >= DHT_RECORD_TTL.as_secs()
+    {
+        return None;
+    }
+    let id = EndpointId::from_bytes(take(&mut bytes, 32)?.try_into().ok()?).ok()?;
+    let mut addr = EndpointAddr::new(id);
+    let len = u16::from_be_bytes(take(&mut bytes, 2)?.try_into().ok()?) as usize;
+    if len > DHT_MAX_RELAY_LEN {
+        return None;
+    }
+    let relay = std::str::from_utf8(take(&mut bytes, len)?).ok()?;
+    if !relay.is_empty() {
+        addr = addr.with_relay_url(relay.parse().ok()?);
+    }
+    let count = *take(&mut bytes, 1)?.first()? as usize;
+    if count > DHT_MAX_ADDRS {
+        return None;
+    }
+    for _ in 0..count {
+        let ip = match *take(&mut bytes, 1)?.first()? {
+            4 => IpAddr::V4(Ipv4Addr::from(
+                <[u8; 4]>::try_from(take(&mut bytes, 4)?).ok()?,
+            )),
+            6 => IpAddr::V6(Ipv6Addr::from(
+                <[u8; 16]>::try_from(take(&mut bytes, 16)?).ok()?,
+            )),
+            _ => return None,
+        };
+        let port = u16::from_be_bytes(take(&mut bytes, 2)?.try_into().ok()?);
+        let socket = SocketAddr::new(ip, port);
+        if usable(&socket, allow_loopback) {
+            addr = addr.with_ip_addr(socket);
+        }
+    }
+    (bytes.is_empty() && !addr.addrs.is_empty()).then_some(addr)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::testing::network;
+
+    fn address() -> EndpointAddr {
+        EndpointAddr::new(iroh::SecretKey::generate().public())
+            .with_ip_addr("127.0.0.1:12345".parse().unwrap())
+    }
+
+    #[test]
+    fn records_reject_expiration_future_dates_truncation_and_oversize() {
+        let addr = address();
+        let bytes = encode_record(&addr, 1_000, true).unwrap();
+        assert_eq!(decode_record(&bytes, 1_100, true).unwrap(), addr);
+        assert!(decode_record(&bytes, 1_901, true).is_none());
+        assert!(decode_record(&bytes, 100, true).is_none());
+        for n in 0..bytes.len() {
+            assert!(decode_record(&bytes[..n], 1_100, true).is_none());
+        }
+        assert!(decode_record(&vec![0; DHT_MAX_VALUE + 1], 1_100, true).is_none());
+        assert!(encode_record(&addr, 1_000, false).is_err());
+        assert!(decode_record(&bytes, 1_100, false).is_none());
+    }
+
+    #[tokio::test]
+    async fn a_stopped_custom_client_never_falls_back_to_public_bootstrap() {
+        let net = mainline::Testnet::builder(3).build().unwrap();
+        let client = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let (name, secret) = network("mainline-shutdown");
+        let keys = NetworkKeys::derive(&name, &secret);
+        let backend = client.for_network(&keys);
+        client.shutdown().await;
+        assert!(backend.resolve(keys.discovery_key()).await.is_err());
+    }
+
+    #[tokio::test]
+    async fn concurrent_writers_share_slots_and_other_secrets_find_nothing() {
+        let net = mainline::Testnet::builder(5).build().unwrap();
+        let client_a = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let client_b = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let (name, secret) = network("mainline-slots");
+        let keys = NetworkKeys::derive(&name, &secret);
+        let a = client_a.for_network(&keys);
+        let b = client_b.for_network(&keys);
+        let first = address();
+        let second = loop {
+            let next = address();
+            if slots_for(next.id) == slots_for(first.id) {
+                break next;
+            }
+        };
+        let key = keys.discovery_key();
+        let (ra, rb) = tokio::join!(
+            a.publish(key, first.clone()),
+            b.publish(key, second.clone())
+        );
+        assert!(ra.is_ok() || rb.is_ok());
+        let (found_a, found_b) = tokio::join!(a.resolve(key), b.resolve(key));
+        let found_a = found_a.unwrap();
+        let found_b = found_b.unwrap();
+        assert!(
+            found_a.iter().any(|c| c.addr == second) || found_b.iter().any(|c| c.addr == first),
+            "at least one writer must discover the other despite colliding in both slots"
+        );
+        assert!(
+            found_a
+                .iter()
+                .chain(&found_b)
+                .all(|c| c.addr == first || c.addr == second)
+        );
+        let other = NetworkKeys::derive(&name, &crate::identity::NetworkSecret::generate());
+        assert!(
+            client_a
+                .for_network(&other)
+                .resolve(other.discovery_key())
+                .await
+                .unwrap()
+                .is_empty()
+        );
+        // Restarting the publisher must read seq from DHT instead of starting at 1.
+        let restarted = client_a.for_network(&keys);
+        restarted.publish(key, first.clone()).await.unwrap();
+        assert!(
+            b.resolve(key)
+                .await
+                .unwrap()
+                .iter()
+                .any(|c| c.addr == first)
+        );
+    }
+}

crates/tsunagi-core/src/discovery/worker.rs

@@ -0,0 +1,251 @@
+//! Per-network discovery tasks. Dropping the owner cancels all backend futures.
+
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::{mpsc, watch};
+use tokio::task::JoinSet;
+use tokio::time::{Instant, timeout};
+
+use super::{Candidate, NetworkDiscovery};
+use crate::config::DiscoveryPolicy;
+use crate::identity::DiscoveryKey;
+use crate::net::EndpointAdapter;
+
+pub(crate) struct DiscoveryWorker {
+    connected: watch::Sender<Connectivity>,
+    tasks: JoinSet<()>,
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    endpoint: iroh::EndpointId,
+}
+
+#[derive(Clone, Copy)]
+enum Connectivity {
+    Initial,
+    Connected,
+    Isolated(Instant),
+}
+
+impl DiscoveryWorker {
+    pub(crate) fn spawn(
+        backend: Arc<dyn NetworkDiscovery>,
+        key: DiscoveryKey,
+        adapter: EndpointAdapter,
+        policy: DiscoveryPolicy,
+        address_check: Duration,
+        candidates: mpsc::Sender<Candidate>,
+    ) -> Self {
+        let (connected, receiver) = watch::channel(Connectivity::Initial);
+        let mut tasks = JoinSet::new();
+        tasks.spawn(publish_loop(
+            backend.clone(),
+            key,
+            adapter.clone(),
+            policy.clone(),
+            address_check,
+        ));
+        tasks.spawn(lookup_loop(
+            backend.clone(),
+            key,
+            policy,
+            receiver,
+            candidates,
+        ));
+        Self {
+            connected,
+            tasks,
+            backend,
+            key,
+            endpoint: adapter.endpoint_id(),
+        }
+    }
+
+    pub(crate) fn set_connected(&self, connected: bool) {
+        self.connected.send_if_modified(|current| {
+            if matches!(*current, Connectivity::Connected) == connected {
+                return false;
+            }
+            *current = if connected {
+                Connectivity::Connected
+            } else {
+                Connectivity::Isolated(Instant::now())
+            };
+            true
+        });
+    }
+
+    pub(crate) async fn stop(mut self) {
+        self.tasks.abort_all();
+        while self.tasks.join_next().await.is_some() {}
+        // BEP44 has no deletion. Memory/static backends can withdraw promptly;
+        // an unresponsive backend must not extend agent shutdown.
+        let _ = timeout(
+            Duration::from_millis(250),
+            self.backend.unpublish(self.key, self.endpoint),
+        )
+        .await;
+    }
+}
+
+fn jitter(delay: Duration) -> Duration {
+    delay
+        .mul_f64(0.8 + 0.4 * rand::random::<f64>())
+        .max(Duration::from_millis(10))
+}
+
+async fn publish_loop(
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    adapter: EndpointAdapter,
+    policy: DiscoveryPolicy,
+    address_check: Duration,
+) {
+    let mut previous = None;
+    let mut due = Instant::now();
+    let mut ticker = tokio::time::interval(address_check.max(Duration::from_millis(10)));
+    ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+    loop {
+        ticker.tick().await;
+        let mut addr = adapter.addr();
+        addr.addrs.extend(adapter.loopback_addr().addrs);
+        if previous.as_ref() != Some(&addr) {
+            previous = Some(addr.clone());
+            due = Instant::now();
+        }
+        if Instant::now() < due {
+            continue;
+        }
+        let result = timeout(policy.request_timeout, backend.publish(key, addr.clone())).await;
+        match result {
+            Ok(Ok(())) => {
+                due = Instant::now() + jitter(policy.publish_interval);
+            }
+            result => {
+                tracing::debug!(?result, "discovery publication failed; will retry");
+                due = Instant::now() + jitter(policy.lookup_interval);
+            }
+        }
+    }
+}
+
+async fn lookup_loop(
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    policy: DiscoveryPolicy,
+    mut connected: watch::Receiver<Connectivity>,
+    candidates: mpsc::Sender<Candidate>,
+) {
+    let mut last_isolation = None;
+    let mut due = Instant::now();
+    let mut delay = policy.lookup_interval;
+    loop {
+        let state = *connected.borrow_and_update();
+        if matches!(state, Connectivity::Connected) {
+            if connected.changed().await.is_err() {
+                break;
+            }
+            continue;
+        }
+        if let Connectivity::Isolated(since) = state
+            && last_isolation != Some(since)
+        {
+            // Preserve the transition time even if a rapid reconnect/disconnect
+            // overwrote a watch value before this task had a chance to run.
+            due = since + policy.reconnect_delay;
+            delay = policy.lookup_interval;
+            last_isolation = Some(since);
+        }
+        tokio::select! {
+            biased;
+            changed = connected.changed() => {
+                if changed.is_err() { break; }
+                continue;
+            }
+            _ = tokio::time::sleep_until(due) => {}
+        }
+        tokio::select! {
+            biased;
+            changed = connected.changed() => {
+                if changed.is_err() { break; }
+                continue;
+            }
+            result = timeout(policy.request_timeout, backend.resolve_into(key, candidates.clone())) => {
+                if !matches!(result, Ok(Ok(()))) {
+                    tracing::debug!(?result, "discovery lookup failed; will retry");
+                }
+            }
+        }
+        due = Instant::now() + jitter(delay);
+        delay = delay.saturating_mul(2).min(policy.max_lookup_interval);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::Result;
+    use crate::discovery::BoxFuture;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    #[derive(Debug, Default)]
+    struct Reads(AtomicUsize);
+    impl NetworkDiscovery for Reads {
+        fn name(&self) -> &str {
+            "reads"
+        }
+        fn publish<'a>(
+            &'a self,
+            _: DiscoveryKey,
+            _: iroh::EndpointAddr,
+        ) -> BoxFuture<'a, Result<()>> {
+            Box::pin(async { Ok(()) })
+        }
+        fn unpublish<'a>(
+            &'a self,
+            _: DiscoveryKey,
+            _: iroh::EndpointId,
+        ) -> BoxFuture<'a, Result<()>> {
+            Box::pin(async { Ok(()) })
+        }
+        fn resolve<'a>(&'a self, _: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+            self.0.fetch_add(1, Ordering::SeqCst);
+            Box::pin(async { Ok(Vec::new()) })
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn recovery_requires_a_full_minute_and_reconnection_restarts_that_minute() {
+        let reads = Arc::new(Reads::default());
+        let (tx, rx) = watch::channel(Connectivity::Connected);
+        let (candidates, _receiver) = mpsc::channel(16);
+        let mut tasks = JoinSet::new();
+        tasks.spawn(lookup_loop(
+            reads.clone(),
+            DiscoveryKey::from_bytes([0; 32]),
+            DiscoveryPolicy::default(),
+            rx,
+            candidates,
+        ));
+        tokio::task::yield_now().await;
+        tx.send(Connectivity::Isolated(Instant::now())).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(59)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 0);
+        tx.send(Connectivity::Connected).unwrap();
+        tx.send(Connectivity::Isolated(Instant::now())).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(59)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 0);
+        tokio::time::advance(Duration::from_secs(1)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 1);
+        tx.send(Connectivity::Connected).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(600)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 1);
+    }
+}

crates/tsunagi-core/src/dns/mod.rs

@@ -0,0 +1,84 @@
+//! A DNS view of the overlay.
+//!
+//! The names and addresses of a network, served to the host that runs the
+//! agent, so members can be reached by name. Answered from signed state, so a
+//! member that is switched off still resolves.
+//!
+//! Three parts, kept apart on purpose:
+//!
+//! * [`zone`] decides what the answer is. Pure, and knows nothing about
+//!   packets or sockets.
+//! * [`server`] puts that on the wire.
+//! * `publish` tells the operating system where to send its questions,
+//!   which is the only part that differs between platforms.
+
+pub mod publish;
+pub mod server;
+pub mod zone;
+
+pub use publish::{DnsPublisher, PublishError, Published};
+
+use std::net::{Ipv4Addr, Ipv6Addr, SocketAddr};
+
+/// Where the server listens, one list per address family.
+///
+/// Loopback, and only loopback. The zones are a view *for the host running
+/// the agent*: it is that host's resolver that is pointed at them. Binding
+/// an overlay address instead would put them in front of the whole mesh —
+/// and an agent in several networks would then answer one network's
+/// questions about another's names, which is precisely what separate
+/// networks are for.
+///
+/// Both families, and independently: a question arrives over whichever one
+/// the resolver happens to use, and on a host with one of them switched off
+/// the other must still answer. So neither list failing says anything about
+/// the other, and a listener is opened from each.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ListenPlan {
+    /// IPv4 candidates, best first.
+    pub v4: Vec<SocketAddr>,
+    /// IPv6 candidates, best first.
+    pub v6: Vec<SocketAddr>,
+}
+
+impl ListenPlan {
+    /// The two lists, so a caller can bind one listener from each.
+    pub fn families(&self) -> [&[SocketAddr]; 2] {
+        [&self.v4, &self.v6]
+    }
+}
+
+/// Builds the listen plan for one port.
+pub fn listen_plan(port: u16) -> ListenPlan {
+    ListenPlan {
+        v4: vec![SocketAddr::from((Ipv4Addr::LOCALHOST, port))],
+        v6: vec![SocketAddr::from((Ipv6Addr::LOCALHOST, port))],
+    }
+}
+
+pub use server::{DnsServer, SharedZone};
+pub use zone::{Answer, Query, Zone, ZoneError, ZoneName, Zones};
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[test]
+    fn both_families_are_planned_for_and_neither_is_reachable_off_this_host() {
+        let plan = listen_plan(5354);
+        assert_eq!(
+            plan.v4,
+            vec!["127.0.0.1:5354".parse::<SocketAddr>().unwrap()]
+        );
+        assert_eq!(plan.v6, vec!["[::1]:5354".parse::<SocketAddr>().unwrap()]);
+        assert_eq!(plan.families().len(), 2);
+        // Nothing here is reachable from another member: the zones are a
+        // view for this host, and an agent in two networks must not answer
+        // one network's questions about the other's names.
+        for address in plan.families().concat() {
+            assert!(address.ip().is_loopback(), "{address} is not loopback");
+        }
+    }
+}

crates/tsunagi-core/src/dns/publish/macos.rs

@@ -0,0 +1,333 @@
+//! Telling the macOS resolver to send some questions here, through
+//! `/etc/resolver`.
+//!
+//! macOS reads one file per domain out of `/etc/resolver`: a file named after
+//! a domain, containing the `nameserver` and `port` to send that domain's
+//! questions to, makes the system resolver a *scoped* resolver for exactly
+//! that suffix and nothing else (see `resolver(5)`). That is the same contract
+//! the systemd-resolved and NRPT publishers arrange on the other platforms:
+//! route these suffixes to this server, and claim nothing else.
+//!
+//! # Why a port, and so no privileged socket
+//!
+//! A resolver file carries a `port`, so the server does not have to sit on 53
+//! and the agent needs nothing like `CAP_NET_BIND_SERVICE`. It can serve on a
+//! high port and still be reached for its suffixes.
+//!
+//! # Tagged, reversible, and never clobbering
+//!
+//! Every file this publisher writes begins with a marker line, so it only ever
+//! replaces or removes files it wrote itself: a pre-existing `/etc/resolver`
+//! entry a user set up by hand is left untouched. The files are removed when
+//! the resolver is turned off and on shutdown; a crash leaves them behind, and
+//! the next apply for the same domains overwrites them.
+//!
+//! # Privilege
+//!
+//! Writing under `/etc/resolver` needs root. A permission error is reported as
+//! a refusal — the kind of failure a person has to act on — rather than a
+//! transient one, so the agent does not retry it at the ordinary pace.
+
+use std::collections::BTreeSet;
+use std::net::IpAddr;
+use std::path::{Path, PathBuf};
+use std::sync::Mutex;
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// The directory macOS reads scoped resolvers from.
+const RESOLVER_DIR: &str = "/etc/resolver";
+
+/// The first line of every file this publisher writes, so it recognises its
+/// own and never touches a file a user created.
+const MARKER: &str = "# managed by tsunagi — do not edit";
+
+/// Writes and removes `/etc/resolver` files for the overlay zones.
+#[derive(Debug)]
+pub struct ResolverDirPublisher {
+    /// The resolver directory. A field rather than a constant so the tests can
+    /// point it at a temporary directory and never touch the real one.
+    dir: PathBuf,
+    /// The domain files written by the last apply, so a later apply can remove
+    /// the ones no longer wanted and `revert` can remove them all.
+    written: Mutex<BTreeSet<String>>,
+}
+
+impl Default for ResolverDirPublisher {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ResolverDirPublisher {
+    /// A publisher writing to the real `/etc/resolver`.
+    pub fn new() -> Self {
+        Self::in_dir(RESOLVER_DIR)
+    }
+
+    /// A publisher writing to a specific directory. For the tests.
+    fn in_dir(dir: impl Into<PathBuf>) -> Self {
+        Self {
+            dir: dir.into(),
+            written: Mutex::new(BTreeSet::new()),
+        }
+    }
+
+    fn apply_inner(&self, published: &Published) -> Result<(), PublishError> {
+        let desired = desired_files(published)?;
+
+        // The directory may not exist yet; creating it needs root.
+        std::fs::create_dir_all(&self.dir).map_err(|err| classify(&self.dir, err))?;
+
+        // Remove the files from a previous apply that are no longer wanted.
+        let previous = lock(&self.written).clone();
+        for name in previous.difference(&keys(&desired)) {
+            remove_if_ours(&self.dir.join(name));
+        }
+
+        for (name, body) in &desired {
+            let path = self.dir.join(name);
+            // Only ever overwrite a file that is already ours, or a name that
+            // is free. A resolver a user set up by hand is left in place and
+            // reported, never clobbered.
+            if path.exists() && !is_ours(&path) {
+                return Err(PublishError::Failed(format!(
+                    "{} already exists and was not created by tsunagi; \
+                     remove it or rename the network to publish its resolver",
+                    path.display()
+                )));
+            }
+            std::fs::write(&path, body).map_err(|err| classify(&path, err))?;
+        }
+
+        *lock(&self.written) = keys(&desired);
+        Ok(())
+    }
+
+    fn revert_inner(&self) {
+        let written = std::mem::take(&mut *lock(&self.written));
+        for name in &written {
+            remove_if_ours(&self.dir.join(name));
+        }
+    }
+}
+
+impl DnsPublisher for ResolverDirPublisher {
+    fn name(&self) -> &str {
+        "resolver-dir"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        // A handful of small file writes, short enough to run in place rather
+        // than hop to the blocking pool.
+        Box::pin(async move { self.apply_inner(published) })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            self.revert_inner();
+            Ok(())
+        })
+    }
+}
+
+/// The resolver files a publication implies: a map of file name to contents.
+fn desired_files(published: &Published) -> Result<Vec<(String, String)>, PublishError> {
+    if published.servers.is_empty() {
+        return Err(PublishError::Failed(
+            "no server address to point the resolver at".to_string(),
+        ));
+    }
+    let body = file_contents(published);
+    let mut files = Vec::new();
+    for domain in &published.domains {
+        let name = resolver_file_name(domain)?;
+        files.push((name, body.clone()));
+    }
+    Ok(files)
+}
+
+/// The contents of one resolver file: the marker, every server, one port.
+fn file_contents(published: &Published) -> String {
+    let mut out = String::from(MARKER);
+    out.push('\n');
+    for server in &published.servers {
+        let ip = match server.ip() {
+            IpAddr::V4(v4) => v4.to_string(),
+            IpAddr::V6(v6) => v6.to_string(),
+        };
+        out.push_str("nameserver ");
+        out.push_str(&ip);
+        out.push('\n');
+    }
+    // One port for the file; every server listens on the same one. The `port`
+    // directive is what lets the server stay off 53.
+    if let Some(first) = published.servers.first() {
+        out.push_str("port ");
+        out.push_str(&first.port().to_string());
+        out.push('\n');
+    }
+    out
+}
+
+/// The file name for a domain, rejecting anything that is not a plain label
+/// path component: a resolver file name must never escape the directory.
+fn resolver_file_name(domain: &str) -> Result<String, PublishError> {
+    let trimmed = domain.trim_end_matches('.');
+    if trimmed.is_empty()
+        || trimmed.contains('/')
+        || trimmed.contains('\\')
+        || trimmed.contains(std::path::MAIN_SEPARATOR)
+        || trimmed == "."
+        || trimmed == ".."
+        || trimmed.starts_with('.')
+    {
+        return Err(PublishError::Failed(format!(
+            "cannot make a resolver file name for the domain {domain:?}"
+        )));
+    }
+    Ok(trimmed.to_string())
+}
+
+/// Whether a file is one this publisher wrote, by its marker first line.
+fn is_ours(path: &Path) -> bool {
+    match std::fs::read_to_string(path) {
+        Ok(text) => text.lines().next() == Some(MARKER),
+        Err(_) => false,
+    }
+}
+
+/// Removes a file, but only if it is one of ours. Best effort.
+fn remove_if_ours(path: &Path) {
+    if is_ours(path)
+        && let Err(err) = std::fs::remove_file(path)
+        && err.kind() != std::io::ErrorKind::NotFound
+    {
+        tracing::debug!(path = %path.display(), %err, "cannot remove the resolver file");
+    }
+}
+
+/// The set of file names in a desired-files list.
+fn keys(files: &[(String, String)]) -> BTreeSet<String> {
+    files.iter().map(|(name, _)| name.clone()).collect()
+}
+
+/// Turns a filesystem error into the right kind of publish error: a permission
+/// problem is a refusal a person must act on, anything else is a plain failure.
+fn classify(path: &Path, err: std::io::Error) -> PublishError {
+    if err.kind() == std::io::ErrorKind::PermissionDenied {
+        PublishError::Refused(format!(
+            "writing {} needs root; start the agent with `sudo` or as a root LaunchDaemon",
+            path.display()
+        ))
+    } else {
+        PublishError::Failed(format!("cannot write {}: {err}", path.display()))
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::net::SocketAddr;
+
+    use super::*;
+
+    fn published(domains: &[&str]) -> Published {
+        Published {
+            interface: "utun5".into(),
+            servers: vec![
+                "10.13.37.1:5353".parse::<SocketAddr>().unwrap(),
+                "[fd00::1]:5353".parse::<SocketAddr>().unwrap(),
+            ],
+            domains: domains.iter().map(|d| d.to_string()).collect(),
+        }
+    }
+
+    #[test]
+    fn a_file_lists_every_server_and_one_port_under_the_marker() {
+        let body = file_contents(&published(&["mynet."]));
+        let mut lines = body.lines();
+        assert_eq!(lines.next(), Some(MARKER));
+        assert_eq!(lines.next(), Some("nameserver 10.13.37.1"));
+        assert_eq!(lines.next(), Some("nameserver fd00::1"));
+        assert_eq!(lines.next(), Some("port 5353"));
+        assert_eq!(lines.next(), None);
+    }
+
+    #[test]
+    fn a_domain_becomes_a_plain_file_name_without_its_trailing_dot() {
+        assert_eq!(resolver_file_name("mynet.").unwrap(), "mynet");
+        assert_eq!(resolver_file_name("my.net").unwrap(), "my.net");
+    }
+
+    #[test]
+    fn a_domain_that_would_escape_the_directory_is_refused() {
+        assert!(resolver_file_name("../etc/passwd").is_err());
+        assert!(resolver_file_name("a/b").is_err());
+        assert!(resolver_file_name("").is_err());
+        assert!(resolver_file_name(".").is_err());
+        assert!(resolver_file_name(".hidden").is_err());
+    }
+
+    #[tokio::test]
+    async fn apply_writes_a_file_per_domain_and_revert_removes_them() {
+        let dir = tempfile::tempdir().unwrap();
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+
+        publisher
+            .apply(&published(&["mynet.", "other."]))
+            .await
+            .unwrap();
+        assert!(dir.path().join("mynet").is_file());
+        assert!(dir.path().join("other").is_file());
+
+        publisher.revert().await.unwrap();
+        assert!(!dir.path().join("mynet").exists());
+        assert!(!dir.path().join("other").exists());
+    }
+
+    #[tokio::test]
+    async fn re_applying_removes_a_domain_that_is_no_longer_wanted() {
+        let dir = tempfile::tempdir().unwrap();
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+
+        publisher.apply(&published(&["a.", "b."])).await.unwrap();
+        publisher.apply(&published(&["a."])).await.unwrap();
+
+        assert!(dir.path().join("a").is_file());
+        assert!(
+            !dir.path().join("b").exists(),
+            "the dropped domain's file is removed"
+        );
+    }
+
+    #[tokio::test]
+    async fn a_file_a_user_created_is_never_clobbered_or_removed() {
+        let dir = tempfile::tempdir().unwrap();
+        let foreign = dir.path().join("mynet");
+        std::fs::write(&foreign, "nameserver 9.9.9.9\n").unwrap();
+
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+        let err = publisher.apply(&published(&["mynet."])).await.unwrap_err();
+        assert!(matches!(err, PublishError::Failed(_)), "{err:?}");
+        // Untouched.
+        assert_eq!(
+            std::fs::read_to_string(&foreign).unwrap(),
+            "nameserver 9.9.9.9\n"
+        );
+
+        // And revert leaves a foreign file alone even if it is in our set.
+        publisher.revert().await.unwrap();
+        assert!(foreign.is_file());
+    }
+}

crates/tsunagi-core/src/dns/publish/mock.rs

@@ -0,0 +1,105 @@
+//! A pretend resolver, so the wiring is tested without touching this host's.
+
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// Records what it was asked to do, and can be told to refuse.
+#[derive(Debug, Clone, Default)]
+pub struct MockPublisher {
+    applied: Arc<Mutex<Option<Published>>>,
+    failure: Option<PublishError>,
+}
+
+impl MockPublisher {
+    /// A publisher that accepts everything.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// A publisher that refuses everything, with this reason.
+    pub fn failing(error: PublishError) -> Self {
+        Self {
+            applied: Arc::new(Mutex::new(None)),
+            failure: Some(error),
+        }
+    }
+
+    /// What is currently applied, if anything.
+    pub fn applied(&self) -> Option<Published> {
+        match self.applied.lock() {
+            Ok(guard) => guard.clone(),
+            Err(poisoned) => poisoned.into_inner().clone(),
+        }
+    }
+}
+
+impl DnsPublisher for MockPublisher {
+    fn name(&self) -> &str {
+        "mock"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        Box::pin(async move {
+            if let Some(failure) = &self.failure {
+                return Err(failure.clone());
+            }
+            match self.applied.lock() {
+                Ok(mut guard) => *guard = Some(published.clone()),
+                Err(poisoned) => *poisoned.into_inner() = Some(published.clone()),
+            }
+            Ok(())
+        })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            match self.applied.lock() {
+                Ok(mut guard) => *guard = None,
+                Err(poisoned) => *poisoned.into_inner() = None,
+            }
+            Ok(())
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn published() -> Published {
+        Published {
+            interface: "tsundemo".into(),
+            servers: vec![
+                "10.13.37.69:5354".parse().unwrap(),
+                "[::1]:5354".parse().unwrap(),
+            ],
+            domains: vec!["lab".into()],
+        }
+    }
+
+    #[tokio::test]
+    async fn applying_and_reverting_are_both_recorded() {
+        let publisher = MockPublisher::new();
+        assert!(publisher.applied().is_none());
+
+        publisher.apply(&published()).await.unwrap();
+        assert_eq!(publisher.applied(), Some(published()));
+
+        publisher.revert().await.unwrap();
+        assert!(publisher.applied().is_none());
+        // Reverting twice is not an error: shutdown must not fail here.
+        publisher.revert().await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn a_refusal_leaves_nothing_applied() {
+        let publisher = MockPublisher::failing(PublishError::Refused("polkit said no".into()));
+        assert!(publisher.apply(&published()).await.is_err());
+        assert!(publisher.applied().is_none());
+    }
+}

crates/tsunagi-core/src/dns/publish/mod.rs

@@ -0,0 +1,215 @@
+//! Telling the operating system where to send its questions.
+//!
+//! The server answers whether or not this works. That is the whole reason it
+//! is a separate thing: if the resolver cannot be configured — no
+//! systemd-resolved, an unwilling polkit, a platform nobody has written this
+//! for — the port is still up and the user can point something at it by
+//! hand. A failure here is a degraded overlay, not a broken one.
+//!
+//! Only the *mechanics* differ between systems. What has to be arranged is
+//! the same everywhere: send questions for these suffixes to this address,
+//! through this interface, and do not make it the resolver for anything
+//! else. That is [`Published`]; the rest is behind [`DnsPublisher`].
+
+use std::net::SocketAddr;
+
+use crate::BoxFuture;
+
+mod mock;
+pub use mock::MockPublisher;
+
+#[cfg(all(feature = "dns-publish", target_os = "linux"))]
+mod resolved;
+#[cfg(all(feature = "dns-publish", target_os = "linux"))]
+pub use resolved::ResolvedPublisher;
+
+#[cfg(all(feature = "dns-publish", target_os = "windows"))]
+mod windows;
+#[cfg(all(feature = "dns-publish", target_os = "windows"))]
+pub use windows::NrptPublisher;
+
+#[cfg(all(feature = "dns-publish", target_os = "macos"))]
+mod macos;
+#[cfg(all(feature = "dns-publish", target_os = "macos"))]
+pub use macos::ResolverDirPublisher;
+
+mod unsupported;
+pub use unsupported::UnsupportedPublisher;
+
+/// What the operating system is asked to do.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Published {
+    /// The interface questions should be sent through.
+    ///
+    /// The server listens on an overlay address, which is only reachable
+    /// over the overlay interface, so the two travel together. It must be an
+    /// interface this agent created: configuring one it did not is
+    /// configuring somebody else's.
+    pub interface: String,
+    /// Every address the server is listening on, best first.
+    ///
+    /// One per address family, so a resolver reaches it over whichever it
+    /// uses. All of them are published together and none is preferred by
+    /// this: which one a resolver picks is its business.
+    pub servers: Vec<SocketAddr>,
+    /// The suffixes that belong to this server.
+    ///
+    /// Routing suffixes only: they say *which questions* come here, never
+    /// that this is the resolver for anything else.
+    pub domains: Vec<String>,
+}
+
+/// Why the resolver could not be told.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum PublishError {
+    /// There is nothing here that can be configured this way.
+    #[error("{0}")]
+    Unavailable(String),
+    /// Something is there, and it declined.
+    ///
+    /// Separate from a plain failure because the answer is different: this
+    /// one is about who the agent is running as, not about whether the thing
+    /// works.
+    #[error("{0}")]
+    Refused(String),
+    /// It was there, it accepted the request, and it went wrong anyway.
+    #[error("{0}")]
+    Failed(String),
+}
+
+impl PublishError {
+    /// Whether waiting will fix it.
+    ///
+    /// A refusal will not change on its own — somebody has to grant
+    /// permission — so retrying it at the pace of everything else is just
+    /// noise. Anything else might be a service still starting.
+    pub fn needs_a_human(&self) -> bool {
+        matches!(self, PublishError::Refused(_))
+    }
+
+    /// One line for a status table.
+    pub fn remedy(&self) -> Option<&'static str> {
+        match self {
+            PublishError::Refused(_) => Some(
+                "grant this user the `org.freedesktop.resolve1.set-*` actions in \
+                 /etc/polkit-1/rules.d, or run the agent as a system service",
+            ),
+            _ => None,
+        }
+    }
+}
+
+/// The polkit rule that lets this user configure the resolver.
+///
+/// Printed in full rather than described, because the point of this feature
+/// is that the user has as little to do as possible, and "write a polkit
+/// rule" is a great deal more work than pasting one.
+///
+/// It grants exactly the four actions this agent calls and nothing else.
+/// polkit decides by user id — a capability does not help here — so there is
+/// no way to do this from inside the process.
+pub fn polkit_recipe(user: &str) -> String {
+    format!(
+        "sudo tee /etc/polkit-1/rules.d/50-tsunagi-resolved.rules > /dev/null <<'RULE'\n\
+         polkit.addRule(function(action, subject) {{\n\
+         \x20   var allowed = [\n\
+         \x20       \"org.freedesktop.resolve1.set-dns-servers\",\n\
+         \x20       \"org.freedesktop.resolve1.set-domains\",\n\
+         \x20       \"org.freedesktop.resolve1.set-default-route\",\n\
+         \x20       \"org.freedesktop.resolve1.revert\"\n\
+         \x20   ];\n\
+         \x20   if (allowed.indexOf(action.id) >= 0 && subject.user == \"{user}\") {{\n\
+         \x20       return polkit.Result.YES;\n\
+         \x20   }}\n\
+         }});\n\
+         RULE"
+    )
+}
+
+/// Arranges for the operating system to ask this server.
+pub trait DnsPublisher: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Applies the setting, replacing whatever this publisher set before.
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>>;
+
+    /// Undoes it.
+    ///
+    /// Reverting something that was never applied succeeds: this runs on the
+    /// shutdown path, where the setting being gone is the point.
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>>;
+}
+
+/// The kernel's index for an interface.
+///
+/// From sysfs, which needs no privileges and no netlink round trip.
+#[cfg(target_os = "linux")]
+pub fn interface_index(name: &str) -> Option<u32> {
+    std::fs::read_to_string(format!("/sys/class/net/{name}/ifindex"))
+        .ok()?
+        .trim()
+        .parse()
+        .ok()
+}
+
+/// The kernel's index for an interface.
+///
+/// Only the systemd-resolved publisher needs one, so off Linux there is
+/// nothing to look up and this is always `None`.
+#[cfg(not(target_os = "linux"))]
+pub fn interface_index(_name: &str) -> Option<u32> {
+    None
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[test]
+    fn only_a_refusal_suggests_what_to_do_about_it() {
+        // The other two are conditions of the host, not of the user.
+        assert!(PublishError::Refused("no".into()).remedy().is_some());
+        assert!(PublishError::Unavailable("none".into()).remedy().is_none());
+        assert!(PublishError::Failed("bang".into()).remedy().is_none());
+    }
+
+    #[test]
+    fn only_a_refusal_waits_for_a_person() {
+        // The rest may come right on their own, so they are worth retrying
+        // at the ordinary pace; a refusal is not.
+        assert!(PublishError::Refused("no".into()).needs_a_human());
+        assert!(!PublishError::Unavailable("none".into()).needs_a_human());
+        assert!(!PublishError::Failed("bang".into()).needs_a_human());
+    }
+
+    #[test]
+    fn the_polkit_recipe_grants_what_is_called_and_no_more() {
+        let recipe = polkit_recipe("ab");
+        for action in [
+            "set-dns-servers",
+            "set-domains",
+            "set-default-route",
+            "revert",
+        ] {
+            assert!(recipe.contains(action), "{action} missing from:\n{recipe}");
+        }
+        // Nothing beyond what the agent calls: a rule that granted the lot
+        // would be handing out more than this feature needs.
+        for other in ["set-dnssec", "set-mdns", "register-service", "set-llmnr"] {
+            assert!(!recipe.contains(other), "{other} should not be granted");
+        }
+        assert!(recipe.contains("subject.user == \"ab\""));
+        // ES5: the rules engine is duktape and has no `startsWith`.
+        assert!(!recipe.contains("startsWith"));
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn an_interface_index_is_read_from_the_running_kernel() {
+        assert_eq!(interface_index("lo"), Some(1));
+        assert_eq!(interface_index("tsunagi-no-such-interface"), None);
+    }
+}

crates/tsunagi-core/src/dns/publish/resolved.rs

@@ -0,0 +1,275 @@
+//! Telling systemd-resolved to send some questions here.
+//!
+//! Three calls on `org.freedesktop.resolve1.Manager`, all scoped to the
+//! overlay interface:
+//!
+//! * `SetLinkDNSEx` — where to send them. The `Ex` form carries a **port**,
+//!   which is why this server does not have to sit on 53 and the agent needs
+//!   no `CAP_NET_BIND_SERVICE`. Systems before systemd 247 have only
+//!   `SetLinkDNS`, which has no port; there the fallback only works if the
+//!   server did get port 53, and it says so rather than appearing to work.
+//! * `SetLinkDomains` with `routing_only` set — a *routing* suffix, the
+//!   `~domain` form. It says which questions come here and claims nothing
+//!   else.
+//! * `SetLinkDefaultRoute(false)` — so this never becomes the resolver for
+//!   anything outside those suffixes. Without it resolved may fall back to
+//!   this link for ordinary names, and this server refuses those.
+//!
+//! # It cleans up by itself
+//!
+//! resolved keys all of this to the interface, and drops it when the
+//! interface goes. The overlay interface belongs to a file descriptor the
+//! agent holds, so it goes when the agent does — however the agent goes. The
+//! explicit `RevertLink` on shutdown only makes that immediate.
+//!
+//! # Privilege
+//!
+//! resolved asks polkit, and polkit decides by user id, not by capability.
+//! So `CAP_NET_ADMIN` does not help here: an ordinary user is prompted or
+//! refused, while a system service running as root is not. That refusal is
+//! reported as its own kind of error, because the answer to it is different
+//! from the answer to "resolved is not installed".
+
+use std::net::IpAddr;
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published, interface_index};
+
+/// `AF_INET`, as resolved wants it.
+const AF_INET: i32 = 2;
+/// `AF_INET6`.
+const AF_INET6: i32 = 10;
+
+#[zbus::proxy(
+    interface = "org.freedesktop.resolve1.Manager",
+    default_service = "org.freedesktop.resolve1",
+    default_path = "/org/freedesktop/resolve1"
+)]
+trait Resolved {
+    /// Servers for a link, with a port and a name. systemd 247 and later.
+    #[zbus(name = "SetLinkDNSEx")]
+    fn set_link_dns_ex(
+        &self,
+        ifindex: i32,
+        addresses: &[(i32, Vec<u8>, u16, String)],
+    ) -> zbus::Result<()>;
+
+    /// Servers for a link, without a port. Always port 53.
+    #[zbus(name = "SetLinkDNS")]
+    fn set_link_dns(&self, ifindex: i32, addresses: &[(i32, Vec<u8>)]) -> zbus::Result<()>;
+
+    /// Suffixes for a link. The flag makes one routing-only.
+    #[zbus(name = "SetLinkDomains")]
+    fn set_link_domains(&self, ifindex: i32, domains: &[(String, bool)]) -> zbus::Result<()>;
+
+    /// Whether this link may answer for names outside its suffixes.
+    #[zbus(name = "SetLinkDefaultRoute")]
+    fn set_link_default_route(&self, ifindex: i32, enable: bool) -> zbus::Result<()>;
+
+    /// Forgets everything set for a link.
+    #[zbus(name = "RevertLink")]
+    fn revert_link(&self, ifindex: i32) -> zbus::Result<()>;
+}
+
+/// Configures systemd-resolved over D-Bus.
+#[derive(Debug, Default)]
+pub struct ResolvedPublisher {
+    /// The link last configured, so shutdown knows what to undo.
+    applied: std::sync::Mutex<Option<u32>>,
+}
+
+impl ResolvedPublisher {
+    /// Creates the publisher. Nothing is contacted until [`Self::apply`].
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    async fn proxy() -> Result<ResolvedProxy<'static>, PublishError> {
+        let connection = zbus::Connection::system().await.map_err(|err| {
+            PublishError::Unavailable(format!("no system D-Bus to talk to resolved on: {err}"))
+        })?;
+        ResolvedProxy::new(&connection).await.map_err(|err| {
+            PublishError::Unavailable(format!("systemd-resolved is not answering: {err}"))
+        })
+    }
+}
+
+/// Turns a D-Bus failure into the kind of failure it actually is.
+fn classify(err: zbus::Error, what: &str) -> PublishError {
+    let name = match &err {
+        zbus::Error::MethodError(name, _, _) => name.as_str().to_string(),
+        _ => String::new(),
+    };
+    match name.as_str() {
+        "org.freedesktop.DBus.Error.InteractiveAuthorizationRequired"
+        | "org.freedesktop.DBus.Error.AccessDenied" => {
+            PublishError::Refused(format!("systemd-resolved refused {what}: {err}"))
+        }
+        "org.freedesktop.DBus.Error.UnknownMethod"
+        | "org.freedesktop.DBus.Error.ServiceUnknown" => {
+            PublishError::Unavailable(format!("systemd-resolved cannot do {what}: {err}"))
+        }
+        _ => PublishError::Failed(format!("systemd-resolved failed {what}: {err}")),
+    }
+}
+
+/// The address in the shape resolved wants: a family and raw octets.
+fn wire_address(address: IpAddr) -> (i32, Vec<u8>) {
+    match address {
+        IpAddr::V4(address) => (AF_INET, address.octets().to_vec()),
+        IpAddr::V6(address) => (AF_INET6, address.octets().to_vec()),
+    }
+}
+
+impl DnsPublisher for ResolvedPublisher {
+    fn name(&self) -> &str {
+        "systemd-resolved"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        Box::pin(async move {
+            let ifindex = interface_index(&published.interface).ok_or_else(|| {
+                PublishError::Unavailable(format!(
+                    "interface `{}` is not on this host, so there is no link to configure",
+                    published.interface
+                ))
+            })?;
+            if published.servers.is_empty() {
+                return Err(PublishError::Unavailable(
+                    "there is no listening address to send questions to".to_string(),
+                ));
+            }
+            let proxy = Self::proxy().await?;
+            let index = ifindex as i32;
+            // Every family in one call, because this replaces the link's
+            // whole list: sending them one at a time would leave only the
+            // last. A resolver then picks whichever it can reach.
+            let servers: Vec<(i32, Vec<u8>, u16, String)> = published
+                .servers
+                .iter()
+                .map(|server| {
+                    let (family, octets) = wire_address(server.ip());
+                    (family, octets, server.port(), String::new())
+                })
+                .collect();
+
+            match proxy.set_link_dns_ex(index, &servers).await {
+                Ok(()) => {}
+                Err(err) => {
+                    let classified = classify(err, "setting the link's DNS servers");
+                    // Older systemd has no `Ex` form, and the plain one is
+                    // always port 53. Falling back to it when the server is
+                    // somewhere else would point resolved at nothing.
+                    if !matches!(classified, PublishError::Unavailable(_)) {
+                        return Err(classified);
+                    }
+                    if let Some(elsewhere) =
+                        published.servers.iter().find(|server| server.port() != 53)
+                    {
+                        return Err(PublishError::Unavailable(format!(
+                            "this systemd-resolved cannot be given a port, and the server is on \
+                             {}. Run the server on port 53, or point your resolver at \
+                             {elsewhere} yourself.",
+                            elsewhere.port()
+                        )));
+                    }
+                    let plain: Vec<(i32, Vec<u8>)> = servers
+                        .into_iter()
+                        .map(|(family, octets, _, _)| (family, octets))
+                        .collect();
+                    proxy
+                        .set_link_dns(index, &plain)
+                        .await
+                        .map_err(|err| classify(err, "setting the link's DNS servers"))?;
+                }
+            }
+
+            let domains: Vec<(String, bool)> = published
+                .domains
+                .iter()
+                .map(|domain| (domain.clone(), true))
+                .collect();
+            proxy
+                .set_link_domains(index, &domains)
+                .await
+                .map_err(|err| classify(err, "setting the link's search domains"))?;
+
+            // Last, and deliberately: until this is off, resolved may send
+            // ordinary names here, and this server refuses them.
+            proxy
+                .set_link_default_route(index, false)
+                .await
+                .map_err(|err| classify(err, "clearing the link's default route"))?;
+
+            match self.applied.lock() {
+                Ok(mut guard) => *guard = Some(ifindex),
+                Err(poisoned) => *poisoned.into_inner() = Some(ifindex),
+            }
+            Ok(())
+        })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            let applied = match self.applied.lock() {
+                Ok(mut guard) => guard.take(),
+                Err(poisoned) => poisoned.into_inner().take(),
+            };
+            let Some(ifindex) = applied else {
+                return Ok(());
+            };
+            let proxy = Self::proxy().await?;
+            match proxy.revert_link(ifindex as i32).await {
+                Ok(()) => Ok(()),
+                // The link going away takes the setting with it, so this is
+                // the outcome asked for rather than a failure.
+                Err(zbus::Error::MethodError(name, _, _))
+                    if name.as_str().contains("NoSuchLink") =>
+                {
+                    Ok(())
+                }
+                Err(err) => Err(classify(err, "reverting the link")),
+            }
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::{Ipv4Addr, SocketAddr};
+
+    #[test]
+    fn an_address_is_encoded_the_way_resolved_expects() {
+        assert_eq!(
+            wire_address(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 69))),
+            (AF_INET, vec![10, 13, 37, 69])
+        );
+        let (family, octets) = wire_address("fd55::1".parse().unwrap());
+        assert_eq!(family, AF_INET6);
+        assert_eq!(octets.len(), 16);
+    }
+
+    #[tokio::test]
+    async fn an_interface_that_is_not_here_is_unavailable_not_a_failure() {
+        // Nothing is contacted: there is no link to configure, and saying so
+        // is the honest answer without bothering the bus.
+        let publisher = ResolvedPublisher::new();
+        let published = Published {
+            interface: "tsunagi-no-such-interface".into(),
+            servers: vec![SocketAddr::from(([10, 13, 37, 69], 5354))],
+            domains: vec!["lab".into()],
+        };
+        let err = publisher.apply(&published).await.unwrap_err();
+        assert!(matches!(err, PublishError::Unavailable(_)), "{err}");
+    }
+
+    #[tokio::test]
+    async fn reverting_without_having_applied_does_nothing_and_succeeds() {
+        // The shutdown path must not fail because there was nothing to undo.
+        ResolvedPublisher::new().revert().await.unwrap();
+    }
+}

crates/tsunagi-core/src/dns/publish/unsupported.rs

@@ -0,0 +1,73 @@
+//! The publisher for systems nobody has written one for.
+//!
+//! macOS needs `SystemConfiguration` and Windows the IP Helper API, and
+//! neither is written. Saying so is more use than appearing to work: the
+//! server is already answering, so all that is missing is the last step, and
+//! the user can take it by hand once they know that is what is needed.
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// Refuses to configure anything, with an explanation.
+#[derive(Debug, Clone)]
+pub struct UnsupportedPublisher {
+    platform: &'static str,
+}
+
+impl Default for UnsupportedPublisher {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl UnsupportedPublisher {
+    /// A publisher naming the platform it stands in for.
+    pub fn new() -> Self {
+        Self {
+            platform: std::env::consts::OS,
+        }
+    }
+}
+
+impl DnsPublisher for UnsupportedPublisher {
+    fn name(&self) -> &str {
+        "unsupported"
+    }
+
+    fn apply<'a>(&'a self, _published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        Box::pin(async move {
+            Err(PublishError::Unavailable(format!(
+                "configuring the system resolver is not implemented on {} yet",
+                self.platform
+            )))
+        })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        // Nothing was set, so there is nothing to undo and no reason to fail
+        // a shutdown over it.
+        Box::pin(async move { Ok(()) })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn it_names_the_platform_and_still_reverts_quietly() {
+        let publisher = UnsupportedPublisher::new();
+        let published = Published {
+            interface: "tsundemo".into(),
+            servers: vec!["10.0.0.1:5354".parse().unwrap()],
+            domains: vec!["lab".into()],
+        };
+        let err = publisher.apply(&published).await.unwrap_err();
+        assert!(matches!(err, PublishError::Unavailable(_)));
+        assert!(err.to_string().contains(std::env::consts::OS));
+        publisher.revert().await.unwrap();
+    }
+}

crates/tsunagi-core/src/dns/publish/windows.rs

@@ -0,0 +1,301 @@
+//! Telling the Windows DNS client to send some questions here, through the
+//! Name Resolution Policy Table.
+//!
+//! The NRPT is how Windows does split DNS: a rule says "names under this
+//! suffix are resolved by these servers", and names outside every rule are
+//! resolved the ordinary way. That is exactly the contract of [`Published`] —
+//! route these suffixes here and claim nothing else — so one NRPT rule per
+//! suffix is the whole of it. The rules are keyed by namespace, not by
+//! interface, so the interface name in [`Published`] is not needed here.
+//!
+//! # Why PowerShell
+//!
+//! The NRPT lives in the registry, but a rule only takes effect once the DNS
+//! client is told to reload its policy, and that notification is an RPC with
+//! no safe wrapper this crate may call — it forbids `unsafe`. The
+//! `DnsClient` PowerShell module does the write *and* the reload, so it is
+//! used by absolute path from `%SystemRoot%`. Every value passed to it is one
+//! this agent produced: the suffixes are validated zone names (letters,
+//! digits, `-`, `_`, `.` only) and the servers are addresses it is listening
+//! on. They are still single-quoted and escaped when built into the script,
+//! so nothing could be read as PowerShell rather than as data.
+//!
+//! # The port
+//!
+//! The Windows DNS client always asks on port 53; it cannot be told another.
+//! So a server on a different port cannot be reached this way, and rather than
+//! configure a rule that points at nothing, this says so — the same honest
+//! failure the systemd-resolved path gives for an old resolver with no port.
+//!
+//! # Privilege
+//!
+//! Writing an NRPT rule needs an elevated process. An ordinary user is
+//! refused, which is reported as its own kind of error because the answer to
+//! it — run elevated, or as a service — differs from "this is not Windows".
+
+use std::collections::BTreeSet;
+use std::net::IpAddr;
+use std::sync::Mutex;
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// The port the Windows DNS client is fixed to.
+const DNS_PORT: u16 = 53;
+
+/// Configures the Name Resolution Policy Table through PowerShell.
+#[derive(Debug, Default)]
+pub struct NrptPublisher {
+    /// The namespaces last configured, so shutdown knows what to undo.
+    applied: Mutex<Vec<String>>,
+}
+
+impl NrptPublisher {
+    /// Creates the publisher. Nothing is contacted until [`Self::apply`].
+    pub fn new() -> Self {
+        Self::default()
+    }
+}
+
+/// A suffix as the NRPT wants it: a leading dot means "this and everything
+/// under it".
+fn namespace(domain: &str) -> String {
+    let trimmed = domain.trim_matches('.');
+    format!(".{trimmed}")
+}
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// A PowerShell array literal, `@('a','b')`, or `@()` when empty.
+fn ps_array(values: &[String]) -> String {
+    let items: Vec<String> = values.iter().map(|value| ps_literal(value)).collect();
+    format!("@({})", items.join(","))
+}
+
+/// The script that clears our rules for these namespaces and then adds them.
+fn apply_script(namespaces: &[String], servers: &[String]) -> String {
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $ns = {ns}\n\
+         $srv = {srv}\n\
+         foreach ($n in $ns) {{ Get-DnsClientNrptRule | Where-Object {{ $_.Namespace -eq $n }} | \
+         ForEach-Object {{ Remove-DnsClientNrptRule -Name $_.Name -Force }} }}\n\
+         foreach ($n in $ns) {{ Add-DnsClientNrptRule -Namespace $n -NameServers $srv }}\n",
+        ns = ps_array(namespaces),
+        srv = ps_array(servers),
+    )
+}
+
+/// The script that clears our rules for these namespaces.
+fn revert_script(namespaces: &[String]) -> String {
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $ns = {ns}\n\
+         foreach ($n in $ns) {{ Get-DnsClientNrptRule | Where-Object {{ $_.Namespace -eq $n }} | \
+         ForEach-Object {{ Remove-DnsClientNrptRule -Name $_.Name -Force }} }}\n",
+        ns = ps_array(namespaces),
+    )
+}
+
+/// Turns what PowerShell said on failure into the kind of failure it is.
+fn classify(text: &str) -> PublishError {
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("access is denied")
+        || lower.contains("requires elevation")
+        || lower.contains("run as administrator")
+        || lower.contains("administrator privilege")
+        || lower.contains("permissiondenied")
+    {
+        PublishError::Refused(format!("the Windows DNS client refused the change: {text}"))
+    } else if lower.contains("is not recognized")
+        || lower.contains("commandnotfoundexception")
+        || lower.contains("not recognized as the name of a cmdlet")
+    {
+        PublishError::Unavailable(format!(
+            "the DnsClient PowerShell module is not available: {text}"
+        ))
+    } else {
+        PublishError::Failed(format!("the Windows DNS client failed the change: {text}"))
+    }
+}
+
+/// Runs a PowerShell script from `%SystemRoot%` and maps a failure to a
+/// [`PublishError`].
+async fn powershell(script: String) -> Result<(), PublishError> {
+    let program = powershell_path();
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program)
+            .args([
+                "-NoProfile",
+                "-NonInteractive",
+                "-ExecutionPolicy",
+                "Bypass",
+                "-Command",
+                &script,
+            ])
+            .output()
+    })
+    .await
+    .map_err(|err| PublishError::Failed(format!("could not run PowerShell: {err}")))?
+    .map_err(|err| PublishError::Unavailable(format!("could not run PowerShell: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = if stderr.trim().is_empty() {
+        String::from_utf8_lossy(&output.stdout).trim().to_string()
+    } else {
+        stderr.trim().to_string()
+    };
+    Err(classify(if text.is_empty() { "no output" } else { &text }))
+}
+
+/// The absolute path to Windows PowerShell, so nothing on `PATH` can stand in
+/// for it.
+fn powershell_path() -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root)
+        .join("System32")
+        .join("WindowsPowerShell")
+        .join("v1.0")
+        .join("powershell.exe")
+}
+
+impl DnsPublisher for NrptPublisher {
+    fn name(&self) -> &str {
+        "windows-nrpt"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        Box::pin(async move {
+            if published.domains.is_empty() {
+                return Err(PublishError::Unavailable(
+                    "there are no suffixes to route, so there is no rule to write".to_string(),
+                ));
+            }
+
+            // Windows can only ask on port 53, so a server anywhere else is
+            // unreachable this way. Keep the ones it can use.
+            let servers: BTreeSet<IpAddr> = published
+                .servers
+                .iter()
+                .filter(|server| server.port() == DNS_PORT)
+                .map(|server| server.ip())
+                .collect();
+            if servers.is_empty() {
+                let elsewhere = published
+                    .servers
+                    .iter()
+                    .map(|server| server.to_string())
+                    .collect::<Vec<_>>()
+                    .join(", ");
+                return Err(PublishError::Unavailable(format!(
+                    "the Windows DNS client only asks on port {DNS_PORT}, and the server is on \
+                     {elsewhere}. Run the server on port {DNS_PORT}, or point your resolver at it \
+                     yourself."
+                )));
+            }
+
+            let namespaces: Vec<String> = published.domains.iter().map(|d| namespace(d)).collect();
+            let servers: Vec<String> = servers.iter().map(|ip| ip.to_string()).collect();
+
+            powershell(apply_script(&namespaces, &servers)).await?;
+
+            match self.applied.lock() {
+                Ok(mut guard) => *guard = namespaces,
+                Err(poisoned) => *poisoned.into_inner() = namespaces,
+            }
+            Ok(())
+        })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            let namespaces = match self.applied.lock() {
+                Ok(mut guard) => std::mem::take(&mut *guard),
+                Err(poisoned) => std::mem::take(&mut *poisoned.into_inner()),
+            };
+            if namespaces.is_empty() {
+                return Ok(());
+            }
+            powershell(revert_script(&namespaces)).await
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::SocketAddr;
+
+    #[test]
+    fn a_suffix_gets_a_leading_dot() {
+        assert_eq!(namespace("lab"), ".lab");
+        assert_eq!(namespace(".lab"), ".lab");
+        assert_eq!(namespace("a.b"), ".a.b");
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        // Zone names never contain a quote, but the escaping is what makes
+        // that a guarantee rather than a hope.
+        assert_eq!(ps_literal("a'b"), "'a''b'");
+        assert_eq!(ps_array(&["x".into(), "y".into()]), "@('x','y')");
+        assert_eq!(ps_array(&[]), "@()");
+    }
+
+    #[test]
+    fn the_apply_script_removes_before_it_adds() {
+        let script = apply_script(&[".lab".into()], &["10.13.37.69".into()]);
+        let remove = script.find("Remove-DnsClientNrptRule").unwrap();
+        let add = script.find("Add-DnsClientNrptRule").unwrap();
+        assert!(
+            remove < add,
+            "a stale rule must go before the new one:\n{script}"
+        );
+        assert!(script.contains("@('.lab')"));
+        assert!(script.contains("@('10.13.37.69')"));
+    }
+
+    #[test]
+    fn an_access_denied_is_a_refusal_a_missing_cmdlet_is_unavailable() {
+        assert!(matches!(
+            classify("Access is denied"),
+            PublishError::Refused(_)
+        ));
+        assert!(matches!(
+            classify("The term 'Add-DnsClientNrptRule' is not recognized"),
+            PublishError::Unavailable(_)
+        ));
+        assert!(matches!(
+            classify("something else"),
+            PublishError::Failed(_)
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_server_only_on_a_nonstandard_port_is_unavailable_not_a_failure() {
+        // Nothing is run: Windows cannot ask there, and saying so is the
+        // honest answer without touching the DNS client.
+        let publisher = NrptPublisher::new();
+        let published = Published {
+            interface: "tsun0".into(),
+            servers: vec![SocketAddr::from(([10, 13, 37, 69], 5354))],
+            domains: vec!["lab".into()],
+        };
+        let err = publisher.apply(&published).await.unwrap_err();
+        assert!(matches!(err, PublishError::Unavailable(_)), "{err}");
+    }
+
+    #[tokio::test]
+    async fn reverting_without_having_applied_does_nothing_and_succeeds() {
+        NrptPublisher::new().revert().await.unwrap();
+    }
+}

crates/tsunagi-core/src/dns/server.rs

@@ -0,0 +1,625 @@
+//! Answering DNS questions about the overlay, over UDP and TCP.
+//!
+//! Authoritative for one zone and nothing else. There is no recursion, no
+//! forwarding and no cache: a question this agent cannot answer from the
+//! signed roster is refused rather than passed anywhere, so pointing a
+//! resolver at this server can never make it a path to the outside.
+//!
+//! The decision of what to answer lives in [`super::zone`]; this module is
+//! only the wire format and the sockets. [`respond`] sits between them and
+//! takes bytes to bytes, so everything the server does to a packet is
+//! testable without opening a socket.
+
+use std::net::SocketAddr;
+use std::sync::{Arc, RwLock};
+
+use simple_dns::rdata::{A, PTR, RData, SOA};
+use simple_dns::{Name, PacketFlag, QCLASS, QTYPE, RCODE, ResourceRecord, TYPE};
+
+use super::zone::{Answer, Query, Zone, Zones};
+
+/// How long an answer may be cached.
+///
+/// Short, because the roster changes when members come and go and a stale
+/// answer is worse than another question.
+pub const TTL: u32 = 30;
+
+/// The largest question this server will read.
+///
+/// A DNS message is 512 bytes without EDNS and 4096 with it; anything past
+/// that is not a question worth answering.
+pub const MAX_MESSAGE_LEN: usize = 4096;
+
+/// The largest answer sent over UDP without the client offering EDNS.
+const CLASSIC_UDP_LIMIT: usize = 512;
+
+/// How long a TCP client may take over one question.
+const TCP_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(5);
+
+/// How many TCP questions may be in flight at once.
+const MAX_TCP_CONNECTIONS: usize = 32;
+
+/// The zones the server answers from, swapped as the rosters change.
+///
+/// Shared rather than copied into the server so that a roster change — or a
+/// network joining or leaving — is one write, not a restart: rebinding the
+/// socket would drop questions in flight for no reason.
+#[derive(Debug, Clone, Default)]
+pub struct SharedZone(Arc<RwLock<Arc<Zones>>>);
+
+impl SharedZone {
+    /// Wraps a set of zones.
+    pub fn new(zones: Zones) -> Self {
+        Self(Arc::new(RwLock::new(Arc::new(zones))))
+    }
+
+    /// Wraps a single zone, which is the common case in a test.
+    pub fn one(zone: Zone) -> Self {
+        Self::new(Zones::new([zone]))
+    }
+
+    /// Replaces them.
+    pub fn set(&self, zones: Zones) {
+        match self.0.write() {
+            Ok(mut guard) => *guard = Arc::new(zones),
+            Err(poisoned) => *poisoned.into_inner() = Arc::new(zones),
+        }
+    }
+
+    /// The zones as they are now.
+    pub fn get(&self) -> Arc<Zones> {
+        match self.0.read() {
+            Ok(guard) => Arc::clone(&guard),
+            Err(poisoned) => Arc::clone(&poisoned.into_inner()),
+        }
+    }
+}
+
+/// Builds the answer to one question.
+///
+/// `None` means say nothing at all: the message was not a question this
+/// server should reply to, and replying anyway would make this a useful
+/// amplifier for somebody spoofing a source address.
+pub fn respond(zones: &Zones, query: &[u8]) -> Option<Vec<u8>> {
+    let packet = simple_dns::Packet::parse(query).ok()?;
+    if packet.has_flags(PacketFlag::RESPONSE) {
+        return None;
+    }
+
+    let mut reply = simple_dns::Packet::new_reply(packet.id());
+    // Recursion is not available here and the flag says so honestly; the
+    // desired bit is echoed because a resolver compares it.
+    if packet.has_flags(PacketFlag::RECURSION_DESIRED) {
+        reply.set_flags(PacketFlag::RECURSION_DESIRED);
+    }
+
+    if packet.opcode() != simple_dns::OPCODE::StandardQuery {
+        *reply.rcode_mut() = RCODE::NotImplemented;
+        return reply.build_bytes_vec().ok();
+    }
+
+    // Exactly one question. Zero is nothing to answer; more than one has no
+    // agreed meaning and every real server rejects it.
+    let [question] = packet.questions.as_slice() else {
+        *reply.rcode_mut() = RCODE::FormatError;
+        return reply.build_bytes_vec().ok();
+    };
+    if !matches!(question.qclass, QCLASS::CLASS(simple_dns::CLASS::IN)) {
+        *reply.rcode_mut() = RCODE::Refused;
+        return reply.build_bytes_vec().ok();
+    }
+
+    let qname = question.qname.to_string();
+    // Which zone answers is part of the answer: a denial is bounded by the
+    // SOA of the zone that denied it, and this server may hold several.
+    let (zone, answer) = match zones.lookup(&qname, query_kind(question.qtype)) {
+        Some(answered) => answered,
+        None => {
+            reply.questions.push(question.clone());
+            *reply.rcode_mut() = RCODE::Refused;
+            return reply.build_bytes_vec_compressed().ok();
+        }
+    };
+    reply.questions.push(question.clone());
+
+    let name = Name::new(&qname).ok()?;
+    match answer {
+        Answer::NotOurs => *reply.rcode_mut() = RCODE::Refused,
+        Answer::Addresses(addresses) => {
+            reply.set_flags(PacketFlag::AUTHORITATIVE_ANSWER);
+            for address in addresses {
+                reply.answers.push(ResourceRecord::new(
+                    name.clone(),
+                    simple_dns::CLASS::IN,
+                    TTL,
+                    RData::A(A::from(address)),
+                ));
+            }
+        }
+        Answer::Name(target) => {
+            reply.set_flags(PacketFlag::AUTHORITATIVE_ANSWER);
+            let target = Name::new(&target).ok()?.into_owned();
+            reply.answers.push(ResourceRecord::new(
+                name.clone(),
+                simple_dns::CLASS::IN,
+                TTL,
+                RData::PTR(PTR(target)),
+            ));
+        }
+        Answer::Soa => {
+            reply.set_flags(PacketFlag::AUTHORITATIVE_ANSWER);
+            reply.answers.push(soa_record(zone)?);
+        }
+        Answer::NoData => {
+            reply.set_flags(PacketFlag::AUTHORITATIVE_ANSWER);
+            // The authority section carries the SOA so a resolver knows how
+            // long it may remember that there is nothing here.
+            reply.name_servers.push(soa_record(zone)?);
+        }
+        Answer::NoSuchName => {
+            reply.set_flags(PacketFlag::AUTHORITATIVE_ANSWER);
+            *reply.rcode_mut() = RCODE::NameError;
+            reply.name_servers.push(soa_record(zone)?);
+        }
+    }
+
+    reply.build_bytes_vec_compressed().ok()
+}
+
+/// The zone's start of authority.
+fn soa_record(zone: &Zone) -> Option<ResourceRecord<'static>> {
+    let origin = Name::new(zone.origin().as_str()).ok()?.into_owned();
+    Some(ResourceRecord::new(
+        origin.clone(),
+        simple_dns::CLASS::IN,
+        TTL,
+        RData::SOA(SOA {
+            mname: origin.clone(),
+            // There is no mailbox behind this zone and inventing one would
+            // be a fiction; the origin itself is the honest answer.
+            rname: origin,
+            serial: zone.serial(),
+            refresh: TTL as i32,
+            retry: TTL as i32,
+            expire: 86_400,
+            minimum: TTL,
+        }),
+    ))
+}
+
+fn query_kind(qtype: QTYPE) -> Query {
+    match qtype {
+        QTYPE::TYPE(TYPE::A) => Query::A,
+        QTYPE::TYPE(TYPE::PTR) => Query::Ptr,
+        QTYPE::TYPE(TYPE::SOA) => Query::Soa,
+        QTYPE::TYPE(TYPE::NS) => Query::Ns,
+        // ANY is answered as an address question rather than by dumping the
+        // zone: an ANY that returns everything is an amplification gift.
+        QTYPE::ANY => Query::A,
+        _ => Query::Other,
+    }
+}
+
+/// Whether the client offered EDNS, and so how large an answer it will take.
+fn udp_limit(query: &[u8]) -> usize {
+    simple_dns::Packet::parse(query)
+        .ok()
+        .and_then(|packet| packet.opt().map(|opt| opt.udp_packet_size as usize))
+        .unwrap_or(CLASSIC_UDP_LIMIT as u16 as usize)
+        .clamp(CLASSIC_UDP_LIMIT, MAX_MESSAGE_LEN)
+}
+
+/// Cuts an answer down to what the client said it would take.
+///
+/// The records are dropped and the truncated bit set, which tells a resolver
+/// to ask again over TCP. Sending a reply it cannot reassemble would just
+/// look like packet loss.
+fn truncate_for_udp(query: &[u8], reply: Vec<u8>) -> Vec<u8> {
+    let limit = udp_limit(query);
+    if reply.len() <= limit {
+        return reply;
+    }
+    let Ok(parsed) = simple_dns::Packet::parse(&reply) else {
+        return reply;
+    };
+    let mut short = simple_dns::Packet::new_reply(parsed.id());
+    short.set_flags(PacketFlag::TRUNCATION | PacketFlag::AUTHORITATIVE_ANSWER);
+    *short.rcode_mut() = parsed.rcode();
+    for question in &parsed.questions {
+        short.questions.push(question.clone());
+    }
+    short.build_bytes_vec().unwrap_or(reply)
+}
+
+/// A running DNS server.
+#[derive(Debug)]
+pub struct DnsServer {
+    local_addr: SocketAddr,
+    tasks: Vec<tokio::task::JoinHandle<()>>,
+}
+
+impl DnsServer {
+    /// Binds and starts answering.
+    ///
+    /// Both transports on the same address and port, as a resolver expects:
+    /// it falls back to TCP when an answer does not fit, and a server that
+    /// only listened on UDP would leave it with nowhere to go.
+    pub async fn bind(addr: SocketAddr, zone: SharedZone) -> std::io::Result<Self> {
+        let udp = tokio::net::UdpSocket::bind(addr).await?;
+        let local_addr = udp.local_addr()?;
+        let tcp = tokio::net::TcpListener::bind(local_addr).await?;
+
+        let udp_zone = zone.clone();
+        let udp_task = tokio::spawn(async move {
+            let mut buffer = vec![0u8; MAX_MESSAGE_LEN];
+            loop {
+                let (read, from) = match udp.recv_from(&mut buffer).await {
+                    Ok(pair) => pair,
+                    Err(err) => {
+                        tracing::debug!(%err, "dns udp receive failed");
+                        continue;
+                    }
+                };
+                let query = &buffer[..read];
+                let Some(reply) = respond(&udp_zone.get(), query) else {
+                    continue;
+                };
+                let reply = truncate_for_udp(query, reply);
+                if let Err(err) = udp.send_to(&reply, from).await {
+                    tracing::debug!(%err, "dns udp reply failed");
+                }
+            }
+        });
+
+        let tcp_zone = zone.clone();
+        let tcp_task = tokio::spawn(async move {
+            let permits = Arc::new(tokio::sync::Semaphore::new(MAX_TCP_CONNECTIONS));
+            loop {
+                let (stream, _) = match tcp.accept().await {
+                    Ok(pair) => pair,
+                    Err(err) => {
+                        tracing::debug!(%err, "dns tcp accept failed");
+                        continue;
+                    }
+                };
+                let Ok(permit) = Arc::clone(&permits).acquire_owned().await else {
+                    return;
+                };
+                let zone = tcp_zone.clone();
+                tokio::spawn(async move {
+                    let _permit = permit;
+                    // Bounded, so a client that connects and says nothing
+                    // cannot hold a slot open.
+                    let _ = tokio::time::timeout(TCP_TIMEOUT, serve_tcp(stream, zone)).await;
+                });
+            }
+        });
+
+        Ok(Self {
+            local_addr,
+            tasks: vec![udp_task, tcp_task],
+        })
+    }
+
+    /// The address it is answering on.
+    pub fn local_addr(&self) -> SocketAddr {
+        self.local_addr
+    }
+}
+
+impl Drop for DnsServer {
+    fn drop(&mut self) {
+        for task in &self.tasks {
+            task.abort();
+        }
+    }
+}
+
+async fn serve_tcp(mut stream: tokio::net::TcpStream, zone: SharedZone) -> std::io::Result<()> {
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+    loop {
+        let mut header = [0u8; 2];
+        if stream.read_exact(&mut header).await.is_err() {
+            return Ok(());
+        }
+        // Checked before the buffer is allocated, as everywhere else that
+        // reads a length off a wire.
+        let len = u16::from_be_bytes(header) as usize;
+        if len == 0 || len > MAX_MESSAGE_LEN {
+            return Ok(());
+        }
+        let mut query = vec![0u8; len];
+        stream.read_exact(&mut query).await?;
+
+        let Some(reply) = respond(&zone.get(), &query) else {
+            return Ok(());
+        };
+        let Ok(len) = u16::try_from(reply.len()) else {
+            return Ok(());
+        };
+        stream.write_all(&len.to_be_bytes()).await?;
+        stream.write_all(&reply).await?;
+        stream.flush().await?;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use crate::dns::zone::ZoneName;
+    use std::net::Ipv4Addr;
+
+    fn zone() -> Zones {
+        Zones::new([one_zone()])
+    }
+
+    fn one_zone() -> Zone {
+        Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [
+                ("music".to_string(), Ipv4Addr::new(10, 13, 37, 237)),
+                ("ai".to_string(), Ipv4Addr::new(10, 13, 37, 69)),
+            ],
+        )
+    }
+
+    fn ask(name: &str, qtype: TYPE) -> Vec<u8> {
+        let mut packet = simple_dns::Packet::new_query(0x1234);
+        packet.questions.push(simple_dns::Question::new(
+            Name::new(name).unwrap(),
+            qtype.into(),
+            QCLASS::CLASS(simple_dns::CLASS::IN),
+            false,
+        ));
+        packet.build_bytes_vec().unwrap()
+    }
+
+    /// The bytes of a reply. Parsed by each caller, because a parsed packet
+    /// borrows from them.
+    fn answer(query: &[u8]) -> Vec<u8> {
+        respond(&zone(), query).expect("a reply")
+    }
+
+    #[test]
+    fn a_member_is_answered_authoritatively() {
+        let bytes = answer(&ask("music.lab", TYPE::A));
+        let reply = simple_dns::Packet::parse(&bytes).unwrap();
+        assert_eq!(reply.rcode(), RCODE::NoError);
+        assert!(reply.has_flags(PacketFlag::RESPONSE));
+        assert!(reply.has_flags(PacketFlag::AUTHORITATIVE_ANSWER));
+        assert!(!reply.has_flags(PacketFlag::RECURSION_AVAILABLE));
+        assert_eq!(reply.answers.len(), 1);
+        assert_eq!(reply.questions.len(), 1, "the question is echoed");
+        match &reply.answers[0].rdata {
+            RData::A(a) => assert_eq!(Ipv4Addr::from(a.address), Ipv4Addr::new(10, 13, 37, 237)),
+            other => panic!("expected an A record, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn an_absent_name_is_denied_with_a_soa_to_cache_the_denial() {
+        let bytes = answer(&ask("nobody.lab", TYPE::A));
+        let reply = simple_dns::Packet::parse(&bytes).unwrap();
+        assert_eq!(reply.rcode(), RCODE::NameError);
+        assert!(reply.answers.is_empty());
+        assert_eq!(
+            reply.name_servers.len(),
+            1,
+            "a SOA bounds the negative cache"
+        );
+    }
+
+    #[test]
+    fn a_name_that_exists_without_that_record_is_not_denied() {
+        // NODATA, not NXDOMAIN: denying the name would stop a resolver
+        // asking for the A record it could have had.
+        let bytes = answer(&ask("music.lab", TYPE::AAAA));
+        let reply = simple_dns::Packet::parse(&bytes).unwrap();
+        assert_eq!(reply.rcode(), RCODE::NoError);
+        assert!(reply.answers.is_empty());
+        assert_eq!(reply.name_servers.len(), 1);
+    }
+
+    #[test]
+    fn anything_outside_the_zone_is_refused_and_never_forwarded() {
+        for name in ["example.com", "evillab", "google.com"] {
+            let bytes = answer(&ask(name, TYPE::A));
+            let reply = simple_dns::Packet::parse(&bytes).unwrap();
+            assert_eq!(reply.rcode(), RCODE::Refused, "{name}");
+            assert!(reply.answers.is_empty());
+        }
+    }
+
+    #[test]
+    fn an_address_is_answered_backwards() {
+        let bytes = answer(&ask("237.37.13.10.in-addr.arpa", TYPE::PTR));
+        let reply = simple_dns::Packet::parse(&bytes).unwrap();
+        assert_eq!(reply.rcode(), RCODE::NoError);
+        match &reply.answers[0].rdata {
+            RData::PTR(ptr) => assert_eq!(ptr.0.to_string(), "music.lab"),
+            other => panic!("expected a PTR record, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn a_reply_is_never_sent_to_something_that_was_not_a_question() {
+        // Answering a response would make this a reflector for anyone who
+        // can spoof a source address.
+        let mut packet = simple_dns::Packet::new_reply(1);
+        packet.set_flags(PacketFlag::RESPONSE);
+        assert!(respond(&zone(), &packet.build_bytes_vec().unwrap()).is_none());
+        assert!(respond(&zone(), b"").is_none());
+        assert!(respond(&zone(), b"not a dns packet at all").is_none());
+    }
+
+    #[test]
+    fn a_question_with_no_question_in_it_is_a_format_error() {
+        let packet = simple_dns::Packet::new_query(7);
+        let bytes = respond(&zone(), &packet.build_bytes_vec().unwrap()).unwrap();
+        let reply = simple_dns::Packet::parse(&bytes).unwrap();
+        assert_eq!(reply.rcode(), RCODE::FormatError);
+    }
+
+    #[test]
+    fn a_class_other_than_internet_is_refused() {
+        let mut packet = simple_dns::Packet::new_query(9);
+        packet.questions.push(simple_dns::Question::new(
+            Name::new("music.lab").unwrap(),
+            TYPE::A.into(),
+            QCLASS::CLASS(simple_dns::CLASS::CH),
+            false,
+        ));
+        let bytes = respond(&zone(), &packet.build_bytes_vec().unwrap()).unwrap();
+        assert_eq!(
+            simple_dns::Packet::parse(&bytes).unwrap().rcode(),
+            RCODE::Refused
+        );
+    }
+
+    #[test]
+    fn an_answer_too_large_for_udp_is_truncated_rather_than_dropped() {
+        // A resolver that gets a truncated reply asks again over TCP; one
+        // that gets nothing back just waits.
+        let many: Vec<(String, Ipv4Addr)> = (0..200)
+            .map(|i| ("host".to_string(), Ipv4Addr::new(10, 13, 37, i as u8)))
+            .collect();
+        let wide = Zones::new([Zone::new(ZoneName::new("lab").unwrap(), many)]);
+        let query = ask("host.lab", TYPE::A);
+        let full = respond(&wide, &query).unwrap();
+        assert!(
+            full.len() > CLASSIC_UDP_LIMIT,
+            "the test needs a big answer"
+        );
+
+        let short = truncate_for_udp(&query, full);
+        assert!(short.len() <= CLASSIC_UDP_LIMIT);
+        let parsed = simple_dns::Packet::parse(&short).unwrap();
+        assert!(parsed.has_flags(PacketFlag::TRUNCATION));
+        assert_eq!(parsed.questions.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn the_server_answers_over_udp_and_tcp_on_one_address() {
+        use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+        let shared = SharedZone::new(zone());
+        let server = DnsServer::bind("127.0.0.1:0".parse().unwrap(), shared.clone())
+            .await
+            .unwrap();
+        let addr = server.local_addr();
+
+        let client = tokio::net::UdpSocket::bind("127.0.0.1:0").await.unwrap();
+        client
+            .send_to(&ask("music.lab", TYPE::A), addr)
+            .await
+            .unwrap();
+        let mut buffer = vec![0u8; MAX_MESSAGE_LEN];
+        let read = client.recv(&mut buffer).await.unwrap();
+        let reply = simple_dns::Packet::parse(&buffer[..read]).unwrap();
+        assert_eq!(reply.answers.len(), 1);
+
+        let mut stream = tokio::net::TcpStream::connect(addr).await.unwrap();
+        let query = ask("ai.lab", TYPE::A);
+        stream
+            .write_all(&(query.len() as u16).to_be_bytes())
+            .await
+            .unwrap();
+        stream.write_all(&query).await.unwrap();
+        let mut header = [0u8; 2];
+        stream.read_exact(&mut header).await.unwrap();
+        let mut body = vec![0u8; u16::from_be_bytes(header) as usize];
+        stream.read_exact(&mut body).await.unwrap();
+        let reply = simple_dns::Packet::parse(&body).unwrap();
+        assert_eq!(reply.answers.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn the_same_zone_is_answered_over_ipv6_as_over_ipv4() {
+        // A question arrives over whichever family the resolver chooses, so
+        // a listener is opened from each list in the plan and both have to
+        // answer the same thing. Answering one family only leaves the other
+        // timing out, which looks like a broken overlay.
+        let shared = SharedZone::new(zone());
+        let plan = crate::dns::listen_plan(0);
+        let mut answered = 0;
+
+        for family in plan.families() {
+            let candidate = family[0];
+            let server = match DnsServer::bind(candidate, shared.clone()).await {
+                Ok(server) => server,
+                // A host with IPv6 switched off in the kernel cannot bind
+                // `::1`, and the agent copes with that by design; so does
+                // this. The other family still has to work.
+                Err(err) if candidate.is_ipv6() => {
+                    eprintln!("no IPv6 loopback on this host ({err}); skipping that family");
+                    continue;
+                }
+                Err(err) => panic!("cannot listen on {candidate}: {err}"),
+            };
+            let addr = server.local_addr();
+            assert_eq!(addr.is_ipv6(), candidate.is_ipv6());
+
+            let client = tokio::net::UdpSocket::bind(if addr.is_ipv6() {
+                "[::1]:0"
+            } else {
+                "127.0.0.1:0"
+            })
+            .await
+            .unwrap();
+            client
+                .send_to(&ask("music.lab", TYPE::A), addr)
+                .await
+                .unwrap();
+            let mut buffer = vec![0u8; MAX_MESSAGE_LEN];
+            let read = client.recv(&mut buffer).await.unwrap();
+            let reply = simple_dns::Packet::parse(&buffer[..read]).unwrap();
+            assert_eq!(reply.rcode(), RCODE::NoError, "over {addr}");
+            match &reply.answers[0].rdata {
+                // The same IPv4 answer either way: the family a question
+                // travelled over says nothing about what the answer is.
+                RData::A(a) => {
+                    assert_eq!(Ipv4Addr::from(a.address), Ipv4Addr::new(10, 13, 37, 237))
+                }
+                other => panic!("expected an A record over {addr}, got {other:?}"),
+            }
+            answered += 1;
+        }
+
+        assert!(answered > 0, "at least one family must have answered");
+    }
+
+    #[tokio::test]
+    async fn replacing_the_zone_changes_what_the_running_server_answers() {
+        let shared = SharedZone::one(Zone::new(ZoneName::new("lab").unwrap(), []));
+        let server = DnsServer::bind("127.0.0.1:0".parse().unwrap(), shared.clone())
+            .await
+            .unwrap();
+        let addr = server.local_addr();
+        let client = tokio::net::UdpSocket::bind("127.0.0.1:0").await.unwrap();
+        let mut buffer = vec![0u8; MAX_MESSAGE_LEN];
+
+        client
+            .send_to(&ask("music.lab", TYPE::A), addr)
+            .await
+            .unwrap();
+        let read = client.recv(&mut buffer).await.unwrap();
+        assert_eq!(
+            simple_dns::Packet::parse(&buffer[..read]).unwrap().rcode(),
+            RCODE::NameError
+        );
+
+        // A member joins: no rebind, no dropped socket.
+        shared.set(zone());
+        client
+            .send_to(&ask("music.lab", TYPE::A), addr)
+            .await
+            .unwrap();
+        let read = client.recv(&mut buffer).await.unwrap();
+        let reply = simple_dns::Packet::parse(&buffer[..read]).unwrap();
+        assert_eq!(reply.rcode(), RCODE::NoError);
+        assert_eq!(reply.answers.len(), 1);
+    }
+}

crates/tsunagi-core/src/dns/zone.rs

@@ -0,0 +1,749 @@
+//! What the overlay answers to, as a DNS zone.
+//!
+//! Nothing here touches a socket or the operating system, and it does not
+//! depend on a DNS wire library either: it takes a roster and a question and
+//! says what the answer is. That is what makes the interesting parts — which
+//! names exist, what "does not exist" means as against "exists with nothing
+//! of that type", and what is outside the zone entirely — testable on their
+//! own.
+//!
+//! # Where the names come from
+//!
+//! From signed state, which is why a member that is switched off still
+//! resolves. Its claim outlived the session, so the name and the address are
+//! both still there to answer with. Nothing is invented for a member that
+//! claimed neither.
+//!
+//! Only IPv4 is served. The IPv6 overlay address is derived from a
+//! WireGuard key that travels in live announcements and is not in signed
+//! state, so it cannot be answered for a member that is away — and answering
+//! for some members and not others depending on whether they happen to be
+//! online is worse than not answering at all.
+
+use std::collections::BTreeMap;
+use std::net::Ipv4Addr;
+
+/// Longest a DNS name may be, in the presentation form used here.
+const MAX_NAME_LEN: usize = 253;
+/// Longest one label may be.
+const MAX_LABEL_LEN: usize = 63;
+
+/// Why a zone name cannot be used.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum ZoneError {
+    /// The name was empty, or became empty once normalised.
+    #[error("a zone name must have at least one label")]
+    Empty,
+    /// One label was unusable.
+    #[error("`{label}` is not a usable DNS label: {reason}")]
+    Label {
+        /// The offending label.
+        label: String,
+        /// What is wrong with it.
+        reason: &'static str,
+    },
+    /// The whole name is too long.
+    #[error("a zone name must be at most {MAX_NAME_LEN} characters")]
+    TooLong,
+}
+
+/// A validated, canonical zone name.
+///
+/// Held without a trailing dot and lower-cased, so comparison is a plain
+/// string comparison rather than a special case at every use.
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct ZoneName(String);
+
+impl ZoneName {
+    /// Validates and normalises a zone name.
+    pub fn new(raw: &str) -> Result<Self, ZoneError> {
+        let trimmed = raw.trim().trim_end_matches('.').to_ascii_lowercase();
+        if trimmed.is_empty() {
+            return Err(ZoneError::Empty);
+        }
+        if trimmed.len() > MAX_NAME_LEN {
+            return Err(ZoneError::TooLong);
+        }
+        for label in trimmed.split('.') {
+            let reason = if label.is_empty() {
+                Some("it is empty")
+            } else if label.len() > MAX_LABEL_LEN {
+                Some("it is longer than 63 characters")
+            } else if !label
+                .bytes()
+                .all(|b| b.is_ascii_alphanumeric() || b == b'-' || b == b'_')
+            {
+                Some("only letters, digits, `-` and `_` are allowed")
+            } else if label.starts_with('-') || label.ends_with('-') {
+                Some("a label may not start or end with `-`")
+            } else {
+                None
+            };
+            if let Some(reason) = reason {
+                return Err(ZoneError::Label {
+                    label: label.to_string(),
+                    reason,
+                });
+            }
+        }
+        Ok(Self(trimmed))
+    }
+
+    /// The name, without a trailing dot.
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+
+    /// The last label, which is what would collide with a real top-level
+    /// domain.
+    pub fn top_label(&self) -> &str {
+        self.0.rsplit('.').next().unwrap_or(&self.0)
+    }
+
+    /// What is worrying about this zone name, if anything.
+    ///
+    /// A warning and never a refusal: the name is the user's to choose, and
+    /// a private zone that shadows a public one is a decision, not a
+    /// mistake. Saying nothing would let it be an accident.
+    pub fn collision(&self) -> Option<String> {
+        let top = self.top_label();
+        // Reserved for exactly this use and never delegated, so nothing to
+        // say. See RFC 6761 and RFC 8375.
+        const RESERVED: &[&str] = &[
+            "internal",
+            "home",
+            "test",
+            "example",
+            "invalid",
+            "localhost",
+        ];
+        if RESERVED.contains(&top) {
+            return None;
+        }
+        if top == "local" {
+            return Some(
+                "`.local` belongs to multicast DNS: on a host running Avahi or \
+                 systemd-resolved's mDNS, names under it are resolved by that and \
+                 not by this agent"
+                    .to_string(),
+            );
+        }
+        // A real top-level domain is the whole of it. A name that is not
+        // delegated today might be one day — and saying so fires on every
+        // private name anybody ever picks, which is a warning about
+        // nothing that teaches people to ignore warnings.
+        tld::exist_case_insensitive(top).then(|| {
+            format!(
+                "`.{top}` is a real top-level domain, so every public name under it \
+                 becomes unreachable from this host while the overlay is up"
+            )
+        })
+    }
+
+    /// Whether `name` is this zone or sits under it.
+    ///
+    /// Compared label-wise, so `evilzone` does not count as being under
+    /// `zone`.
+    pub fn covers(&self, name: &str) -> bool {
+        let name = name.trim_end_matches('.').to_ascii_lowercase();
+        name == self.0
+            || name
+                .strip_suffix(&self.0)
+                .is_some_and(|rest| rest.ends_with('.'))
+    }
+
+    /// The part of `name` below this zone, if it is under it.
+    fn relative(&self, name: &str) -> Option<String> {
+        let name = name.trim_end_matches('.').to_ascii_lowercase();
+        if name == self.0 {
+            return Some(String::new());
+        }
+        let rest = name.strip_suffix(&self.0)?;
+        let rest = rest.strip_suffix('.')?;
+        (!rest.is_empty()).then(|| rest.to_string())
+    }
+}
+
+/// The kinds of question this zone knows how to answer.
+///
+/// Its own enum rather than the wire library's, so the decision of what to
+/// answer does not depend on how a packet is encoded.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Query {
+    /// An IPv4 address.
+    A,
+    /// A name for an address.
+    Ptr,
+    /// The zone's start of authority.
+    Soa,
+    /// The zone's name servers.
+    Ns,
+    /// Anything else, including AAAA.
+    Other,
+}
+
+/// What the zone has to say.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum Answer {
+    /// Addresses for the name asked about.
+    Addresses(Vec<Ipv4Addr>),
+    /// A name for the address asked about.
+    Name(String),
+    /// The zone's start of authority.
+    Soa,
+    /// The name exists here but has nothing of the type asked for.
+    ///
+    /// Distinct from [`Answer::NoSuchName`] because the two are different
+    /// answers on the wire: this one is a success with no records, and a
+    /// resolver must not take it as proof the name is absent. Getting them
+    /// the wrong way round teaches a resolver to cache the wrong thing.
+    NoData,
+    /// No such name in this zone.
+    NoSuchName,
+    /// Not a name this zone is responsible for.
+    NotOurs,
+}
+
+/// The names and addresses of one network, ready to answer questions.
+#[derive(Debug, Clone)]
+pub struct Zone {
+    origin: ZoneName,
+    /// Name relative to the origin, to the addresses it answers to.
+    hosts: BTreeMap<String, Vec<Ipv4Addr>>,
+    /// Address to the name that holds it.
+    names: BTreeMap<Ipv4Addr, String>,
+    /// Changes whenever the contents do.
+    serial: u32,
+}
+
+impl Zone {
+    /// Builds a zone from the members that have both a name and an address.
+    ///
+    /// A member with one but not the other contributes nothing: a name with
+    /// no address cannot be answered, and an address with no name has nothing
+    /// to be asked about.
+    pub fn new(origin: ZoneName, members: impl IntoIterator<Item = (String, Ipv4Addr)>) -> Self {
+        let mut hosts: BTreeMap<String, Vec<Ipv4Addr>> = BTreeMap::new();
+        let mut names: BTreeMap<Ipv4Addr, String> = BTreeMap::new();
+        for (hostname, address) in members {
+            let hostname = hostname.trim_matches('.').to_ascii_lowercase();
+            if hostname.is_empty() {
+                continue;
+            }
+            let addresses = hosts.entry(hostname.clone()).or_default();
+            if !addresses.contains(&address) {
+                addresses.push(address);
+            }
+            // First name wins, and the map is ordered, so the reverse answer
+            // is the same on every replica rather than depending on the order
+            // records happened to arrive in.
+            names.entry(address).or_insert(hostname);
+        }
+        for addresses in hosts.values_mut() {
+            addresses.sort();
+        }
+
+        let serial = content_serial(&hosts);
+        Self {
+            origin,
+            hosts,
+            names,
+            serial,
+        }
+    }
+
+    /// The zone's origin.
+    pub fn origin(&self) -> &ZoneName {
+        &self.origin
+    }
+
+    /// A number that changes whenever the contents do.
+    pub fn serial(&self) -> u32 {
+        self.serial
+    }
+
+    /// How many names it answers for.
+    pub fn len(&self) -> usize {
+        self.hosts.len()
+    }
+
+    /// Whether it answers for nothing.
+    pub fn is_empty(&self) -> bool {
+        self.hosts.is_empty()
+    }
+
+    /// Every name it answers for, with its addresses.
+    pub fn entries(&self) -> impl Iterator<Item = (&str, &[Ipv4Addr])> {
+        self.hosts
+            .iter()
+            .map(|(name, addresses)| (name.as_str(), addresses.as_slice()))
+    }
+
+    /// The reverse zones this zone is authoritative for.
+    ///
+    /// Only when the range lands on an octet boundary. Claiming a reverse
+    /// zone larger than the range would shadow reverse lookups for addresses
+    /// that are nothing to do with us, which is worse than not answering.
+    pub fn reverse_origin(base: Ipv4Addr, prefix_len: u8) -> Option<String> {
+        let octets = base.octets();
+        match prefix_len {
+            8 => Some(format!("{}.in-addr.arpa", octets[0])),
+            16 => Some(format!("{}.{}.in-addr.arpa", octets[1], octets[0])),
+            24 => Some(format!(
+                "{}.{}.{}.in-addr.arpa",
+                octets[2], octets[1], octets[0]
+            )),
+            _ => None,
+        }
+    }
+
+    /// Answers one question.
+    pub fn lookup(&self, qname: &str, query: Query) -> Answer {
+        if let Some(address) = reverse_address(qname) {
+            return match self.names.get(&address) {
+                Some(name) if query == Query::Ptr => {
+                    Answer::Name(format!("{name}.{}", self.origin.as_str()))
+                }
+                Some(_) => Answer::NoData,
+                None => Answer::NoSuchName,
+            };
+        }
+
+        let Some(relative) = self.origin.relative(qname) else {
+            return Answer::NotOurs;
+        };
+
+        // The apex: the zone itself exists whether or not anybody is in it.
+        if relative.is_empty() {
+            return match query {
+                Query::Soa => Answer::Soa,
+                Query::Ns => Answer::NoData,
+                _ => Answer::NoData,
+            };
+        }
+
+        match self.hosts.get(&relative) {
+            Some(addresses) if query == Query::A => Answer::Addresses(addresses.clone()),
+            // The name is here, it just has no AAAA and never will while
+            // only IPv4 is served. Saying "no such name" instead would tell
+            // a resolver to stop asking for the A record too.
+            Some(_) => Answer::NoData,
+            None => Answer::NoSuchName,
+        }
+    }
+}
+
+/// Every zone one agent serves.
+///
+/// One agent has one identity and as many networks as it likes, and each of
+/// them is a zone of its own named after the network. They are served
+/// together because they arrive at the same socket: a question is not
+/// labelled with the network it belongs to, only with a name, and the
+/// suffix is what decides. Nothing is shared between them — a member of one
+/// network is not a name in another.
+#[derive(Debug, Clone, Default)]
+pub struct Zones {
+    zones: Vec<Zone>,
+}
+
+impl Zones {
+    /// Collects zones, dropping any that repeats an origin already taken.
+    ///
+    /// Two networks can be given the same name, and then only one of them
+    /// can own the suffix. First wins, deterministically, rather than one
+    /// shadowing the other depending on the order they happened to start.
+    pub fn new(zones: impl IntoIterator<Item = Zone>) -> Self {
+        let mut taken: Vec<Zone> = Vec::new();
+        for zone in zones {
+            if taken.iter().any(|other| other.origin() == zone.origin()) {
+                continue;
+            }
+            taken.push(zone);
+        }
+        Self { zones: taken }
+    }
+
+    /// The zones, in the order they will be consulted.
+    pub fn iter(&self) -> impl Iterator<Item = &Zone> {
+        self.zones.iter()
+    }
+
+    /// How many zones are served.
+    pub fn len(&self) -> usize {
+        self.zones.len()
+    }
+
+    /// Whether nothing is served.
+    pub fn is_empty(&self) -> bool {
+        self.zones.is_empty()
+    }
+
+    /// How many names are answered for, across every zone.
+    pub fn names(&self) -> usize {
+        self.zones.iter().map(Zone::len).sum()
+    }
+
+    /// Answers one question, and says which zone answered it.
+    ///
+    /// The answer has to travel with its zone because the SOA that bounds a
+    /// denial belongs to the zone that denied it, and serving several means
+    /// that is no longer a foregone conclusion.
+    ///
+    /// A reverse question carries no suffix to match on, so every zone is
+    /// asked and the one that holds the address answers; if none does, the
+    /// denial comes from whichever is authoritative for that reverse zone,
+    /// and from nowhere at all when none is.
+    pub fn lookup(&self, qname: &str, query: Query) -> Option<(&Zone, Answer)> {
+        let mut denial: Option<(&Zone, Answer)> = None;
+        for zone in &self.zones {
+            match zone.lookup(qname, query) {
+                // Not this zone's question. Another may still own it.
+                Answer::NotOurs => continue,
+                // A zone that owns the suffix but not the name can still be
+                // overruled by one that has an answer: only reverse
+                // questions reach more than one zone, and there the zone
+                // holding the address is the one that knows.
+                answer @ (Answer::NoSuchName | Answer::NoData) => {
+                    denial.get_or_insert((zone, answer));
+                }
+                answer => return Some((zone, answer)),
+            }
+        }
+        denial
+    }
+}
+
+/// The address a reverse name asks about, if it is one.
+fn reverse_address(qname: &str) -> Option<Ipv4Addr> {
+    let name = qname.trim_end_matches('.').to_ascii_lowercase();
+    let rest = name.strip_suffix(".in-addr.arpa")?;
+    let mut octets = [0u8; 4];
+    let mut seen = 0;
+    for (index, part) in rest.split('.').enumerate() {
+        if index >= 4 {
+            return None;
+        }
+        octets[3 - index] = part.parse().ok()?;
+        seen += 1;
+    }
+    (seen == 4).then(|| Ipv4Addr::from(octets))
+}
+
+/// A serial that changes with the contents and not otherwise.
+///
+/// Derived rather than counted, so two agents holding the same roster agree,
+/// and a restart does not go backwards.
+fn content_serial(hosts: &BTreeMap<String, Vec<Ipv4Addr>>) -> u32 {
+    use std::hash::{Hash, Hasher};
+    let mut hasher = std::collections::hash_map::DefaultHasher::new();
+    for (name, addresses) in hosts {
+        name.hash(&mut hasher);
+        for address in addresses {
+            address.octets().hash(&mut hasher);
+        }
+    }
+    // Never zero: a zero serial is legal but reads like "unset" in a log.
+    (hasher.finish() as u32).max(1)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn zone() -> Zone {
+        Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [
+                ("music".to_string(), Ipv4Addr::new(10, 13, 37, 237)),
+                ("ai".to_string(), Ipv4Addr::new(10, 13, 37, 69)),
+            ],
+        )
+    }
+
+    #[test]
+    fn a_zone_name_is_normalised_and_checked() {
+        assert_eq!(ZoneName::new("LAB.").unwrap().as_str(), "lab");
+        assert_eq!(
+            ZoneName::new(" lab.internal ").unwrap().as_str(),
+            "lab.internal"
+        );
+        assert_eq!(
+            ZoneName::new("lab.internal").unwrap().top_label(),
+            "internal"
+        );
+
+        assert_eq!(ZoneName::new("").unwrap_err(), ZoneError::Empty);
+        assert_eq!(ZoneName::new(".").unwrap_err(), ZoneError::Empty);
+        assert!(matches!(
+            ZoneName::new("a..b").unwrap_err(),
+            ZoneError::Label { .. }
+        ));
+        assert!(matches!(
+            ZoneName::new("-lab").unwrap_err(),
+            ZoneError::Label { .. }
+        ));
+        assert!(matches!(
+            ZoneName::new("la b").unwrap_err(),
+            ZoneError::Label { .. }
+        ));
+        assert!(matches!(
+            ZoneName::new(&"x".repeat(64)).unwrap_err(),
+            ZoneError::Label { .. }
+        ));
+    }
+
+    #[test]
+    fn a_zone_name_that_shadows_a_public_one_is_flagged_but_allowed() {
+        // Flagged, never refused: shadowing is the user's decision to make.
+        let ru = ZoneName::new("ru").unwrap();
+        let warning = ru.collision().expect("a real TLD is worth mentioning");
+        assert!(warning.contains("real top-level domain"), "{warning}");
+
+        assert!(ZoneName::new("com").unwrap().collision().is_some());
+        assert!(ZoneName::new("lab.com").unwrap().collision().is_some());
+
+        // Reserved for private use, so nothing to say.
+        for quiet in ["internal", "lab.internal", "home", "test", "invalid"] {
+            assert_eq!(ZoneName::new(quiet).unwrap().collision(), None, "{quiet}");
+        }
+
+        // `.local` is not a delegated TLD, but it is not free either.
+        let local = ZoneName::new("local").unwrap().collision().unwrap();
+        assert!(local.contains("multicast DNS"), "{local}");
+
+        // A name that collides with nothing is said nothing about. It
+        // might be delegated one day, and warning about that fires on
+        // every private name anybody picks — a warning about nothing,
+        // which is how people learn to ignore warnings.
+        for quiet in ["lab", "a.lab", "pidar", "twar", "kitchen-table"] {
+            assert_eq!(ZoneName::new(quiet).unwrap().collision(), None, "{quiet}");
+        }
+    }
+
+    #[test]
+    fn a_neighbouring_name_is_not_inside_the_zone() {
+        // `evillab` ends with `lab`, and a suffix comparison that forgot the
+        // label boundary would hand it to us.
+        let origin = ZoneName::new("lab").unwrap();
+        assert!(origin.covers("lab"));
+        assert!(origin.covers("music.lab."));
+        assert!(origin.covers("a.b.lab"));
+        assert!(!origin.covers("evillab"));
+        assert!(!origin.covers("lab.example.com"));
+        assert!(!origin.covers("example.com"));
+    }
+
+    #[test]
+    fn a_member_resolves_by_name() {
+        let zone = zone();
+        assert_eq!(
+            zone.lookup("music.lab", Query::A),
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 237)])
+        );
+        // Case and a trailing dot are the same question.
+        assert_eq!(
+            zone.lookup("MUSIC.LAB.", Query::A),
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 237)])
+        );
+    }
+
+    #[test]
+    fn a_name_with_no_record_of_that_type_is_not_a_missing_name() {
+        // AAAA for a member that exists must be NODATA, not NXDOMAIN: an
+        // NXDOMAIN would tell the resolver the name is absent and stop it
+        // asking for the A record.
+        let zone = zone();
+        assert_eq!(zone.lookup("music.lab", Query::Other), Answer::NoData);
+        assert_eq!(zone.lookup("nobody.lab", Query::A), Answer::NoSuchName);
+    }
+
+    #[test]
+    fn questions_outside_the_zone_are_refused_rather_than_denied() {
+        // Denying them would be a lie: this server knows nothing about them.
+        let zone = zone();
+        assert_eq!(zone.lookup("example.com", Query::A), Answer::NotOurs);
+        assert_eq!(zone.lookup("evillab", Query::A), Answer::NotOurs);
+    }
+
+    #[test]
+    fn the_apex_exists_even_with_nobody_in_the_network() {
+        let empty = Zone::new(ZoneName::new("lab").unwrap(), []);
+        assert!(empty.is_empty());
+        assert_eq!(empty.lookup("lab", Query::Soa), Answer::Soa);
+        assert_eq!(empty.lookup("lab", Query::A), Answer::NoData);
+        assert_eq!(empty.lookup("music.lab", Query::A), Answer::NoSuchName);
+    }
+
+    #[test]
+    fn an_address_resolves_back_to_its_name() {
+        let zone = zone();
+        assert_eq!(
+            zone.lookup("237.37.13.10.in-addr.arpa", Query::Ptr),
+            Answer::Name("music.lab".into())
+        );
+        assert_eq!(
+            zone.lookup("9.37.13.10.in-addr.arpa", Query::Ptr),
+            Answer::NoSuchName
+        );
+        // A reverse name asked for the wrong type is still a name we know.
+        assert_eq!(
+            zone.lookup("237.37.13.10.in-addr.arpa", Query::A),
+            Answer::NoData
+        );
+    }
+
+    #[test]
+    fn a_reverse_zone_is_claimed_only_when_it_is_exactly_ours() {
+        // Claiming a reverse zone wider than the range would shadow lookups
+        // for addresses that have nothing to do with this network.
+        assert_eq!(
+            Zone::reverse_origin(Ipv4Addr::new(10, 13, 37, 0), 24).as_deref(),
+            Some("37.13.10.in-addr.arpa")
+        );
+        assert_eq!(
+            Zone::reverse_origin(Ipv4Addr::new(10, 13, 0, 0), 16).as_deref(),
+            Some("13.10.in-addr.arpa")
+        );
+        assert_eq!(Zone::reverse_origin(Ipv4Addr::new(10, 13, 37, 0), 25), None);
+        assert_eq!(Zone::reverse_origin(Ipv4Addr::new(100, 64, 0, 0), 10), None);
+    }
+
+    #[test]
+    fn members_with_nothing_to_say_are_left_out() {
+        let zone = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [(String::new(), Ipv4Addr::new(10, 0, 0, 1))],
+        );
+        assert!(zone.is_empty());
+    }
+
+    #[test]
+    fn two_members_sharing_a_name_both_answer_and_the_order_is_stable() {
+        // The state layer resolves name ownership; if two records still
+        // reach here, answering with both beats picking one at random.
+        let build = |flip: bool| {
+            let members = if flip {
+                vec![
+                    ("music".to_string(), Ipv4Addr::new(10, 0, 0, 2)),
+                    ("music".to_string(), Ipv4Addr::new(10, 0, 0, 1)),
+                ]
+            } else {
+                vec![
+                    ("music".to_string(), Ipv4Addr::new(10, 0, 0, 1)),
+                    ("music".to_string(), Ipv4Addr::new(10, 0, 0, 2)),
+                ]
+            };
+            Zone::new(ZoneName::new("lab").unwrap(), members)
+        };
+        assert_eq!(
+            build(false).lookup("music.lab", Query::A),
+            build(true).lookup("music.lab", Query::A)
+        );
+        assert_eq!(build(false).serial(), build(true).serial());
+    }
+
+    #[test]
+    fn the_serial_follows_the_contents() {
+        let a = zone();
+        let b = zone();
+        assert_eq!(a.serial(), b.serial(), "the same roster, the same serial");
+
+        let changed = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [("music".to_string(), Ipv4Addr::new(10, 13, 37, 238))],
+        );
+        assert_ne!(a.serial(), changed.serial());
+    }
+
+    #[test]
+    fn each_network_answers_only_for_its_own_zone() {
+        // One agent, several networks, one socket. A question carries a
+        // name and not the network it belongs to, so the suffix is what
+        // decides — and a member of one network is not a name in another.
+        let zones = Zones::new([
+            Zone::new(
+                ZoneName::new("lab").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+            ),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        assert_eq!(zones.len(), 2);
+        assert_eq!(zones.names(), 2);
+
+        let (zone, answer) = zones.lookup("music.lab", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "lab");
+        assert_eq!(
+            answer,
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 2)])
+        );
+
+        let (zone, answer) = zones.lookup("music.home", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "home");
+        assert_eq!(answer, Answer::Addresses(vec![Ipv4Addr::new(10, 20, 0, 2)]));
+
+        // A name under neither is nobody's business here.
+        assert!(zones.lookup("music.example.com", Query::A).is_none());
+        assert!(Zones::default().lookup("music.lab", Query::A).is_none());
+    }
+
+    #[test]
+    fn a_denial_comes_from_the_zone_that_owns_the_name() {
+        // The SOA that bounds a negative answer belongs to the zone that
+        // denied it; with several served that is no longer a given.
+        let zones = Zones::new([
+            Zone::new(ZoneName::new("lab").unwrap(), []),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        let (zone, answer) = zones.lookup("nobody.lab", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "lab");
+        assert_eq!(answer, Answer::NoSuchName);
+    }
+
+    #[test]
+    fn a_reverse_question_is_answered_by_whichever_zone_holds_the_address() {
+        // It carries no suffix to match on, so every zone is asked.
+        let zones = Zones::new([
+            Zone::new(
+                ZoneName::new("lab").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+            ),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("kitchen".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        let (_, answer) = zones.lookup("2.0.20.10.in-addr.arpa", Query::Ptr).unwrap();
+        assert_eq!(answer, Answer::Name("kitchen.home".to_string()));
+        let (_, answer) = zones.lookup("2.37.13.10.in-addr.arpa", Query::Ptr).unwrap();
+        assert_eq!(answer, Answer::Name("music.lab".to_string()));
+    }
+
+    #[test]
+    fn two_networks_of_one_name_do_not_shadow_each_other_at_random() {
+        // Exactly the two-`LAB` case. Only one can own the suffix; which
+        // one must not depend on the order they happened to start in.
+        let first = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+        );
+        let second = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [("music".to_string(), Ipv4Addr::new(10, 99, 0, 2))],
+        );
+        let zones = Zones::new([first, second]);
+        assert_eq!(zones.len(), 1, "one suffix, one zone");
+        let (_, answer) = zones.lookup("music.lab", Query::A).unwrap();
+        assert_eq!(
+            answer,
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 2)])
+        );
+    }
+}

crates/tsunagi-core/src/error.rs

@@ -15,6 +15,12 @@ pub type Result<T, E = Error> = std::result::Result<T, E>;
 #[derive(Debug, thiserror::Error)]
 #[non_exhaustive]
 pub enum Error {
+    /// A node-history cleanup is unsafe, stale or cannot be performed.
+    #[error("cannot forget node: {0}")]
+    ForgetNode(String),
+    /// An ordered local data-plane protocol selection is invalid.
+    #[error("invalid protocol selection: {0}")]
+    InvalidProtocols(String),
     /// The supplied network name does not satisfy the documented rules.
     #[error("invalid network name: {0}")]
     InvalidNetworkName(&'static str),
@@ -97,6 +103,19 @@ pub enum Error {
     #[error("network {0} is not configured")]
     NetworkUnknown(NetworkId),
 
+    /// An overlay range overlaps networks already active on this agent.
+    #[error(
+        "network {network} cannot start: {range} overlaps active networks {conflicts:?}; switch to it to suspend those networks"
+    )]
+    NetworkRangeConflict {
+        /// Network that could not start.
+        network: NetworkId,
+        /// Its agreed or proposed range.
+        range: crate::state::Ipv4Range,
+        /// Active networks that must be suspended first.
+        conflicts: Vec<NetworkId>,
+    },
+
     /// No session with that peer exists in the given network.
     #[error("no authenticated session with peer {peer} in network {network}")]
     NoSuchPeer {
@@ -110,6 +129,23 @@ pub enum Error {
     #[error("agent is stopped")]
     Stopped,
 
+    /// Something that had to answer in bounded time did not.
+    ///
+    /// Reported instead of waited out: an agent that has wedged must not
+    /// leave a command hanging with nothing on screen and no way out.
+    #[error("{what} did not answer in time")]
+    Timeout {
+        /// What was asked, phrased so the message stands on its own.
+        what: String,
+    },
+
+    /// The request is understood and cannot be done as things stand.
+    #[error("{reason}")]
+    Rejected {
+        /// Why, phrased so the message stands on its own.
+        reason: String,
+    },
+
     /// Discovery backend failure. Never fatal for the agent.
     #[error("discovery error: {0}")]
     Discovery(String),

crates/tsunagi-core/src/identity/mod.rs

@@ -49,15 +49,28 @@ impl DeviceIdentity {
         self.secret.public()
     }
 
-    /// The raw secret key bytes, for persistence only.
-    pub(crate) fn secret_bytes(&self) -> [u8; 32] {
+    /// The raw secret key bytes, for persistence and transport certificate generation.
+    pub fn secret_bytes(&self) -> [u8; 32] {
         self.secret.to_bytes()
     }
 
+    /// Signs a message with the device's persistent secret key.
+    pub fn sign(&self, message: &[u8]) -> iroh::Signature {
+        self.secret.sign(message)
+    }
+
     /// A clone of the iroh secret key, for endpoint construction only.
     pub(crate) fn secret_key(&self) -> SecretKey {
         self.secret.clone()
     }
+
+    /// A clone of the key used to sign this device's own state records.
+    ///
+    /// The same persistent identity the control plane authenticates, so a
+    /// record signed today is still attributable after any absence.
+    pub(crate) fn signing_key(&self) -> SecretKey {
+        self.secret.clone()
+    }
 }
 
 impl std::fmt::Debug for DeviceIdentity {

crates/tsunagi-core/src/identity/network.rs

@@ -340,6 +340,8 @@ pub struct NetworkKeys {
     network_id: NetworkId,
     discovery_key: DiscoveryKey,
     auth_key: Zeroizing<[u8; 32]>,
+    // Separate from the semi-public DiscoveryKey. Existing derivations stay frozen.
+    dht_write_key: Zeroizing<[u8; 32]>,
     name: NetworkName,
 }
 
@@ -373,6 +375,7 @@ impl NetworkKeys {
             network_id: NetworkId(expand("network-id")),
             discovery_key: DiscoveryKey(expand("discovery-key")),
             auth_key: Zeroizing::new(expand("handshake-auth")),
+            dht_write_key: Zeroizing::new(expand("mainline-rendezvous-write-v1")),
             name: name.clone(),
         }
     }
@@ -405,6 +408,11 @@ impl NetworkKeys {
     pub(crate) fn auth_key(&self) -> &[u8; 32] {
         &self.auth_key
     }
+
+    /// Secret signing material for network rendezvous; never exported or logged.
+    pub(crate) fn dht_write_key(&self) -> &[u8; 32] {
+        &self.dht_write_key
+    }
 }
 
 impl std::fmt::Debug for NetworkKeys {

crates/tsunagi-core/src/ipc/unix.rs

@@ -0,0 +1,227 @@
+//! A Unix socket adapter for the local control interface.
+//!
+//! One of the per-platform adapters; see [`super`]. It provides only what is
+//! particular to a Unix socket — binding a listener, connecting a client, and
+//! the owner-only permissions — and hands every accepted connection to the
+//! shared [`serve_connection`](super::serve_connection). The request framing,
+//! the dispatch and the client wrappers are the same on every platform and
+//! live in [`super`], re-exported here so `ipc::unix::request_status` and its
+//! siblings keep resolving.
+//!
+//! It is reachable only by a process that can open a file inside the agent's
+//! owner-only state directory, and the socket itself is created mode `0600`.
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use tokio::net::{UnixListener, UnixStream};
+use tokio::task::JoinHandle;
+
+use crate::error::{Error, Result};
+
+// The transport-agnostic surface, re-exported so this module is a complete
+// view of the local control interface on its own.
+pub use super::{
+    CONTROL_PROTOCOL, ControlSocketAccess, EXCHANGE_TIMEOUT, ReportSource, is_serving,
+    join_network, leave_network, request_status, set_active, set_dns, set_hostname,
+};
+
+/// Serves the local control interface on a Unix socket.
+#[derive(Debug)]
+pub struct ControlSocket {
+    path: PathBuf,
+    task: Option<JoinHandle<()>>,
+}
+
+impl ControlSocket {
+    /// Binds the socket and starts serving.
+    ///
+    /// A socket file left behind by a crashed agent is replaced, but only
+    /// after checking that nothing is listening on it, so two live agents
+    /// never fight over one path.
+    pub async fn bind(
+        path: impl AsRef<Path>,
+        source: Arc<dyn ReportSource>,
+        access: ControlSocketAccess,
+    ) -> Result<Self> {
+        let path = path.as_ref().to_path_buf();
+
+        if let Some(parent) = path.parent() {
+            create_parent(parent, &access)?;
+        }
+
+        if path.exists() {
+            if is_serving(&path).await {
+                return Err(Error::StateLocked { path: path.clone() });
+            }
+            // Nothing is listening, so the file is a leftover.
+            std::fs::remove_file(&path).map_err(|source| Error::Io {
+                path: path.clone(),
+                source,
+            })?;
+        }
+
+        let listener = UnixListener::bind(&path).map_err(|source| Error::Io {
+            path: path.clone(),
+            source,
+        })?;
+        apply_access(&path, &access)?;
+
+        let task = tokio::spawn(serve(listener, source));
+        Ok(Self {
+            path,
+            task: Some(task),
+        })
+    }
+
+    /// The path being served.
+    pub fn path(&self) -> &Path {
+        &self.path
+    }
+
+    /// Stops serving and removes the socket file.
+    pub async fn shutdown(mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+            let _ = task.await;
+        }
+        let _ = std::fs::remove_file(&self.path);
+    }
+}
+
+impl Drop for ControlSocket {
+    fn drop(&mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+        }
+        let _ = std::fs::remove_file(&self.path);
+    }
+}
+
+/// Creates the socket's parent directory.
+///
+/// A group-shared socket must sit in a directory the group can traverse, so it
+/// is `0755`; a private socket keeps the owner-only state-directory default.
+fn create_parent(parent: &Path, access: &ControlSocketAccess) -> Result<()> {
+    match access {
+        ControlSocketAccess::Private => crate::storage::create_dir(parent),
+        ControlSocketAccess::Group(_) => {
+            std::fs::create_dir_all(parent).map_err(|source| Error::Io {
+                path: parent.to_path_buf(),
+                source,
+            })?;
+            set_mode(parent, 0o755)
+        }
+    }
+}
+
+/// Applies the access policy to the bound socket.
+///
+/// Private is owner-only `0600`. Group sets the socket's group to the gid and
+/// the mode to `0660`; if the group cannot be set — the agent is not a member,
+/// say — it falls back to owner-only rather than leaving the group wrong, so
+/// the socket is never more open than intended.
+fn apply_access(path: &Path, access: &ControlSocketAccess) -> Result<()> {
+    match access {
+        ControlSocketAccess::Private => set_mode(path, 0o600),
+        ControlSocketAccess::Group(gid) => {
+            if let Err(source) = std::os::unix::fs::chown(path, None, Some(*gid)) {
+                tracing::warn!(
+                    path = %path.display(), gid, %source,
+                    "cannot set the control socket group; keeping it owner-only"
+                );
+                return set_mode(path, 0o600);
+            }
+            set_mode(path, 0o660)
+        }
+    }
+}
+
+fn set_mode(path: &Path, mode: u32) -> Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    std::fs::set_permissions(path, std::fs::Permissions::from_mode(mode)).map_err(|source| {
+        Error::Io {
+            path: path.to_path_buf(),
+            source,
+        }
+    })
+}
+
+async fn serve(listener: UnixListener, source: Arc<dyn ReportSource>) {
+    loop {
+        let Ok((stream, _)) = listener.accept().await else {
+            continue;
+        };
+        let source = Arc::clone(&source);
+        tokio::spawn(async move {
+            if let Err(err) = super::serve_connection(stream, source).await {
+                tracing::debug!(%err, "local control request failed");
+            }
+        });
+    }
+}
+
+/// Connects a client to the socket at `path`.
+pub(crate) async fn connect(path: &Path) -> Result<UnixStream> {
+    UnixStream::connect(path).await.map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })
+}
+
+/// Whether an agent is listening on the socket at `path`.
+pub(crate) async fn probe(path: &Path) -> bool {
+    UnixStream::connect(path).await.is_ok()
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::time::Duration;
+
+    use super::super::{Request, Response, StatusReport, exchange};
+    use super::*;
+    use crate::BoxFuture;
+
+    #[tokio::test]
+    async fn a_silent_agent_is_reported_rather_than_waited_out() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("control.sock");
+        // A socket that accepts and then says nothing, like a wedged agent:
+        // the connection succeeds, the answer never comes. Unbounded, this
+        // call would never return.
+        let _listener = UnixListener::bind(&path).unwrap();
+
+        let error = exchange(&path, &Request::Status, Duration::from_millis(100))
+            .await
+            .expect_err("a silent agent cannot be reported as healthy");
+
+        match error {
+            Error::Timeout { what } => assert!(
+                what.contains("control.sock"),
+                "the message names what did not answer: {what}"
+            ),
+            other => panic!("expected a timeout, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn an_answered_request_is_not_affected_by_the_bound() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("control.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+        let control = ControlSocket::bind(&path, source, ControlSocketAccess::Private)
+            .await
+            .unwrap();
+
+        let answer = exchange(&path, &Request::Status, EXCHANGE_TIMEOUT)
+            .await
+            .unwrap();
+        assert!(matches!(answer, Response::Status(_)));
+
+        control.shutdown().await;
+    }
+}

crates/tsunagi-core/src/ipc/windows.rs

@@ -0,0 +1,453 @@
+//! A named-pipe adapter for the local control interface.
+//!
+//! The Windows counterpart of [`super::unix`]. It provides only what is
+//! particular to a named pipe — creating the server instance, opening a client
+//! and turning the state directory path into a pipe name — and hands every
+//! accepted connection to the shared [`serve_connection`](super::serve_connection).
+//! The framing, the dispatch and the client wrappers are the same on every
+//! platform and live in [`super`], re-exported here so `ipc::windows::request_status`
+//! and its siblings resolve just as their Unix equivalents do.
+//!
+//! # From a path to a pipe
+//!
+//! The rest of the agent addresses the control interface by a filesystem path,
+//! the same one a Unix socket would live at. A named pipe has no filesystem
+//! path, so the path is hashed into a stable name under `\\.\pipe\`. The agent
+//! and a client derive it the same way from the same path, so neither has to
+//! be told where the other put it.
+//!
+//! # Who may connect
+//!
+//! The pipe rejects clients from other machines (`reject_remote_clients(true)`).
+//! An explicit SDDL security descriptor grants Full Control to Administrators
+//! and LocalSystem, while granting Read and Write (`GRGW`) to all local users
+//! with a Medium Integrity SACL label (`S:(ML;;NW;;;ME)`). This allows unprivileged
+//! user-space tools (such as tray applications, status monitors, or CLI commands
+//! launched by regular users) to connect to the agent even when the agent was
+//! started by an elevated Administrator or as a system service.
+//! A permission denial is reported as such, never as an absent agent.
+
+#![allow(unsafe_code)]
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::Duration;
+
+use sha2::{Digest, Sha256};
+use tokio::net::windows::named_pipe::{ClientOptions, NamedPipeClient, ServerOptions};
+use tokio::task::JoinHandle;
+
+use crate::error::{Error, Result};
+
+pub use super::{
+    CONTROL_PROTOCOL, ControlSocketAccess, EXCHANGE_TIMEOUT, ReportSource, is_serving,
+    join_network, leave_network, request_status, set_active, set_dns, set_hostname,
+};
+
+/// `ERROR_ACCESS_DENIED`: creating a first instance finds an existing pipe,
+/// or a client cannot access an existing pipe with its current privileges.
+const ERROR_ACCESS_DENIED: i32 = 5;
+/// `ERROR_PIPE_BUSY`: every instance is serving a client right now. A server
+/// is there; the client only has to wait for a free instance.
+const ERROR_PIPE_BUSY: i32 = 231;
+
+/// RAII wrapper for Windows named pipe security attributes with permissive local ACL.
+///
+/// SDDL specification:
+/// - `D:(A;;GRGW;;;WD)`: DACL allows Generic Read and Generic Write to Everyone (`WD`).
+///   Full control (`GA`) remains reserved to Administrators and LocalSystem.
+/// - `(A;;GA;;;BA)`: DACL allows Generic All to Builtin Administrators (`BA`).
+/// - `(A;;GA;;;SY)`: DACL allows Generic All to LocalSystem (`SY`).
+/// - `S:(ML;;NW;;;ME)`: SACL Mandatory Label with Medium Integrity (`ME`) and No Write Up (`NW`).
+///   This allows unprivileged user-space processes (Medium Integrity) to write to the
+///   named pipe even when the agent runs as an elevated Administrator (High Integrity).
+struct PipeSecurityAttributes {
+    attrs: windows_sys::Win32::Security::SECURITY_ATTRIBUTES,
+}
+
+impl std::fmt::Debug for PipeSecurityAttributes {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PipeSecurityAttributes")
+            .field("nLength", &self.attrs.nLength)
+            .field("lpSecurityDescriptor", &self.attrs.lpSecurityDescriptor)
+            .field("bInheritHandle", &self.attrs.bInheritHandle)
+            .finish()
+    }
+}
+
+impl PipeSecurityAttributes {
+    fn new() -> Result<Self> {
+        use std::ffi::OsStr;
+        use std::os::windows::ffi::OsStrExt;
+        use windows_sys::Win32::Security::Authorization::{
+            ConvertStringSecurityDescriptorToSecurityDescriptorW, SDDL_REVISION_1,
+        };
+
+        const SDDL: &str = "D:(A;;GRGW;;;WD)(A;;GA;;;BA)(A;;GA;;;SY)S:(ML;;NW;;;ME)";
+        let wide: Vec<u16> = OsStr::new(SDDL)
+            .encode_wide()
+            .chain(std::iter::once(0))
+            .collect();
+
+        let mut p_sd = std::ptr::null_mut();
+        // SAFETY: We pass a valid null-terminated wide string, standard revision 1,
+        // and a valid pointer to receive the descriptor.
+        let success = unsafe {
+            ConvertStringSecurityDescriptorToSecurityDescriptorW(
+                wide.as_ptr(),
+                SDDL_REVISION_1,
+                &mut p_sd,
+                std::ptr::null_mut(),
+            )
+        };
+
+        if success == 0 || p_sd.is_null() {
+            let err = std::io::Error::last_os_error();
+            return Err(Error::Storage(format!(
+                "cannot create pipe security descriptor: {err}"
+            )));
+        }
+
+        let attrs = windows_sys::Win32::Security::SECURITY_ATTRIBUTES {
+            nLength: std::mem::size_of::<windows_sys::Win32::Security::SECURITY_ATTRIBUTES>()
+                as u32,
+            lpSecurityDescriptor: p_sd,
+            bInheritHandle: 0,
+        };
+
+        Ok(Self { attrs })
+    }
+
+    fn as_raw_mut(&mut self) -> *mut std::ffi::c_void {
+        &mut self.attrs as *mut _ as *mut std::ffi::c_void
+    }
+}
+
+impl Drop for PipeSecurityAttributes {
+    fn drop(&mut self) {
+        if !self.attrs.lpSecurityDescriptor.is_null() {
+            // SAFETY: The descriptor was allocated by ConvertStringSecurityDescriptorToSecurityDescriptorW
+            // and must be freed with LocalFree.
+            unsafe {
+                windows_sys::Win32::Foundation::LocalFree(self.attrs.lpSecurityDescriptor);
+            }
+            self.attrs.lpSecurityDescriptor = std::ptr::null_mut();
+        }
+    }
+}
+
+/// Serves the local control interface on a named pipe.
+#[derive(Debug)]
+pub struct ControlSocket {
+    path: PathBuf,
+    task: Option<JoinHandle<()>>,
+}
+
+impl ControlSocket {
+    /// Binds the pipe and starts serving.
+    ///
+    /// Claiming the first instance of the name is what detects a second agent:
+    /// if the name already exists, the create is refused and that is reported
+    /// as the state being locked, so two live agents never share one pipe.
+    ///
+    /// `_access` is accepted for parity with the Unix socket; granting a group
+    /// here means a pipe security descriptor and is a separate follow-up, so for
+    /// now the pipe keeps its default (creator plus the built-in access the
+    /// `PipeSecurityAttributes` set).
+    pub async fn bind(
+        path: impl AsRef<Path>,
+        source: Arc<dyn ReportSource>,
+        _access: ControlSocketAccess,
+    ) -> Result<Self> {
+        let path = path.as_ref().to_path_buf();
+        let name = pipe_name(&path);
+
+        let mut sec_attrs = PipeSecurityAttributes::new()?;
+        // SAFETY: sec_attrs points to valid initialized SECURITY_ATTRIBUTES.
+        let server = match unsafe {
+            ServerOptions::new()
+                .first_pipe_instance(true)
+                .reject_remote_clients(true)
+                .create_with_security_attributes_raw(&name, sec_attrs.as_raw_mut())
+        } {
+            Ok(server) => server,
+            Err(err) if err.raw_os_error() == Some(ERROR_ACCESS_DENIED) => {
+                return Err(Error::StateLocked { path });
+            }
+            Err(source) => return Err(Error::Io { path, source }),
+        };
+
+        let task = tokio::spawn(serve(name, server, source));
+        Ok(Self {
+            path,
+            task: Some(task),
+        })
+    }
+
+    /// The path the pipe name was derived from.
+    pub fn path(&self) -> &Path {
+        &self.path
+    }
+
+    /// Stops serving. The pipe goes with the server, so there is nothing to
+    /// remove.
+    pub async fn shutdown(mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+            let _ = task.await;
+        }
+    }
+}
+
+impl Drop for ControlSocket {
+    fn drop(&mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+        }
+    }
+}
+
+/// Accepts one client at a time, keeping a fresh instance ready for the next.
+///
+/// A named pipe server instance serves a single client, so a new instance is
+/// created as soon as one is taken — otherwise a second `tsng status` while
+/// the first is mid-flight would find nothing listening.
+async fn serve(
+    name: String,
+    first: tokio::net::windows::named_pipe::NamedPipeServer,
+    source: Arc<dyn ReportSource>,
+) {
+    let mut server = first;
+    loop {
+        if server.connect().await.is_err() {
+            match next_instance(&name) {
+                Some(next) => {
+                    server = next;
+                    continue;
+                }
+                None => return,
+            }
+        }
+        let connected = server;
+        match next_instance(&name) {
+            Some(next) => server = next,
+            None => {
+                // Nothing left to accept the next client on, but the one in
+                // hand is still answered before the loop ends.
+                let source = Arc::clone(&source);
+                tokio::spawn(async move {
+                    let _ = super::serve_connection(connected, source).await;
+                });
+                return;
+            }
+        }
+        let source = Arc::clone(&source);
+        tokio::spawn(async move {
+            if let Err(err) = super::serve_connection(connected, source).await {
+                tracing::debug!(%err, "local control request failed");
+            }
+        });
+    }
+}
+
+/// Creates the next pipe instance, or `None` if the name can no longer be
+/// served.
+fn next_instance(name: &str) -> Option<tokio::net::windows::named_pipe::NamedPipeServer> {
+    let mut sec_attrs = match PipeSecurityAttributes::new() {
+        Ok(attrs) => attrs,
+        Err(err) => {
+            tracing::debug!(%err, "cannot build pipe security attributes for next instance");
+            return None;
+        }
+    };
+    // SAFETY: sec_attrs points to valid initialized SECURITY_ATTRIBUTES.
+    match unsafe {
+        ServerOptions::new()
+            .reject_remote_clients(true)
+            .create_with_security_attributes_raw(name, sec_attrs.as_raw_mut())
+    } {
+        Ok(server) => Some(server),
+        Err(err) => {
+            tracing::debug!(%err, "cannot create the next control pipe instance");
+            None
+        }
+    }
+}
+
+/// Connects a client to the pipe for `path`, waiting out a busy pipe.
+///
+/// The wait is bounded by the caller's exchange timeout, so a pipe that is
+/// busy forever is given up on rather than spun on.
+pub(crate) async fn connect(path: &Path) -> Result<NamedPipeClient> {
+    let name = pipe_name(path);
+    loop {
+        match ClientOptions::new().open(&name) {
+            Ok(client) => return Ok(client),
+            Err(err) if err.raw_os_error() == Some(ERROR_PIPE_BUSY) => {
+                tokio::time::sleep(Duration::from_millis(50)).await;
+            }
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => {
+                return Err(Error::Io {
+                    path: path.to_path_buf(),
+                    source: std::io::Error::new(
+                        std::io::ErrorKind::PermissionDenied,
+                        format!(
+                            "the agent's control pipe exists, but Windows denied access. \
+                             Run this command as the same Windows user and with the same \
+                             elevation as `tsng up` (use an administrator terminal if \
+                             the agent is elevated). Windows error: {err}"
+                        ),
+                    ),
+                });
+            }
+            Err(source) => {
+                return Err(Error::Io {
+                    path: path.to_path_buf(),
+                    source,
+                });
+            }
+        }
+    }
+}
+
+/// Whether an agent is serving the pipe for `path`.
+///
+/// A busy or inaccessible pipe still means a server is there. In particular,
+/// access denied must not send a mutating command down the offline path,
+/// where it would only produce a misleading state-directory lock error.
+pub(crate) async fn probe(path: &Path) -> bool {
+    let name = pipe_name(path);
+    match ClientOptions::new().open(&name) {
+        Ok(_) => true,
+        Err(err) => {
+            err.raw_os_error() == Some(ERROR_PIPE_BUSY)
+                || err.kind() == std::io::ErrorKind::PermissionDenied
+        }
+    }
+}
+
+/// The pipe name for a control-socket path.
+///
+/// A hash rather than the path itself, because a pipe name may not contain the
+/// separators and drive letters a path does, and because two agents with
+/// different state directories must never collide.
+fn pipe_name(path: &Path) -> String {
+    let digest = Sha256::digest(path.as_os_str().as_encoded_bytes());
+    format!(r"\\.\pipe\tsunagi-{}", hex::encode(&digest[..16]))
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::time::Duration;
+
+    use super::super::{Request, Response, StatusReport, exchange};
+    use super::*;
+    use crate::BoxFuture;
+
+    #[test]
+    fn one_path_maps_to_one_name_and_two_paths_do_not_collide() {
+        let a = pipe_name(Path::new(r"C:\a\agent.sock"));
+        let b = pipe_name(Path::new(r"C:\b\agent.sock"));
+        assert!(a.starts_with(r"\\.\pipe\tsunagi-"), "{a}");
+        assert_eq!(a, pipe_name(Path::new(r"C:\a\agent.sock")));
+        assert_ne!(a, b);
+    }
+
+    #[tokio::test]
+    async fn a_request_makes_the_round_trip_over_a_real_pipe() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+
+        assert!(!is_serving(&path).await, "nothing serves the pipe yet");
+        let control = ControlSocket::bind(&path, source).await.unwrap();
+        assert!(is_serving(&path).await, "the agent serves it now");
+
+        let answer = exchange(&path, &Request::Status, EXCHANGE_TIMEOUT)
+            .await
+            .unwrap();
+        assert!(matches!(answer, Response::Status(_)));
+
+        control.shutdown().await;
+    }
+
+    #[tokio::test]
+    async fn an_inaccessible_agent_is_not_mistaken_for_an_absent_one() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        // An outbound-only pipe denies the client's duplex open even when
+        // both processes have the same privileges. This exercises the same
+        // Windows error as a client unable to access an elevated agent,
+        // without requiring elevation or changing any ACLs.
+        let _server = ServerOptions::new()
+            .first_pipe_instance(true)
+            .access_inbound(false)
+            .create(pipe_name(&path))
+            .unwrap();
+
+        assert!(is_serving(&path).await, "access denied is not absence");
+        for error in [
+            request_status(&path).await.unwrap_err(),
+            join_network(&path, "test", "local-ipc-test-secret")
+                .await
+                .unwrap_err(),
+            set_dns(&path, false, None).await.unwrap_err(),
+        ] {
+            let Error::Io { source, .. } = error else {
+                panic!("expected an access error, got {error}");
+            };
+            assert_eq!(source.kind(), std::io::ErrorKind::PermissionDenied);
+            let message = source.to_string();
+            assert!(message.contains("same Windows user"), "{message}");
+            assert!(message.contains("administrator"), "{message}");
+            assert!(!message.contains("local-ipc-test-secret"), "{message}");
+        }
+    }
+
+    #[tokio::test]
+    async fn a_silent_agent_is_reported_rather_than_waited_out() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        // A server instance that accepts and then never answers, like a
+        // wedged agent. Unbounded, the exchange below would never return.
+        let name = pipe_name(&path);
+        let server = ServerOptions::new()
+            .first_pipe_instance(true)
+            .create(&name)
+            .unwrap();
+        let _accept = tokio::spawn(async move {
+            let _ = server.connect().await;
+            // Hold the connection open, answering nothing.
+            tokio::time::sleep(Duration::from_secs(30)).await;
+        });
+
+        let error = exchange(&path, &Request::Status, Duration::from_millis(200))
+            .await
+            .expect_err("a silent agent cannot be reported as healthy");
+        assert!(matches!(error, Error::Timeout { .. }), "{error:?}");
+    }
+
+    #[tokio::test]
+    async fn a_second_agent_on_the_same_pipe_is_refused() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+
+        let first = ControlSocket::bind(&path, Arc::clone(&source))
+            .await
+            .unwrap();
+        let second = ControlSocket::bind(&path, source).await;
+        assert!(
+            matches!(second, Err(Error::StateLocked { .. })),
+            "{second:?}"
+        );
+
+        first.shutdown().await;
+    }
+}

crates/tsunagi-core/src/lib.rs

@@ -14,7 +14,12 @@
 //! * [`proto`] — the control protocol: framing, messages, handshake.
 //! * [`net`] — the iroh connectivity adapter and its observability surface.
 //! * [`agent`] — the runtime: agent lifecycle, per-network runtimes, reconnect.
-//! * [`dataplane`] — the minimal contract future IP plugins must satisfy.
+//! * [`dataplane`] — the contract IP plugins satisfy, the packet transport,
+//!   and the WireGuard data plane.
+//! * [`state`] — signed records that outlive a session, and the rules for
+//!   merging them between replicas.
+//! * [`ipc`] — the local control interface a command line tool talks to. An
+//!   adapter over the public API; the core does not know it exists.
 //!
 //! # What this library deliberately does not do
 //!
@@ -34,11 +39,19 @@ pub mod agent;
 pub mod config;
 pub mod dataplane;
 pub mod discovery;
+pub mod dns;
 pub mod error;
 pub mod identity;
+pub mod ipc;
 pub mod net;
+pub mod overlay;
 pub mod proto;
+pub mod state;
 pub mod storage;
+pub mod task;
+#[cfg(feature = "testing")]
+pub mod testing;
+pub mod vip;
 
 pub use agent::{Agent, AgentStatus, Event, NetworkStatus, PeerStatus};
 pub use config::{AgentConfig, Limits, ReconnectPolicy, StoragePaths, TransportPolicy};

crates/tsunagi-core/src/net.rs

@@ -45,6 +45,16 @@ pub enum PathAddr {
     Other(String),
 }
 
+impl std::fmt::Display for PathAddr {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            PathAddr::Ip(addr) => write!(f, "{addr}"),
+            PathAddr::Relay(url) => write!(f, "relay {url}"),
+            PathAddr::Other(what) => write!(f, "{what}"),
+        }
+    }
+}
+
 /// One verified network path of a connection.
 #[derive(Debug, Clone)]
 pub struct PathInfo {
@@ -81,6 +91,17 @@ pub enum TransportKind {
     Unknown,
 }
 
+impl std::fmt::Display for TransportKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let word = match self {
+            TransportKind::Direct => "direct",
+            TransportKind::Relay => "relay",
+            TransportKind::Unknown => "unknown",
+        };
+        f.write_str(word)
+    }
+}
+
 /// Counters for one connection.
 #[derive(Debug, Clone, Default, PartialEq, Eq)]
 pub struct ConnectionCounters {
@@ -219,6 +240,62 @@ impl EndpointAdapter {
             TransportPolicy::N0Defaults => builder.preset(presets::N0),
         };
 
+        if matches!(
+            config.transport,
+            TransportPolicy::N0Defaults | TransportPolicy::DirectOnly
+        ) {
+            let filter = iroh::address_lookup::AddrFilter::new(|addrs| {
+                let mut overlay_ips = std::collections::HashSet::new();
+                for iface in netdev::get_interfaces() {
+                    let name = iface.name.to_lowercase();
+                    let friendly = iface
+                        .friendly_name
+                        .as_deref()
+                        .map(str::to_lowercase)
+                        .unwrap_or_default();
+                    let desc = iface
+                        .description
+                        .as_deref()
+                        .map(str::to_lowercase)
+                        .unwrap_or_default();
+                    if name.starts_with("tsun")
+                        || name.contains("tsunagi")
+                        || name.contains("wintun")
+                        || friendly.starts_with("tsun")
+                        || friendly.contains("tsunagi")
+                        || friendly.contains("wintun")
+                        || desc.contains("wintun")
+                        || desc.contains("tsunagi")
+                    {
+                        for ip in iface.ipv4 {
+                            overlay_ips.insert(std::net::IpAddr::V4(ip.addr()));
+                        }
+                        for ip in iface.ipv6 {
+                            overlay_ips.insert(std::net::IpAddr::V6(ip.addr()));
+                        }
+                    }
+                }
+                if overlay_ips.is_empty() {
+                    return std::borrow::Cow::Borrowed(addrs);
+                }
+                let filtered: Vec<iroh::TransportAddr> = addrs
+                    .iter()
+                    .filter(|addr| match addr {
+                        iroh::TransportAddr::Ip(sock) => !overlay_ips.contains(&sock.ip()),
+                        _ => true,
+                    })
+                    .cloned()
+                    .collect();
+                std::borrow::Cow::Owned(filtered)
+            });
+            builder = builder.addr_filter(filter);
+        }
+
+        #[cfg(feature = "testing")]
+        if let Some(transport) = &config.test_quic_transport {
+            builder = builder.transport_config(transport.clone());
+        }
+
         if !config.bind_addrs.is_empty() {
             builder = builder.clear_ip_transports();
             for addr in &config.bind_addrs {
@@ -233,6 +310,91 @@ impl EndpointAdapter {
             .await
             .map_err(|err| Error::Endpoint(format!("cannot bind endpoint: {err}")))?;
 
+        tracing::debug!(
+            endpoint = %endpoint.id().fmt_short(),
+            bound_sockets = ?endpoint.bound_sockets(),
+            "iroh endpoint bound locally"
+        );
+
+        let ep_id = endpoint.id();
+        let mut addr_watcher = endpoint.watch_addr();
+        let mut relay_watcher = endpoint.home_relay_status();
+        let close_token = endpoint.clone();
+        tokio::spawn(async move {
+            use iroh::Watcher;
+            let mut last_addrs: Vec<SocketAddr> = Vec::new();
+            let mut last_relays: Vec<String> = Vec::new();
+
+            let warn_timer = tokio::time::sleep(Duration::from_secs(12));
+            tokio::pin!(warn_timer);
+            let mut warned = false;
+
+            loop {
+                tokio::select! {
+                    biased;
+                    _ = close_token.closed() => {
+                        break;
+                    }
+                    () = &mut warn_timer, if !warned => {
+                        warned = true;
+                        if last_addrs.is_empty() && !last_relays.is_empty() {
+                            tracing::warn!(
+                                endpoint = %ep_id.fmt_short(),
+                                relays = ?last_relays,
+                                "No direct reflexive or UPnP addresses discovered after probe; communications will rely on relay servers"
+                            );
+                        } else if last_addrs.is_empty() && last_relays.is_empty() {
+                            tracing::warn!(
+                                endpoint = %ep_id.fmt_short(),
+                                "No reachability discovered (neither direct reflexive/UPnP addresses nor relay servers available)"
+                            );
+                        }
+                    }
+                    res = addr_watcher.updated() => {
+                        let addr = match res {
+                            Ok(a) => a,
+                            Err(_) => break,
+                        };
+                        let addrs: Vec<SocketAddr> = addr.ip_addrs().copied().collect();
+                        let relays: Vec<String> = addr.relay_urls().map(|u| u.to_string()).collect();
+
+                        if addrs != last_addrs || relays != last_relays {
+                            tracing::debug!(
+                                endpoint = %ep_id.fmt_short(),
+                                direct_addrs = ?addrs,
+                                relays = ?relays,
+                                "Discovered endpoint reachability (STUN/UPnP/local/relay)"
+                            );
+                            last_addrs = addrs;
+                            last_relays = relays;
+                        }
+                    }
+                    res = relay_watcher.updated() => {
+                        let relays = match res {
+                            Ok(r) => r,
+                            Err(_) => break,
+                        };
+                        for relay in &relays {
+                            if relay.is_connected() {
+                                tracing::debug!(
+                                    endpoint = %ep_id.fmt_short(),
+                                    relay = %relay.url(),
+                                    "Connected to home relay server"
+                                );
+                            } else if let Some(err) = relay.last_error() {
+                                tracing::debug!(
+                                    endpoint = %ep_id.fmt_short(),
+                                    relay = %relay.url(),
+                                    error = %err,
+                                    "Relay connection attempt failed"
+                                );
+                            }
+                        }
+                    }
+                }
+            }
+        });
+
         Ok(Self { endpoint })
     }
 

crates/tsunagi-core/src/overlay/broadcast.rs

@@ -0,0 +1,231 @@
+//! Scoped LAN discovery fanout at IP ingress. A domain is one logical network;
+//! routed copies use ordinary encrypted peer links and are never reflooded.
+//! Future subnet exporters must supply explicitly authorized ingress domains
+//! here, not teach the encrypted transit router to inspect application bytes.
+
+use super::router::{NetworkRoutes, Route};
+use crate::NetworkId;
+use iroh::EndpointId;
+use std::collections::{HashMap, HashSet};
+use std::net::Ipv4Addr;
+use std::sync::Arc;
+
+/// Local participation and the currently authenticated participants in a domain.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastPolicy {
+    /// Whether this agent originates and accepts LAN broadcasts in this network.
+    pub enabled: bool,
+    /// Live members explicitly advertising that they accept broadcasts.
+    pub peers: HashSet<EndpointId>,
+}
+impl Default for BroadcastPolicy {
+    fn default() -> Self {
+        Self {
+            enabled: true,
+            peers: HashSet::new(),
+        }
+    }
+}
+
+#[derive(Debug)]
+struct Domain {
+    network: NetworkId,
+    directed: Ipv4Addr,
+    enabled: bool,
+    recipients: Arc<[Route]>,
+    members: HashSet<EndpointId>,
+}
+
+/// Immutable ingress-domain index rebuilt with address/policy changes.
+#[derive(Debug, Default)]
+pub(crate) struct BroadcastTable {
+    sources: HashMap<Ipv4Addr, Arc<Domain>>,
+    networks: HashMap<NetworkId, Arc<Domain>>,
+    destinations: HashSet<Ipv4Addr>,
+}
+impl BroadcastTable {
+    pub fn build(networks: &HashMap<NetworkId, NetworkRoutes>) -> Self {
+        let mut table = Self::default();
+        for (&network, routes) in networks {
+            let (Some(range), Some(local)) = (routes.range, routes.local) else {
+                continue;
+            };
+            if range.prefix_len > 30 {
+                continue;
+            }
+            let directed = Ipv4Addr::from(u32::from(range.base) | (u32::MAX >> range.prefix_len));
+            let mut recipients: Vec<_> = routes
+                .peers
+                .iter()
+                .filter(|(_, peer)| routes.broadcast.peers.contains(peer))
+                .map(|(_, peer)| Route {
+                    network,
+                    peer: *peer,
+                })
+                .collect();
+            recipients.sort_by_key(|route| route.peer);
+            recipients.dedup();
+            let domain = Arc::new(Domain {
+                network,
+                directed,
+                enabled: routes.broadcast.enabled,
+                members: recipients.iter().map(|route| route.peer).collect(),
+                recipients: recipients.into(),
+            });
+            table.sources.insert(local, domain.clone());
+            table.networks.insert(network, domain);
+            table.destinations.insert(directed);
+        }
+        table
+    }
+    pub fn is_destination(&self, destination: Ipv4Addr) -> bool {
+        destination.is_broadcast() || self.destinations.contains(&destination)
+    }
+    pub fn outgoing(&self, source: Ipv4Addr, destination: Ipv4Addr) -> Option<Arc<[Route]>> {
+        let domain = self.sources.get(&source)?;
+        (domain.enabled && (destination.is_broadcast() || destination == domain.directed))
+            .then(|| domain.recipients.clone())
+    }
+    pub fn accepts(&self, network: NetworkId, peer: EndpointId, destination: Ipv4Addr) -> bool {
+        self.networks.get(&network).is_some_and(|domain| {
+            domain.network == network
+                && domain.enabled
+                && domain.members.contains(&peer)
+                && (destination.is_broadcast() || destination == domain.directed)
+        })
+    }
+}
+
+/// Validates IPv4/UDP lengths before broadcast fanout. No game-specific ports,
+/// payload rewriting or checksum changes. IP fragments retain their bytes and
+/// are reassembled by the destination OS; even later fragments have protocol 17.
+pub(crate) fn valid_udp(packet: &[u8]) -> bool {
+    if packet.len() < 20 || packet[0] >> 4 != 4 || packet[9] != 17 {
+        return false;
+    }
+    let header = usize::from(packet[0] & 15) * 4;
+    let total = usize::from(u16::from_be_bytes([packet[2], packet[3]]));
+    if header < 20 || total <= header || total > packet.len() {
+        return false;
+    }
+    let fragment = u16::from_be_bytes([packet[6], packet[7]]);
+    if fragment & 0x1fff != 0 {
+        return true;
+    }
+    if total < header + 8 {
+        return false;
+    }
+    let udp = usize::from(u16::from_be_bytes([packet[header + 4], packet[header + 5]]));
+    udp >= 8
+        && if fragment & 0x2000 != 0 {
+            udp >= total - header
+        } else {
+            udp == total - header
+        }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+    fn id(name: &str) -> NetworkId {
+        NetworkKeys::derive(
+            &NetworkName::new(name).unwrap(),
+            &NetworkSecret::from_bytes([4; 32]).unwrap(),
+        )
+        .network_id()
+    }
+    #[test]
+    fn limited_and_directed_broadcasts_never_cross_domains_or_target_disabled_peers() {
+        let peer = iroh::SecretKey::from_bytes(&[8; 32]).public();
+        let a = NetworkRoutes {
+            exit: Default::default(),
+            range: Some("10.0.0.0/24".parse().unwrap()),
+            local: Some("10.0.0.1".parse().unwrap()),
+            peers: vec![("10.0.0.2".parse().unwrap(), peer)],
+            broadcast: BroadcastPolicy {
+                enabled: true,
+                peers: HashSet::from([peer]),
+            },
+        };
+        let b = NetworkRoutes {
+            exit: Default::default(),
+            range: Some("10.1.0.0/24".parse().unwrap()),
+            local: Some("10.1.0.1".parse().unwrap()),
+            peers: vec![("10.1.0.2".parse().unwrap(), peer)],
+            broadcast: BroadcastPolicy::default(),
+        };
+        let table = BroadcastTable::build(&HashMap::from([(id("a"), a.clone()), (id("b"), b)]));
+        let routes = table
+            .outgoing(a.local.unwrap(), Ipv4Addr::BROADCAST)
+            .unwrap();
+        assert_eq!(
+            &*routes,
+            &[Route {
+                network: id("a"),
+                peer
+            }]
+        );
+        assert!(
+            table
+                .outgoing(a.local.unwrap(), "10.1.0.255".parse().unwrap())
+                .is_none()
+        );
+        assert!(
+            table
+                .outgoing("192.168.1.2".parse().unwrap(), Ipv4Addr::BROADCAST)
+                .is_none()
+        );
+        assert!(
+            table
+                .outgoing("10.1.0.1".parse().unwrap(), Ipv4Addr::BROADCAST)
+                .unwrap()
+                .is_empty()
+        );
+        assert!(!table.accepts(id("a"), peer, "10.1.0.255".parse().unwrap()));
+        assert!(!table.accepts(id("b"), peer, Ipv4Addr::BROADCAST));
+        assert!(table.accepts(id("a"), peer, Ipv4Addr::BROADCAST));
+        let disabled = NetworkRoutes {
+            exit: Default::default(),
+            broadcast: BroadcastPolicy {
+                enabled: false,
+                ..a.broadcast.clone()
+            },
+            ..a.clone()
+        };
+        let table = BroadcastTable::build(&HashMap::from([(id("a"), disabled)]));
+        assert!(
+            table
+                .outgoing(a.local.unwrap(), Ipv4Addr::BROADCAST)
+                .is_none()
+        );
+        assert!(!table.accepts(id("a"), peer, Ipv4Addr::BROADCAST));
+    }
+    #[test]
+    fn malformed_udp_is_not_amplified_but_ipv4_fragments_are_supported() {
+        let mut packet = vec![0u8; 36];
+        packet[0] = 0x45;
+        packet[3] = 36;
+        packet[9] = 17;
+        packet[25] = 16;
+        assert!(valid_udp(&packet));
+        for length in 0..36 {
+            assert!(!valid_udp(&packet[..length]));
+        }
+        packet[9] = 6;
+        assert!(!valid_udp(&packet));
+        packet[9] = 17;
+        packet[0] = 0x44;
+        assert!(!valid_udp(&packet));
+        packet[0] = 0x45;
+        packet[25] = 8;
+        assert!(!valid_udp(&packet));
+        packet[6] = 0x20;
+        packet[25] = 64;
+        assert!(valid_udp(&packet));
+        packet[6] = 0;
+        packet[7] = 2;
+        assert!(valid_udp(&packet));
+    }
+}

crates/tsunagi-core/src/overlay/config.rs

@@ -11,10 +11,11 @@
 
 use std::net::{IpAddr, Ipv6Addr};
 
-use crate::dataplane::PluginError;
 use crate::identity::NetworkId;
+use crate::overlay::OverlayError;
 
-use super::overlay::OVERLAY_HOST_PREFIX_LEN;
+/// Prefix length of a single IPv6 host address.
+const HOST_PREFIX_LEN: u8 = 128;
 
 /// Longest interface name Linux accepts, excluding the terminating NUL.
 pub const MAX_INTERFACE_NAME_LEN: usize = 15;
@@ -33,13 +34,13 @@ pub struct Cidr {
 
 impl Cidr {
     /// Builds a CIDR, rejecting an impossible prefix length.
-    pub fn new(addr: IpAddr, prefix_len: u8) -> Result<Self, PluginError> {
+    pub fn new(addr: IpAddr, prefix_len: u8) -> Result<Self, OverlayError> {
         let max = match addr {
             IpAddr::V4(_) => 32,
             IpAddr::V6(_) => 128,
         };
         if prefix_len > max {
-            return Err(PluginError::Other(format!(
+            return Err(OverlayError::Other(format!(
                 "prefix length /{prefix_len} is impossible for {addr}"
             )));
         }
@@ -50,7 +51,7 @@ impl Cidr {
     pub fn host(addr: Ipv6Addr) -> Self {
         Self {
             addr: IpAddr::V6(addr),
-            prefix_len: OVERLAY_HOST_PREFIX_LEN,
+            prefix_len: HOST_PREFIX_LEN,
         }
     }
 }
@@ -66,9 +67,9 @@ impl std::fmt::Display for Cidr {
 /// The name is stable across restarts and short enough for the platform. Two
 /// agents on the same host in the same network must be given different
 /// prefixes, or they would derive the same name.
-pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, PluginError> {
+pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, OverlayError> {
     if prefix.is_empty() {
-        return Err(PluginError::Other(
+        return Err(OverlayError::Other(
             "interface prefix must not be empty".into(),
         ));
     }
@@ -76,12 +77,12 @@ pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, Plugin
         .chars()
         .all(|c| c.is_ascii_lowercase() || c.is_ascii_digit())
     {
-        return Err(PluginError::Other(
+        return Err(OverlayError::Other(
             "interface prefix must be lowercase ASCII letters and digits".into(),
         ));
     }
     if prefix.len() >= MAX_INTERFACE_NAME_LEN {
-        return Err(PluginError::Other(format!(
+        return Err(OverlayError::Other(format!(
             "interface prefix must be shorter than {MAX_INTERFACE_NAME_LEN} characters"
         )));
     }

crates/tsunagi-core/src/overlay/exit.rs

@@ -0,0 +1,32 @@
+//! Exit-node host rules: what makes this device an exit node for others, and
+//! what sends this device's own traffic through one.
+//!
+//! The packet side lives in [`super::router`]; this is the side the operating
+//! system sees. It is kept apart from the broadcast rules, which have a
+//! different lifetime and different tags.
+
+/// How one set of host rules went.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct RuleSetReport {
+    /// Whether every step took.
+    pub ok: bool,
+    /// One line: what is in place, or what is missing and why.
+    pub detail: String,
+}
+
+/// What the host rules of one network's exit-node settings are doing.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitRulesReport {
+    /// Masquerading and forwarding for members' traffic, when this agent
+    /// offers itself as an exit node here. `None` means none are wanted.
+    pub offer: Option<RuleSetReport>,
+    /// Whether the kernel forwards packets for the overlay interface.
+    ///
+    /// `Some(false)` means the rules are in place but nothing passes until
+    /// forwarding is turned on, which the agent never does itself. `None`
+    /// when it could not be read.
+    pub forwarding: Option<bool>,
+    /// The routes that send this device's traffic through its exit node.
+    /// `None` means it uses none through this network.
+    pub client: Option<RuleSetReport>,
+}

crates/tsunagi-core/src/overlay/hostrules/exit.rs

@@ -0,0 +1,268 @@
+//! Host rules for exit nodes: what the operating system needs for this agent
+//! to carry other members' internet traffic, and for this device's own
+//! traffic to go through a member.
+//!
+//! Two independent halves, planned together because one interface carries
+//! both:
+//!
+//! * **Offering** — one set of rules per network that offers this agent as an
+//!   exit node, chosen by that network's range: masquerade the range out of
+//!   every other interface and let it through the forward chain. A network
+//!   that does not offer has no rules, so its members' packets are dropped by
+//!   the firewall as well as by the agent.
+//! * **Using** — policy routing that sends everything the host does not have
+//!   a more specific route for into the overlay interface, while the agent's
+//!   own sockets keep the ordinary route (see `linux_exit`).
+//!
+//! Nothing a peer announces becomes a host setting. Offering is the owner's
+//! switch; using is the user's choice of a member. Every object is tagged,
+//! reversible, and removed when it is no longer wanted, when the network
+//! leaves or stops, and when the agent stops.
+
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+
+/// What the host should hold for one interface's exit settings.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitHostPlan {
+    /// The overlay interface.
+    pub interface: String,
+    /// The ranges of every network that offers this agent as an exit node.
+    pub offer: Vec<Ipv4Range>,
+    /// Whether this device sends its internet traffic through a member.
+    pub client: bool,
+    /// The range of every network on the interface. Only the using side
+    /// reads it, to keep this agent's own traffic *to* the overlay inside it
+    /// where the host has no per-user routing (macOS).
+    pub overlay: Vec<Ipv4Range>,
+    /// Underlay addresses this agent talks to directly and must keep reaching
+    /// outside the overlay's default route. Only filled while `client` is set
+    /// and only for hosts that report [`ExitHostRules::needs_bypass`].
+    pub bypass: Vec<std::net::Ipv4Addr>,
+    /// Host names (relay servers) that must stay reachable the same way; the
+    /// host rules resolve them.
+    pub bypass_hosts: Vec<String>,
+}
+
+impl ExitHostPlan {
+    /// Whether there is nothing to hold.
+    pub fn is_empty(&self) -> bool {
+        self.offer.is_empty() && !self.client
+    }
+}
+
+/// What applying a plan achieved.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitHostReport {
+    /// One outcome per offered range: its masquerade and forward rules.
+    pub offer: Vec<(Ipv4Range, RuleOutcome)>,
+    /// Whether the kernel forwards packets for the interface. `None` when it
+    /// could not be read. The agent only reads this: turning forwarding on is
+    /// the owner's decision.
+    pub forwarding: Option<bool>,
+    /// The routes and rules that send this device's traffic through the
+    /// interface, when the plan asked for them.
+    pub client: Option<RuleOutcome>,
+}
+
+/// Installs and removes exit-node rules for an overlay interface.
+///
+/// Like the broadcast rules, no call fails as a whole: a step that cannot be
+/// done is in the report, because the caller's only useful response is to
+/// say so.
+pub trait ExitHostRules: Send + Sync + std::fmt::Debug + 'static {
+    /// Makes the host match the plan, replacing what this agent put there
+    /// for the interface before.
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport>;
+
+    /// Removes everything this agent put there for the interface. Succeeds
+    /// when nothing is there: it runs on the shutdown path, and once at
+    /// startup to take away what a crashed run left.
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()>;
+
+    /// Refreshes volatile local socket protection without replacing exit routes.
+    fn refresh<'a>(&'a self, _plan: &'a ExitHostPlan) -> BoxFuture<'a, Option<RuleOutcome>> {
+        Box::pin(async { None })
+    }
+
+    /// Whether using an exit node on this host needs the underlay addresses
+    /// the agent talks to, because the host cannot exempt the agent's own
+    /// traffic from the default route by user id (Windows). The default is no.
+    fn needs_bypass(&self) -> bool {
+        false
+    }
+}
+
+/// The tag every firewall object of one offered range carries.
+pub fn exit_tag(interface: &str, range: &Ipv4Range) -> String {
+    format!("{}{range}", exit_tag_prefix(interface))
+}
+
+/// The start of every tag for an interface; what cleanup selects by.
+pub fn exit_tag_prefix(interface: &str) -> String {
+    format!("tsunagi-exit:{interface}:")
+}
+
+/// The one-line detail for a set of offer outcomes.
+pub fn offer_detail(outcome: &RuleOutcome, range: &Ipv4Range) -> String {
+    match outcome {
+        RuleOutcome::Applied => format!("masquerade + forward for {range} in place"),
+        RuleOutcome::Failed(reason) => reason.clone(),
+    }
+}
+
+/// An in-memory host for exit rules, for tests: nothing is touched.
+#[derive(Debug, Clone, Default)]
+pub struct MockExitRules {
+    state: Arc<Mutex<Option<ExitHostPlan>>>,
+    calls: Arc<Mutex<Vec<String>>>,
+    forwarding: Arc<Mutex<Option<bool>>>,
+    failure: Arc<Mutex<Option<String>>>,
+    needs_bypass: Arc<Mutex<bool>>,
+}
+
+impl MockExitRules {
+    /// An empty host with forwarding turned on.
+    pub fn new() -> Self {
+        let rules = Self::default();
+        *lock(&rules.forwarding) = Some(true);
+        rules
+    }
+
+    /// What the host reports for kernel forwarding from now on.
+    pub fn set_forwarding(&self, forwarding: Option<bool>) {
+        *lock(&self.forwarding) = forwarding;
+    }
+
+    /// Whether this host asks for the underlay addresses, like Windows.
+    pub fn set_needs_bypass(&self, needs: bool) {
+        *lock(&self.needs_bypass) = needs;
+    }
+
+    /// Makes every step fail with this reason from now on.
+    pub fn fail_with(&self, reason: Option<String>) {
+        *lock(&self.failure) = reason;
+    }
+
+    /// What is installed now.
+    pub fn installed(&self) -> Option<ExitHostPlan> {
+        lock(&self.state).clone()
+    }
+
+    /// Every `apply:<interface>` / `clear:<interface>` call, in order.
+    pub fn calls(&self) -> Vec<String> {
+        lock(&self.calls).clone()
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+impl ExitHostRules for MockExitRules {
+    fn needs_bypass(&self) -> bool {
+        *lock(&self.needs_bypass)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("apply:{}", plan.interface));
+            let outcome = match lock(&self.failure).clone() {
+                Some(reason) => RuleOutcome::Failed(reason),
+                None => {
+                    *lock(&self.state) = Some(plan.clone());
+                    RuleOutcome::Applied
+                }
+            };
+            ExitHostReport {
+                offer: plan
+                    .offer
+                    .iter()
+                    .map(|range| (*range, outcome.clone()))
+                    .collect(),
+                forwarding: *lock(&self.forwarding),
+                client: plan.client.then_some(outcome),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("clear:{interface}"));
+            *lock(&self.state) = None;
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    #[test]
+    fn tags_name_the_interface_and_the_range() {
+        assert_eq!(
+            exit_tag("tsun0", &range()),
+            "tsunagi-exit:tsun0:10.13.37.0/24"
+        );
+        assert!(exit_tag("tsun0", &range()).starts_with(&exit_tag_prefix("tsun0")));
+        // Another interface's tag never matches this one's prefix.
+        assert!(!exit_tag("tsun01", &range()).starts_with(&exit_tag_prefix("tsun0")));
+    }
+
+    #[test]
+    fn an_empty_plan_is_nothing_to_hold() {
+        let mut plan = ExitHostPlan::default();
+        assert!(plan.is_empty());
+        plan.client = true;
+        assert!(!plan.is_empty());
+    }
+
+    #[tokio::test]
+    async fn the_mock_applies_reports_and_clears() {
+        let rules = MockExitRules::new();
+        let plan = ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: vec![range()],
+            client: true,
+            overlay: vec![range()],
+            ..Default::default()
+        };
+        let report = rules.apply(&plan).await;
+        assert_eq!(report.offer[0].1, RuleOutcome::Applied);
+        assert_eq!(report.forwarding, Some(true));
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        assert_eq!(rules.installed(), Some(plan));
+
+        rules.set_forwarding(Some(false));
+        rules.fail_with(Some("denied".into()));
+        let report = rules
+            .apply(&ExitHostPlan {
+                interface: "tsun0".into(),
+                offer: vec![range()],
+                client: false,
+                overlay: Vec::new(),
+                ..Default::default()
+            })
+            .await;
+        assert_eq!(report.forwarding, Some(false));
+        assert!(matches!(report.offer[0].1, RuleOutcome::Failed(_)));
+        assert_eq!(report.client, None);
+
+        rules.clear("tsun0").await;
+        assert_eq!(rules.installed(), None);
+        assert_eq!(rules.calls(), ["apply:tsun0", "apply:tsun0", "clear:tsun0"]);
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/linux.rs

@@ -0,0 +1,710 @@
+//! Broadcast host rules on Linux: a netlink route and an `iptables` rule.
+//!
+//! The route is added by the netlink thread the interface provisioner already
+//! owns (see [`super::super::provision`]), because that thread is the only one
+//! that ever holds `CAP_NET_ADMIN`. It belongs to the interface and goes away
+//! with it.
+//!
+//! The firewall allowance is a rule in the `INPUT` chain, inserted at the top
+//! so a default-deny chain further down cannot drop discovery replies first:
+//!
+//! ```text
+//! iptables -I INPUT 1 -i <if> -p udp -s <range> -m comment --comment tsunagi:<if> -j ACCEPT
+//! ```
+//!
+//! `iptables` is a program, not a netlink call, so it is looked up in fixed
+//! system directories rather than on `PATH`, and its arguments are built from
+//! values this agent computed. It works with both the legacy backend and
+//! `iptables-nft`. A host with only `nft` or `firewalld` has no `iptables`;
+//! that is reported on the firewall half and the route is still installed.
+//!
+//! The rule carries a comment tag. That is how a rule left behind by a crashed
+//! run is found and replaced, and why nothing else in the chain is ever
+//! deleted.
+
+use std::net::Ipv4Addr;
+
+use crate::BoxFuture;
+use crate::overlay::provision::RouteHandle;
+
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome, tag_for};
+
+const CHAIN: &str = "INPUT";
+
+/// Fixed places `iptables` is installed. Never `PATH`: the agent is
+/// privileged and a directory a user controls must not decide what it runs.
+const IPTABLES_DIRECTORIES: [&str; 6] = [
+    "/usr/sbin/iptables",
+    "/sbin/iptables",
+    "/usr/bin/iptables",
+    "/bin/iptables",
+    "/usr/local/sbin/iptables",
+    "/run/current-system/sw/bin/iptables",
+];
+
+/// The first candidate that exists.
+fn pick_iptables(exists: impl Fn(&str) -> bool) -> Option<&'static str> {
+    IPTABLES_DIRECTORIES
+        .into_iter()
+        .find(|candidate| exists(candidate))
+}
+
+/// What running `iptables` produced.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) enum Ran {
+    /// There is no `iptables` to run.
+    Missing,
+    /// It ran, and exited like this.
+    Exited {
+        success: bool,
+        stdout: String,
+        stderr: String,
+    },
+}
+
+/// Runs `iptables` with these arguments. A function value so the flow below
+/// can be tested without touching the host.
+pub(super) type Run<'a> = &'a dyn Fn(&[String]) -> Ran;
+
+/// The match and target of the rule, without the chain or position.
+///
+/// `comment` false leaves the tag off, for a kernel that has no `comment`
+/// match; the rule is then recognised by its shape, see [`tagged_deletions`].
+fn rule_spec(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    let range = plan.range.to_string();
+    let tag = plan.tag();
+    let mut words = vec!["-i", plan.interface.as_str(), "-p", "udp", "-s", &range];
+    if comment {
+        words.extend(["-m", "comment", "--comment", &tag]);
+    }
+    words.extend(["-j", "ACCEPT"]);
+    words.into_iter().map(str::to_string).collect()
+}
+
+/// Whether the kernel has the `comment` match, which a rule needs to carry
+/// its tag. Asked of iptables itself: checking a rule that does not exist
+/// fails either way, but only a missing match says so.
+pub(super) fn comment_supported(run: Run<'_>) -> bool {
+    let probe = locked(
+        [
+            "-t",
+            "filter",
+            "-C",
+            "FORWARD",
+            "-m",
+            "comment",
+            "--comment",
+            "tsunagi-probe",
+            "-j",
+            "ACCEPT",
+        ]
+        .map(str::to_string),
+    );
+    match run(&probe) {
+        Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } => !stderr.contains("No chain/target/match by that name"),
+        _ => true,
+    }
+}
+
+/// `iptables -S` words as option clauses: an option with its value, a negated
+/// one with its `!`. It prints the options in its own order, not the one a
+/// rule was written in, so rules are compared by their clauses, sorted.
+fn clauses(words: &[String]) -> Vec<String> {
+    let mut out = Vec::new();
+    let mut at = 0;
+    while at < words.len() {
+        let take = if words[at] == "!" { 3 } else { 2 };
+        let end = (at + take).min(words.len());
+        out.push(words[at..end].join(" "));
+        at = end;
+    }
+    out.sort();
+    out
+}
+
+/// Whether two rules say the same thing, whatever order they list it in.
+pub(super) fn same_rule(a: &[String], b: &[String]) -> bool {
+    clauses(a) == clauses(b)
+}
+
+/// Waits briefly for the xtables lock instead of failing when another tool
+/// holds it.
+pub(super) fn locked(args: impl IntoIterator<Item = String>) -> Vec<String> {
+    ["-w", "5"]
+        .into_iter()
+        .map(str::to_string)
+        .chain(args)
+        .collect()
+}
+
+fn check_args(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    locked(
+        ["-C".to_string(), CHAIN.to_string()]
+            .into_iter()
+            .chain(rule_spec(plan, comment)),
+    )
+}
+
+fn insert_args(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    locked(
+        ["-I".to_string(), CHAIN.to_string(), "1".to_string()]
+            .into_iter()
+            .chain(rule_spec(plan, comment)),
+    )
+}
+
+fn list_args() -> Vec<String> {
+    locked(["-S".to_string(), CHAIN.to_string()])
+}
+
+/// Splits one `iptables -S` line into words, honouring double quotes, which
+/// `iptables` puts around comments that need them.
+pub(super) fn split_rule_line(line: &str) -> Vec<String> {
+    let mut words = Vec::new();
+    let mut current = String::new();
+    let mut quoted = false;
+    let mut started = false;
+    for c in line.chars() {
+        match c {
+            '"' => {
+                quoted = !quoted;
+                started = true;
+            }
+            c if c.is_whitespace() && !quoted => {
+                if started {
+                    words.push(std::mem::take(&mut current));
+                    started = false;
+                }
+            }
+            c => {
+                current.push(c);
+                started = true;
+            }
+        }
+    }
+    if started {
+        words.push(current);
+    }
+    words
+}
+
+/// Whether a rule is the one this module writes for some range of the
+/// interface the tag belongs to, with no comment on it.
+fn is_uncommented_rule(tag: &str, rest: &[String]) -> bool {
+    let Some(interface) = tag.strip_prefix("tsunagi:") else {
+        return false;
+    };
+    let Some(range) = rest
+        .windows(2)
+        .find(|pair| pair[0] == "-s")
+        .map(|pair| pair[1].as_str())
+    else {
+        return false;
+    };
+    let ours: Vec<String> = ["-i", interface, "-p", "udp", "-s", range, "-j", "ACCEPT"]
+        .map(str::to_string)
+        .to_vec();
+    same_rule(&ours, rest)
+}
+
+/// The delete commands for every rule in an `iptables -S INPUT` listing that
+/// carries exactly this tag, or, with no `comment` match in the kernel, is
+/// exactly the rule written for some range. Anything else is left alone.
+fn tagged_deletions(listing: &str, tag: &str) -> Vec<Vec<String>> {
+    listing
+        .lines()
+        .filter_map(|line| {
+            let words = split_rule_line(line);
+            let [append, chain, rest @ ..] = words.as_slice() else {
+                return None;
+            };
+            if append != "-A" || chain != CHAIN {
+                return None;
+            }
+            let tagged = rest
+                .windows(2)
+                .any(|pair| pair[0] == "--comment" && pair[1] == tag)
+                || is_uncommented_rule(tag, rest);
+            tagged.then(|| {
+                locked(
+                    ["-D".to_string(), CHAIN.to_string()]
+                        .into_iter()
+                        .chain(rest.iter().cloned()),
+                )
+            })
+        })
+        .collect()
+}
+
+/// A failure message with the likely fix for the likely cause.
+pub(super) fn failure(stderr: &str) -> String {
+    let text = stderr.trim();
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("permission denied") || lower.contains("must be root") {
+        format!("iptables needs root: {text}")
+    } else if text.is_empty() {
+        "iptables failed without saying why".to_string()
+    } else {
+        format!("iptables failed: {text}")
+    }
+}
+
+fn missing(plan: &BroadcastRulesPlan) -> String {
+    format!(
+        "iptables not found; install it (the iptables-nft package is enough) or allow inbound \
+         UDP from {} on `{}` in your firewall",
+        plan.range, plan.interface
+    )
+}
+
+/// Makes the firewall hold exactly one rule for this interface, the planned one.
+fn apply_firewall(run: Run<'_>, plan: &BroadcastRulesPlan) -> RuleOutcome {
+    let comment = comment_supported(run);
+    let present = match run(&check_args(plan, comment)) {
+        Ran::Missing => return RuleOutcome::Failed(missing(plan)),
+        Ran::Exited { success, .. } => success,
+    };
+    let listing = match run(&list_args()) {
+        Ran::Missing => return RuleOutcome::Failed(missing(plan)),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => stdout,
+        Ran::Exited { stderr, .. } => return RuleOutcome::Failed(failure(&stderr)),
+    };
+    let stale = tagged_deletions(&listing, &plan.tag());
+    // `-C` matched an identical rule and nothing else carries our tag, so
+    // there is nothing to do. Otherwise start from a clean slate for this
+    // interface: a rule from an earlier run may name another range.
+    if present && stale.len() == 1 {
+        return RuleOutcome::Applied;
+    }
+    for delete in &stale {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(delete)
+        {
+            return RuleOutcome::Failed(failure(&stderr));
+        }
+    }
+    match run(&insert_args(plan, comment)) {
+        Ran::Missing => RuleOutcome::Failed(missing(plan)),
+        Ran::Exited { success: true, .. } => RuleOutcome::Applied,
+        Ran::Exited { stderr, .. } => RuleOutcome::Failed(failure(&stderr)),
+    }
+}
+
+/// Removes every rule tagged for this interface. Best effort.
+fn clear_firewall(run: Run<'_>, interface: &str) {
+    let Ran::Exited {
+        success: true,
+        stdout,
+        ..
+    } = run(&list_args())
+    else {
+        return;
+    };
+    for delete in tagged_deletions(&stdout, &tag_for(interface)) {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(&delete)
+        {
+            tracing::warn!(
+                interface,
+                "cannot remove the broadcast firewall rule: {}",
+                failure(&stderr)
+            );
+        }
+    }
+}
+
+/// Runs the real `iptables`.
+pub(super) fn run_iptables(args: &[String]) -> Ran {
+    run_firewall(false, args)
+}
+
+pub(super) fn run_firewall(ipv6: bool, args: &[String]) -> Ran {
+    let v6_paths: Vec<_> = IPTABLES_DIRECTORIES
+        .iter()
+        .map(|p| p.replace("iptables", "ip6tables"))
+        .collect();
+    let v6_program = v6_paths
+        .iter()
+        .find(|path| std::path::Path::new(path).is_file());
+    let program = if ipv6 {
+        v6_program.map(String::as_str)
+    } else {
+        pick_iptables(|path| std::path::Path::new(path).is_file())
+    };
+    let Some(program) = program else {
+        return Ran::Missing;
+    };
+    match std::process::Command::new(program).args(args).output() {
+        Ok(output) => Ran::Exited {
+            success: output.status.success(),
+            stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
+            stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
+        },
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ran::Missing,
+        Err(err) => Ran::Exited {
+            success: false,
+            stdout: String::new(),
+            stderr: err.to_string(),
+        },
+    }
+}
+
+/// The Linux implementation: netlink for the route, `iptables` for the rule.
+#[derive(Debug)]
+pub struct LinuxHostRules {
+    pub(super) route: RouteHandle,
+    /// Interface, sorted socket snapshot and whether exit routes committed.
+    pub(super) underlay_sockets:
+        std::sync::Mutex<Option<(String, Vec<super::linux_exit_mark::Listener>, bool)>>,
+}
+
+impl LinuxHostRules {
+    pub(crate) fn new(route: RouteHandle) -> Self {
+        Self {
+            route,
+            underlay_sockets: std::sync::Mutex::new(None),
+        }
+    }
+
+    async fn set_route(&self, interface: String, source: Ipv4Addr) -> RuleOutcome {
+        let route = self.route.clone();
+        match tokio::task::spawn_blocking(move || route.set(&interface, source)).await {
+            Ok(Ok(())) => RuleOutcome::Applied,
+            Ok(Err(err)) => RuleOutcome::Failed(err.to_string()),
+            Err(err) => RuleOutcome::Failed(format!("the route task failed: {err}")),
+        }
+    }
+}
+
+impl BroadcastHostRules for LinuxHostRules {
+    fn name(&self) -> &str {
+        "netlink+iptables"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn super::ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = self.set_route(plan.interface.clone(), plan.source).await;
+            let owned = plan.clone();
+            let firewall =
+                tokio::task::spawn_blocking(move || apply_firewall(&run_iptables, &owned))
+                    .await
+                    .unwrap_or_else(|err| {
+                        RuleOutcome::Failed(format!("the firewall task failed: {err}"))
+                    });
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let route = self.route.clone();
+            let name = interface.to_string();
+            let route_name = name.clone();
+            if let Ok(Err(err)) =
+                tokio::task::spawn_blocking(move || route.clear(&route_name)).await
+            {
+                tracing::warn!(interface = %name, %err, "cannot remove the broadcast route");
+            }
+            let name = interface.to_string();
+            let _ = tokio::task::spawn_blocking(move || clear_firewall(&run_iptables, &name)).await;
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::cell::RefCell;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    fn words(args: &[String]) -> String {
+        args.join(" ")
+    }
+
+    #[test]
+    fn the_rule_is_udp_only_from_the_overlay_range_on_this_interface() {
+        let spec = words(&rule_spec(&plan(), true));
+        assert_eq!(
+            spec,
+            "-i tsun0 -p udp -s 10.13.37.0/24 -m comment --comment tsunagi:tsun0 -j ACCEPT"
+        );
+        // Nothing here may widen to other protocols or every source.
+        assert!(spec.contains("-p udp"));
+        assert!(!spec.contains("0.0.0.0/0"));
+        assert!(words(&insert_args(&plan(), true)).starts_with("-w 5 -I INPUT 1 -i tsun0"));
+        assert!(words(&check_args(&plan(), true)).starts_with("-w 5 -C INPUT -i tsun0"));
+    }
+
+    #[test]
+    fn iptables_is_found_in_fixed_directories_never_on_path() {
+        assert_eq!(pick_iptables(|_| false), None);
+        assert_eq!(
+            pick_iptables(|p| p == "/sbin/iptables"),
+            Some("/sbin/iptables")
+        );
+        assert_eq!(pick_iptables(|_| true), Some("/usr/sbin/iptables"));
+    }
+
+    #[test]
+    fn only_rules_with_our_exact_tag_are_selected_for_deletion() {
+        let listing = "\
+-P INPUT DROP
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment \"tsunagi:tsun0\" -j ACCEPT
+-A INPUT -i tsun0 -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -i tsun1 -p udp -m comment --comment tsunagi:tsun1 -j ACCEPT
+-A INPUT -p tcp --dport 22 -m comment --comment \"allow ssh tsunagi:tsun0\" -j ACCEPT
+-A FORWARD -i tsun0 -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -p tcp --dport 22 -j ACCEPT
+";
+        let deletions = tagged_deletions(listing, "tsunagi:tsun0");
+        assert_eq!(deletions.len(), 2, "{deletions:?}");
+        assert_eq!(
+            words(&deletions[0]),
+            "-w 5 -D INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT"
+        );
+        assert!(words(&deletions[1]).contains("-D INPUT -i tsun0 -m comment"));
+        assert!(tagged_deletions("", "tsunagi:tsun0").is_empty());
+    }
+
+    #[test]
+    fn a_kernel_without_the_comment_match_says_so_in_its_error() {
+        let answer = |stderr: &'static str| {
+            move |_: &[String]| Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: stderr.to_string(),
+            }
+        };
+        assert!(!comment_supported(&answer(
+            "iptables: No chain/target/match by that name."
+        )));
+        assert!(comment_supported(&answer(
+            "iptables: Bad rule (does a matching rule exist in that chain?)."
+        )));
+        assert!(comment_supported(&|_: &[String]| Ran::Missing));
+    }
+
+    #[test]
+    fn a_rule_written_without_a_comment_is_found_by_its_shape() {
+        // Listed the way iptables prints it: its own option order.
+        let listing = "\
+-P INPUT ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -j ACCEPT
+-A INPUT -s 10.99.0.0/16 -i tsun0 -p udp -j ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun1 -p udp -j ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -j DROP
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p tcp -j ACCEPT
+";
+        let deletions = tagged_deletions(listing, "tsunagi:tsun0");
+        assert_eq!(deletions.len(), 2, "{deletions:?}");
+        assert!(words(&deletions[0]).ends_with("-s 10.13.37.0/24 -i tsun0 -p udp -j ACCEPT"));
+        assert!(words(&deletions[1]).contains("10.99.0.0/16"));
+    }
+
+    #[test]
+    fn without_a_comment_the_rule_has_none() {
+        let with = words(&rule_spec(&plan(), true));
+        let without = words(&rule_spec(&plan(), false));
+        assert!(with.contains("-m comment --comment tsunagi:tsun0"));
+        assert_eq!(without, "-i tsun0 -p udp -s 10.13.37.0/24 -j ACCEPT");
+    }
+
+    /// A scripted `iptables`: answers by the leading verb, records every call.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        check_ok: bool,
+        listing: String,
+        insert_stderr: Option<&'static str>,
+        missing: bool,
+    }
+
+    impl Fake {
+        fn new(check_ok: bool, listing: &str) -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                check_ok,
+                listing: listing.to_string(),
+                insert_stderr: None,
+                missing: false,
+            }
+        }
+        fn run(&self, args: &[String]) -> Ran {
+            // The probe for the `comment` match: not part of what is tested.
+            if args.iter().any(|word| word == "tsunagi-probe") {
+                return Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "Bad rule".into(),
+                };
+            }
+            self.calls.borrow_mut().push(words(args));
+            if self.missing {
+                return Ran::Missing;
+            }
+            let ok = |stdout: &str| Ran::Exited {
+                success: true,
+                stdout: stdout.to_string(),
+                stderr: String::new(),
+            };
+            match args.get(2).map(String::as_str) {
+                Some("-C") if self.check_ok => ok(""),
+                Some("-C") => Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "Bad rule".into(),
+                },
+                Some("-S") => ok(&self.listing),
+                Some("-I") => match self.insert_stderr {
+                    Some(stderr) => Ran::Exited {
+                        success: false,
+                        stdout: String::new(),
+                        stderr: stderr.into(),
+                    },
+                    None => ok(""),
+                },
+                _ => ok(""),
+            }
+        }
+        fn verbs(&self) -> Vec<String> {
+            self.calls
+                .borrow()
+                .iter()
+                .map(|call| call.split(' ').nth(2).unwrap().to_string())
+                .collect()
+        }
+    }
+
+    const ONE_OURS: &str =
+        "-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT\n";
+
+    #[test]
+    fn an_identical_rule_already_in_place_is_left_alone() {
+        let fake = Fake::new(true, ONE_OURS);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S"], "no insert, no delete");
+    }
+
+    #[test]
+    fn a_missing_rule_is_inserted_once() {
+        let fake = Fake::new(false, "-P INPUT ACCEPT\n");
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-I"]);
+    }
+
+    #[test]
+    fn a_stale_rule_for_another_range_is_replaced_not_stacked() {
+        // A crashed run left a rule for an old range under our tag.
+        let stale = "-A INPUT -s 10.99.0.0/16 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT\n";
+        let fake = Fake::new(false, stale);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-D", "-I"]);
+    }
+
+    #[test]
+    fn duplicates_of_the_rule_collapse_to_one() {
+        let doubled = format!("{ONE_OURS}{ONE_OURS}");
+        let fake = Fake::new(true, &doubled);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-D", "-D", "-I"]);
+    }
+
+    #[test]
+    fn a_missing_iptables_says_what_to_install_and_what_to_allow() {
+        let mut fake = Fake::new(false, "");
+        fake.missing = true;
+        let RuleOutcome::Failed(reason) = apply_firewall(&|a| fake.run(a), &plan()) else {
+            panic!("expected a failure");
+        };
+        assert!(reason.contains("iptables not found"), "{reason}");
+        assert!(reason.contains("10.13.37.0/24"), "{reason}");
+        assert!(reason.contains("tsun0"), "{reason}");
+    }
+
+    #[test]
+    fn a_refused_insert_is_reported_with_the_cause() {
+        let mut fake = Fake::new(false, "");
+        fake.insert_stderr = Some(
+            "iptables v1.8.10: can't initialize iptables table `filter': Permission denied (you must be root)",
+        );
+        let RuleOutcome::Failed(reason) = apply_firewall(&|a| fake.run(a), &plan()) else {
+            panic!("expected a failure");
+        };
+        assert!(reason.starts_with("iptables needs root"), "{reason}");
+    }
+
+    #[test]
+    fn clearing_removes_only_tagged_rules_and_ignores_a_missing_iptables() {
+        let listing = format!(
+            "-A INPUT -p tcp --dport 22 -j ACCEPT\n{ONE_OURS}-A INPUT -i tsun1 -m comment --comment tsunagi:tsun1 -j ACCEPT\n"
+        );
+        let fake = Fake::new(true, &listing);
+        clear_firewall(&|a| fake.run(a), "tsun0");
+        assert_eq!(fake.verbs(), ["-S", "-D"]);
+
+        let mut gone = Fake::new(true, "");
+        gone.missing = true;
+        clear_firewall(&|a| gone.run(a), "tsun0");
+        assert_eq!(gone.verbs(), ["-S"]);
+    }
+
+    #[test]
+    fn rule_listing_words_honour_quotes() {
+        assert_eq!(
+            split_rule_line("-A INPUT -m comment --comment \"a b\" -j ACCEPT"),
+            [
+                "-A",
+                "INPUT",
+                "-m",
+                "comment",
+                "--comment",
+                "a b",
+                "-j",
+                "ACCEPT"
+            ]
+        );
+        assert!(split_rule_line("   ").is_empty());
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/linux_exit.rs

@@ -0,0 +1,968 @@
+//! Exit-node host rules on Linux: `iptables` for the offering side, policy
+//! routing over netlink for the using side.
+//!
+//! # Offering
+//!
+//! For every network that offers this agent as an exit node, three rules,
+//! selected by that network's range and carrying the tag
+//! `tsunagi-exit:<if>:<range>`:
+//!
+//! ```text
+//! -t nat    POSTROUTING  -s <range> ! -o <if> -j MASQUERADE
+//! -t filter FORWARD      -i <if> -s <range> ! -o <if> -j ACCEPT
+//! -t filter FORWARD      -o <if> -d <range> -m conntrack --ctstate RELATED,ESTABLISHED -j ACCEPT
+//! ```
+//!
+//! `! -o <if>` is the egress without naming it: whatever interface the
+//! default route uses, now or after it changes. Kernel forwarding is only
+//! *read*. Turning it on changes how the whole host behaves, which is the
+//! owner's decision, so when it is off the rules are installed anyway and the
+//! report says so.
+//!
+//! # Using
+//!
+//! Policy routing, with no `/1` routes and nothing in the main table:
+//!
+//! ```text
+//! table 0x7473:  default dev <if>
+//! 5205: from all fwmark 0x40000000/0x40000000 lookup main
+//! 5206: from all uidrange <euid>-<euid> lookup main
+//! 5290: from all lookup main suppress_prefixlength 0
+//! 5300: from all lookup 0x7473
+//! ```
+//!
+//! The first two keep the agent's *own* sockets and marked replies off the tunnel: its QUIC, relay,
+//! DHT and DNS traffic would otherwise be routed into the interface it is
+//! carrying, a loop. The specific-route rule lets the LAN, another
+//! VPN, a tunnel to somewhere else — keep working, because only the default
+//! route is suppressed. Mark/UID bypass precedes Tailscale; application rules follow it.
+//!
+//! The UID rule exempts a *user id*, so it is only right when the agent has
+//! one of its own, as the packaged service does. An agent started from a login
+//! session would take that user's own applications out of the tunnel with it,
+//! and the exit node would silently do nothing for them. That is refused, not
+//! guessed at: see [`client_refusal`].
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules, exit_tag, exit_tag_prefix};
+use super::linux::{
+    LinuxHostRules, Ran, Run, comment_supported, failure, locked, run_firewall, run_iptables,
+    same_rule, split_rule_line,
+};
+use super::linux_exit_mark;
+
+/// The chains the offering rules live in: `(table, chain)`.
+const CHAINS: [(&str, &str); 2] = [("nat", "POSTROUTING"), ("filter", "FORWARD")];
+
+/// One firewall rule: where it goes, and its match and target.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Rule {
+    table: &'static str,
+    chain: &'static str,
+    spec: Vec<String>,
+}
+
+fn owned(words: &[&str]) -> Vec<String> {
+    words.iter().map(|word| word.to_string()).collect()
+}
+
+/// The three rules of one offered range.
+///
+/// Each carries a comment naming its interface and range, which is how it is
+/// found again. A kernel without the `comment` match (small router builds
+/// leave it out) cannot take that, so `comment` false leaves it off and the
+/// rules are recognised by their shape instead; see [`tagged_rules`].
+fn offer_rules(interface: &str, range: &Ipv4Range, comment: bool) -> Vec<Rule> {
+    let range_text = range.to_string();
+    let tag = exit_tag(interface, range);
+    let comment: Vec<String> = if comment {
+        owned(&["-m", "comment", "--comment", tag.as_str()])
+    } else {
+        Vec::new()
+    };
+    let spec = |head: &[&str], tail: &[&str]| -> Vec<String> {
+        owned(head)
+            .into_iter()
+            .chain(comment.iter().cloned())
+            .chain(owned(tail))
+            .collect()
+    };
+    vec![
+        Rule {
+            table: "nat",
+            chain: "POSTROUTING",
+            spec: spec(
+                &["-s", &range_text, "!", "-o", interface],
+                &["-j", "MASQUERADE"],
+            ),
+        },
+        Rule {
+            table: "filter",
+            chain: "FORWARD",
+            spec: spec(
+                &["-i", interface, "-s", &range_text, "!", "-o", interface],
+                &["-j", "ACCEPT"],
+            ),
+        },
+        Rule {
+            table: "filter",
+            chain: "FORWARD",
+            spec: spec(
+                &[
+                    "-o",
+                    interface,
+                    "-d",
+                    &range_text,
+                    "-m",
+                    "conntrack",
+                    "--ctstate",
+                    "RELATED,ESTABLISHED",
+                ],
+                &["-j", "ACCEPT"],
+            ),
+        },
+    ]
+}
+
+fn check_args(rule: &Rule) -> Vec<String> {
+    locked(
+        owned(&["-t", rule.table, "-C", rule.chain])
+            .into_iter()
+            .chain(rule.spec.iter().cloned()),
+    )
+}
+
+fn insert_args(rule: &Rule) -> Vec<String> {
+    locked(
+        owned(&["-t", rule.table, "-I", rule.chain, "1"])
+            .into_iter()
+            .chain(rule.spec.iter().cloned()),
+    )
+}
+
+fn list_args(table: &str, chain: &str) -> Vec<String> {
+    locked(owned(&["-t", table, "-S", chain]))
+}
+
+/// A rule found in a listing whose comment starts with our prefix.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Found {
+    /// The whole tag.
+    tag: String,
+    /// The words after `-A <chain>`, which is what `-D` needs.
+    rest: Vec<String>,
+}
+
+/// Every rule in an `iptables -S <chain>` listing that is ours: carrying a
+/// tag that starts with `prefix`, or, on a kernel with no `comment` match,
+/// being exactly one of the three rules this module writes for some range.
+/// Anything else in the chain is left alone.
+fn tagged_rules(listing: &str, chain: &str, prefix: &str) -> Vec<Found> {
+    // `tsunagi-exit:<interface>:`
+    let interface = prefix
+        .strip_prefix("tsunagi-exit:")
+        .and_then(|rest| rest.strip_suffix(':'))
+        .unwrap_or_default();
+    listing
+        .lines()
+        .filter_map(|line| {
+            let words = split_rule_line(line);
+            let [append, found_chain, rest @ ..] = words.as_slice() else {
+                return None;
+            };
+            if append != "-A" || found_chain != chain {
+                return None;
+            }
+            let tag = rest
+                .windows(2)
+                .find(|pair| pair[0] == "--comment" && pair[1].starts_with(prefix))
+                .map(|pair| pair[1].clone())
+                .or_else(|| uncommented_tag(interface, chain, rest))?;
+            Some(Found {
+                tag,
+                rest: rest.to_vec(),
+            })
+        })
+        .collect()
+}
+
+/// The tag a rule would have carried, when it is one of ours written without
+/// a comment: its words are those of the rule for the range it names.
+fn uncommented_tag(interface: &str, chain: &str, rest: &[String]) -> Option<String> {
+    if interface.is_empty() {
+        return None;
+    }
+    let range: Ipv4Range = rest
+        .windows(2)
+        .find(|pair| pair[0] == "-s" || pair[0] == "-d")
+        .and_then(|pair| pair[1].parse().ok())?;
+    offer_rules(interface, &range, false)
+        .iter()
+        .any(|rule| rule.chain == chain && same_rule(&rule.spec, rest))
+        .then(|| exit_tag(interface, &range))
+}
+
+fn delete_args(table: &str, chain: &str, found: &Found) -> Vec<String> {
+    locked(
+        owned(&["-t", table, "-D", chain])
+            .into_iter()
+            .chain(found.rest.iter().cloned()),
+    )
+}
+
+fn missing_iptables(range: &Ipv4Range, interface: &str) -> String {
+    format!(
+        "iptables not found; install it (the iptables-nft package is enough), or masquerade \
+         {range} out of every interface but `{interface}` and forward it yourself"
+    )
+}
+
+/// Reads one chain, or says why not.
+fn read_chain(run: Run<'_>, table: &str, chain: &str) -> Result<String, Failure> {
+    match run(&list_args(table, chain)) {
+        Ran::Missing => Err(Failure::Missing),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => Ok(stdout),
+        Ran::Exited { stderr, .. } => Err(Failure::Failed(failure(&stderr))),
+    }
+}
+
+enum Failure {
+    Missing,
+    Failed(String),
+}
+
+/// Deletes found rules; the first failure ends it.
+fn delete_all(
+    run: Run<'_>,
+    table: &str,
+    chain: &str,
+    found: &[Found],
+    keep: impl Fn(&Found) -> bool,
+) -> Result<(), String> {
+    for rule in found.iter().filter(|rule| !keep(rule)) {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(&delete_args(table, chain, rule))
+        {
+            return Err(failure(&stderr));
+        }
+    }
+    Ok(())
+}
+
+/// Makes the firewall hold exactly the offering rules of these ranges for the
+/// interface, and no others of ours.
+fn apply_offer(
+    run: Run<'_>,
+    interface: &str,
+    ranges: &[Ipv4Range],
+) -> Vec<(Ipv4Range, RuleOutcome)> {
+    let fail_all = |reason: String| -> Vec<(Ipv4Range, RuleOutcome)> {
+        ranges
+            .iter()
+            .map(|range| (*range, RuleOutcome::Failed(reason.clone())))
+            .collect()
+    };
+    let prefix = exit_tag_prefix(interface);
+    let wanted: Vec<String> = ranges.iter().map(|r| exit_tag(interface, r)).collect();
+    let comment = comment_supported(run);
+
+    // What is there now, per chain.
+    let mut present: Vec<(&str, &str, Vec<Found>)> = Vec::new();
+    for (table, chain) in CHAINS {
+        match read_chain(run, table, chain) {
+            Ok(listing) => present.push((table, chain, tagged_rules(&listing, chain, &prefix))),
+            Err(Failure::Missing) => {
+                return match ranges.first() {
+                    Some(range) => fail_all(missing_iptables(range, interface)),
+                    None => Vec::new(),
+                };
+            }
+            Err(Failure::Failed(reason)) => return fail_all(reason),
+        }
+    }
+
+    // Rules for a range no longer offered, or left by an earlier run.
+    for (table, chain, found) in &present {
+        if let Err(reason) = delete_all(run, table, chain, found, |rule| wanted.contains(&rule.tag))
+        {
+            return fail_all(reason);
+        }
+    }
+
+    ranges
+        .iter()
+        .map(|range| {
+            let rules = offer_rules(interface, range, comment);
+            let tag = exit_tag(interface, range);
+            let in_place = rules
+                .iter()
+                .all(|rule| matches!(run(&check_args(rule)), Ran::Exited { success: true, .. }))
+                && present.iter().all(|(_, chain, found)| {
+                    let have = found.iter().filter(|rule| rule.tag == tag).count();
+                    have == rules.iter().filter(|rule| rule.chain == *chain).count()
+                });
+            if in_place {
+                return (*range, RuleOutcome::Applied);
+            }
+            // Start this range from a clean slate: a half-installed set, or
+            // one that differs from the plan, is replaced as a whole.
+            for (table, chain, found) in &present {
+                if let Err(reason) = delete_all(run, table, chain, found, |rule| rule.tag != tag) {
+                    return (*range, RuleOutcome::Failed(reason));
+                }
+            }
+            for rule in &rules {
+                match run(&insert_args(rule)) {
+                    Ran::Missing => {
+                        return (
+                            *range,
+                            RuleOutcome::Failed(missing_iptables(range, interface)),
+                        );
+                    }
+                    Ran::Exited { success: true, .. } => {}
+                    Ran::Exited { stderr, .. } => {
+                        return (*range, RuleOutcome::Failed(failure(&stderr)));
+                    }
+                }
+            }
+            (*range, RuleOutcome::Applied)
+        })
+        .collect()
+}
+
+/// Removes every offering rule of ours for the interface. Best effort.
+fn clear_offer(run: Run<'_>, interface: &str) {
+    let prefix = exit_tag_prefix(interface);
+    for (table, chain) in CHAINS {
+        let Ok(listing) = read_chain(run, table, chain) else {
+            continue;
+        };
+        let found = tagged_rules(&listing, chain, &prefix);
+        if let Err(reason) = delete_all(run, table, chain, &found, |_| false) {
+            tracing::warn!(
+                interface,
+                "cannot remove an exit node firewall rule: {reason}"
+            );
+        }
+    }
+}
+
+/// The kernel's forwarding switch for packets arriving on the interface,
+/// read from the text of the sysctl files. Never written.
+fn forwarding_state(read: impl Fn(&str) -> Option<String>, interface: &str) -> Option<bool> {
+    let flag = |text: String| match text.trim() {
+        "1" => Some(true),
+        "0" => Some(false),
+        _ => None,
+    };
+    // The name is derived here, but it ends up in a path.
+    if !interface.is_empty() && !interface.contains(['/', '\0']) && interface != ".." {
+        let per_interface = format!("/proc/sys/net/ipv4/conf/{interface}/forwarding");
+        if let Some(state) = read(&per_interface).and_then(flag) {
+            return Some(state);
+        }
+    }
+    read("/proc/sys/net/ipv4/ip_forward").and_then(flag)
+}
+
+fn read_sysctl(path: &str) -> Option<String> {
+    std::fs::read_to_string(path).ok()
+}
+
+fn check_kernel_rpf(overlay: &str) -> Result<(), String> {
+    let directory = "/proc/sys/net/ipv4/conf";
+    let value = |name: &str, setting: &str| -> Result<u32, String> {
+        read_sysctl(&format!("{directory}/{name}/{setting}"))
+            .and_then(|s| s.trim().parse().ok())
+            .ok_or_else(|| format!("cannot check {name}/{setting} before installing exit routes"))
+    };
+    let mut interfaces = Vec::new();
+    for entry in std::fs::read_dir(directory).map_err(|err| err.to_string())? {
+        let entry = entry.map_err(|err| err.to_string())?;
+        let name = entry.file_name().to_string_lossy().into_owned();
+        if name != "all" && name != "default" && name != "lo" && name != overlay {
+            interfaces.push((
+                name.clone(),
+                value(&name, "rp_filter")?,
+                value(&name, "src_valid_mark")?,
+            ));
+        }
+    }
+    linux_exit_mark::check_kernel_rpf(
+        value("all", "rp_filter")?,
+        value("all", "src_valid_mark")?,
+        &interfaces,
+    )
+}
+
+/// The effective user id from the text of `/proc/self/status`.
+fn parse_euid(status: &str) -> Option<u32> {
+    status
+        .lines()
+        .find_map(|line| line.strip_prefix("Uid:"))
+        .and_then(|rest| rest.split_whitespace().nth(1))
+        .and_then(|effective| effective.parse().ok())
+}
+
+/// Why the using side must not be installed for this process, if it must not.
+///
+/// The bypass rule exempts this process's *user id* from the tunnel. That is
+/// right for a service with an id of its own and wrong for a process started
+/// from somebody's login session, which shares its id with every application
+/// that person runs. A login session is recognised by `XDG_RUNTIME_DIR`,
+/// which `systemd --user` and the login managers set and a system service
+/// does not. Crude on purpose: refusing a service that was started oddly is
+/// a message, and exempting a user's applications is a silent failure.
+fn client_refusal(xdg_runtime_dir: Option<&str>) -> Option<String> {
+    xdg_runtime_dir.filter(|dir| !dir.is_empty()).map(|_| {
+        "the exit node needs the agent to run as a system service with a user of its own: \
+         this agent was started from a login session, so routing everything through the \
+         overlay would take that user's own applications out of the tunnel with it"
+            .to_string()
+    })
+}
+
+impl LinuxHostRules {
+    async fn blocking<T: Send + 'static>(
+        work: impl FnOnce() -> T + Send + 'static,
+    ) -> Result<T, String> {
+        tokio::task::spawn_blocking(work)
+            .await
+            .map_err(|err| format!("the task failed: {err}"))
+    }
+
+    async fn apply_client(&self, interface: &str) -> RuleOutcome {
+        let xdg = std::env::var("XDG_RUNTIME_DIR").ok();
+        if let Some(reason) = client_refusal(xdg.as_deref()) {
+            // Whatever an earlier run left must not stay in force.
+            let _ = self.clear_client(interface).await;
+            return RuleOutcome::Failed(reason);
+        }
+        let Some(euid) = std::fs::read_to_string("/proc/self/status")
+            .ok()
+            .and_then(|status| parse_euid(&status))
+        else {
+            return RuleOutcome::Failed("cannot tell which user this agent runs as".to_string());
+        };
+        let installed = self
+            .underlay_sockets
+            .lock()
+            .unwrap_or_else(|p| p.into_inner())
+            .as_ref()
+            .is_some_and(|(name, _, routed)| name == interface && *routed);
+        if installed {
+            // A failed socket refresh must never remove an active exit route.
+            return self
+                .refresh_sockets(interface)
+                .await
+                .unwrap_or(RuleOutcome::Applied);
+        }
+        if let Err(reason) = self.clear_client(interface).await {
+            return RuleOutcome::Failed(reason);
+        }
+        let name = interface.to_string();
+        let prepared = Self::blocking(move || {
+            check_kernel_rpf(&name)?;
+            linux_exit_mark::install(
+                &run_firewall,
+                &name,
+                euid,
+                std::path::Path::new("/proc/sys/net/ipv6").is_dir(),
+            )
+        })
+        .await;
+        match prepared {
+            Ok(Ok(sockets)) => {
+                *self
+                    .underlay_sockets
+                    .lock()
+                    .unwrap_or_else(|p| p.into_inner()) = Some((interface.into(), sockets, false));
+            }
+            Ok(Err(reason)) | Err(reason) => return RuleOutcome::Failed(reason),
+        }
+        let route = self.route.clone();
+        let name = interface.to_string();
+        let outcome = match Self::blocking(move || route.set_exit_client(&name, euid)).await {
+            Ok(Ok(())) => RuleOutcome::Applied,
+            Ok(Err(err)) => RuleOutcome::Failed(err.to_string()),
+            Err(reason) => RuleOutcome::Failed(reason),
+        };
+        if let RuleOutcome::Failed(reason) = outcome {
+            let reason = match self.clear_client(interface).await {
+                Ok(()) => reason,
+                Err(cleanup) => format!("{reason}; cannot roll back exit rules: {cleanup}"),
+            };
+            RuleOutcome::Failed(reason)
+        } else {
+            if let Some((_, _, routed)) = self
+                .underlay_sockets
+                .lock()
+                .unwrap_or_else(|p| p.into_inner())
+                .as_mut()
+            {
+                *routed = true;
+            }
+            outcome
+        }
+    }
+
+    async fn clear_client(&self, interface: &str) -> Result<(), String> {
+        let route = self.route.clone();
+        Self::blocking(move || route.clear_exit_client())
+            .await?
+            .map_err(|err| err.to_string())?;
+        // Route removal must precede removing reply protection.
+        let name = interface.to_string();
+        Self::blocking(move || {
+            linux_exit_mark::clear(
+                &run_firewall,
+                &name,
+                std::path::Path::new("/proc/sys/net/ipv6").is_dir(),
+            )
+        })
+        .await??;
+        *self
+            .underlay_sockets
+            .lock()
+            .unwrap_or_else(|p| p.into_inner()) = None;
+        Ok(())
+    }
+
+    async fn refresh_sockets(&self, interface: &str) -> Option<RuleOutcome> {
+        let previous = self
+            .underlay_sockets
+            .lock()
+            .unwrap_or_else(|p| p.into_inner())
+            .clone();
+        if previous
+            .as_ref()
+            .is_none_or(|(name, _, routed)| name != interface || !routed)
+        {
+            return None;
+        }
+        let name = interface.to_string();
+        let work_name = name.clone();
+        let result = Self::blocking(move || {
+            let uid = std::fs::read_to_string("/proc/self/status")
+                .ok()
+                .and_then(|s| parse_euid(&s))
+                .ok_or("cannot read agent UID")?;
+            check_kernel_rpf(&work_name)?;
+            let v6 = std::path::Path::new("/proc/sys/net/ipv6").is_dir();
+            let sockets = linux_exit_mark::local_sockets(uid, v6)?;
+            if previous
+                .as_ref()
+                .is_some_and(|(name, old, _)| name == &work_name && old == &sockets)
+            {
+                return Ok(None);
+            }
+            linux_exit_mark::refresh(&run_firewall, &work_name, v6, &sockets)?;
+            Ok(Some(sockets))
+        })
+        .await;
+        match result {
+            Ok(Ok(None)) => None,
+            Ok(Ok(Some(sockets))) => {
+                *self
+                    .underlay_sockets
+                    .lock()
+                    .unwrap_or_else(|p| p.into_inner()) = Some((name, sockets, true));
+                Some(RuleOutcome::Applied)
+            }
+            Ok(Err(reason)) | Err(reason) => Some(RuleOutcome::Failed(reason)),
+        }
+    }
+}
+
+impl ExitHostRules for LinuxHostRules {
+    fn refresh<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, Option<RuleOutcome>> {
+        Box::pin(async move {
+            if plan.client {
+                self.refresh_sockets(&plan.interface).await
+            } else {
+                None
+            }
+        })
+    }
+
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let interface = plan.interface.clone();
+            let ranges = plan.offer.clone();
+            let offer = {
+                let (job_interface, job_ranges) = (interface.clone(), ranges.clone());
+                Self::blocking(move || apply_offer(&run_iptables, &job_interface, &job_ranges))
+                    .await
+                    .unwrap_or_else(|reason| {
+                        ranges
+                            .iter()
+                            .map(|range| (*range, RuleOutcome::Failed(reason.clone())))
+                            .collect()
+                    })
+            };
+            let forwarding = if ranges.is_empty() {
+                None
+            } else {
+                forwarding_state(read_sysctl, &interface)
+            };
+            let client = if plan.client {
+                Some(self.apply_client(&interface).await)
+            } else {
+                if let Err(reason) = self.clear_client(&interface).await {
+                    tracing::warn!(%reason, "cannot remove exit-client rules");
+                }
+                None
+            };
+            ExitHostReport {
+                offer,
+                forwarding,
+                client,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let _ = Self::blocking(move || clear_offer(&run_iptables, &name)).await;
+            if let Err(reason) = self.clear_client(interface).await {
+                tracing::warn!(%reason, "cannot remove exit-client rules");
+            }
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    fn words(args: &[String]) -> String {
+        args.join(" ")
+    }
+
+    #[test]
+    fn the_rules_masquerade_the_range_out_of_every_other_interface_and_forward_it() {
+        let rules = offer_rules("tsun0", &range(), true);
+        assert_eq!(rules.len(), 3);
+        assert_eq!(
+            words(&rules[0].spec),
+            "-s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j MASQUERADE"
+        );
+        assert_eq!((rules[0].table, rules[0].chain), ("nat", "POSTROUTING"));
+        assert_eq!(
+            words(&rules[1].spec),
+            "-i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        assert_eq!((rules[1].table, rules[1].chain), ("filter", "FORWARD"));
+        assert_eq!(
+            words(&rules[2].spec),
+            "-o tsun0 -d 10.13.37.0/24 -m conntrack --ctstate RELATED,ESTABLISHED -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        // Nothing may widen to every source or to the other direction.
+        for rule in &rules {
+            assert!(!words(&rule.spec).contains("0.0.0.0/0"));
+        }
+        assert!(words(&insert_args(&rules[1])).starts_with("-w 5 -t filter -I FORWARD 1 -i tsun0"));
+        assert!(words(&check_args(&rules[0])).starts_with("-w 5 -t nat -C POSTROUTING -s"));
+    }
+
+    #[test]
+    fn only_rules_with_our_tag_for_this_interface_are_selected() {
+        let listing = "\
+-P FORWARD DROP
+-A FORWARD -i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT
+-A FORWARD -i tsun1 -m comment --comment tsunagi-exit:tsun1:10.99.0.0/24 -j ACCEPT
+-A FORWARD -i tsun01 -m comment --comment tsunagi-exit:tsun01:10.5.0.0/24 -j ACCEPT
+-A FORWARD -m comment --comment \"allow tsunagi-exit:tsun0:x\" -j ACCEPT
+-A FORWARD -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT
+-A FORWARD -j ACCEPT
+";
+        let found = tagged_rules(listing, "FORWARD", &exit_tag_prefix("tsun0"));
+        assert_eq!(found.len(), 1, "{found:?}");
+        assert_eq!(found[0].tag, "tsunagi-exit:tsun0:10.13.37.0/24");
+        assert_eq!(
+            words(&delete_args("filter", "FORWARD", &found[0])),
+            "-w 5 -t filter -D FORWARD -i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        assert!(tagged_rules("", "FORWARD", "tsunagi-exit:tsun0:").is_empty());
+    }
+
+    /// A scripted `iptables` that remembers which rules it was given.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        /// Rules present, as `-S` lines per `(table, chain)`.
+        listing: RefCell<Vec<(String, String)>>,
+        missing: bool,
+        check_ok: bool,
+        /// A kernel with no `comment` match, as small router builds have.
+        no_comment: bool,
+    }
+
+    impl Fake {
+        fn new(check_ok: bool) -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                listing: RefCell::new(Vec::new()),
+                missing: false,
+                check_ok,
+                no_comment: false,
+            }
+        }
+
+        fn with_rule(self, table: &str, line: &str) -> Self {
+            self.listing
+                .borrow_mut()
+                .push((table.to_string(), line.to_string()));
+            self
+        }
+
+        fn run(&self, args: &[String]) -> Ran {
+            if self.missing {
+                return Ran::Missing;
+            }
+            let line = args.join(" ");
+            self.calls.borrow_mut().push(line.clone());
+            let table = args
+                .iter()
+                .position(|word| word == "-t")
+                .map(|at| args[at + 1].clone())
+                .unwrap_or_default();
+            let ok = |stdout: String| Ran::Exited {
+                success: true,
+                stdout,
+                stderr: String::new(),
+            };
+            if self.no_comment && line.contains("-m comment") {
+                return Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "iptables: No chain/target/match by that name.".to_string(),
+                };
+            }
+            if line.contains(" -S ") {
+                let listing: Vec<String> = self
+                    .listing
+                    .borrow()
+                    .iter()
+                    .filter(|(t, _)| *t == table)
+                    .map(|(_, l)| l.clone())
+                    .collect();
+                return ok(listing.join("\n"));
+            }
+            if line.contains(" -C ") {
+                return Ran::Exited {
+                    success: self.check_ok,
+                    stdout: String::new(),
+                    stderr: String::new(),
+                };
+            }
+            ok(String::new())
+        }
+
+        fn verbs(&self, verb: &str) -> Vec<String> {
+            self.calls
+                .borrow()
+                .iter()
+                .filter(|call| call.contains(&format!(" {verb} ")))
+                .cloned()
+                .collect()
+        }
+    }
+
+    #[test]
+    fn a_kernel_without_the_comment_match_gets_the_rules_without_comments() {
+        let mut fake = Fake::new(false);
+        fake.no_comment = true;
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let inserted = fake.verbs("-I");
+        assert_eq!(inserted.len(), 3, "{inserted:?}");
+        assert!(
+            inserted.iter().all(|call| !call.contains("comment")),
+            "{inserted:?}"
+        );
+    }
+
+    #[test]
+    fn rules_written_without_a_comment_are_ours_and_are_removed_with_them() {
+        // The way iptables prints them: its own option order.
+        let listing = "\
+-P FORWARD ACCEPT
+-A FORWARD -s 10.13.37.0/24 -i tsun0 ! -o tsun0 -j ACCEPT
+-A FORWARD -d 10.13.37.0/24 -o tsun0 -m conntrack --ctstate RELATED,ESTABLISHED -j ACCEPT
+-A FORWARD -i tsun0 -j ACCEPT
+-A FORWARD -s 10.13.37.0/24 -i tsun1 ! -o tsun1 -j ACCEPT
+";
+        let found = tagged_rules(listing, "FORWARD", &exit_tag_prefix("tsun0"));
+        assert_eq!(found.len(), 2, "{found:?}");
+        assert!(
+            found
+                .iter()
+                .all(|rule| rule.tag == exit_tag("tsun0", &range()))
+        );
+        let nat = "-A POSTROUTING -s 10.13.37.0/24 ! -o tsun0 -j MASQUERADE\n";
+        assert_eq!(
+            tagged_rules(nat, "POSTROUTING", &exit_tag_prefix("tsun0")).len(),
+            1
+        );
+    }
+
+    #[test]
+    fn a_fresh_host_gets_the_three_rules() {
+        let fake = Fake::new(false);
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let inserted = fake.verbs("-I");
+        assert_eq!(inserted.len(), 3, "{inserted:?}");
+        assert!(inserted[0].contains("-t nat -I POSTROUTING 1"));
+        assert!(fake.verbs("-D").is_empty());
+    }
+
+    #[test]
+    fn rules_already_in_place_are_left_alone() {
+        let tag = exit_tag("tsun0", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -s 10.13.37.0/24 ! -o tsun0 -m comment --comment {tag} -j MASQUERADE"),
+            )
+            .with_rule("filter", &format!("-A FORWARD -i tsun0 -m comment --comment {tag} -j ACCEPT"))
+            .with_rule("filter", &format!("-A FORWARD -o tsun0 -m comment --comment {tag} -j ACCEPT"));
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        assert!(fake.verbs("-I").is_empty());
+        assert!(fake.verbs("-D").is_empty());
+    }
+
+    #[test]
+    fn a_stale_range_is_removed_and_a_half_installed_set_is_replaced() {
+        let old: Ipv4Range = "10.99.0.0/24".parse().unwrap();
+        let old_tag = exit_tag("tsun0", &old);
+        let tag = exit_tag("tsun0", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!(
+                    "-A POSTROUTING -s 10.99.0.0/24 -m comment --comment {old_tag} -j MASQUERADE"
+                ),
+            )
+            // Only one of the three for the current range.
+            .with_rule(
+                "filter",
+                &format!("-A FORWARD -i tsun0 -m comment --comment {tag} -j ACCEPT"),
+            )
+            // Somebody else's rule, which must survive.
+            .with_rule("filter", "-A FORWARD -i eth0 -j ACCEPT");
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let deleted = fake.verbs("-D");
+        assert!(
+            deleted.iter().any(|call| call.contains(&old_tag)),
+            "{deleted:?}"
+        );
+        assert!(
+            deleted.iter().any(|call| call.contains(&tag)),
+            "{deleted:?}"
+        );
+        assert!(
+            deleted.iter().all(|call| !call.contains("eth0")),
+            "{deleted:?}"
+        );
+        assert_eq!(fake.verbs("-I").len(), 3);
+    }
+
+    #[test]
+    fn clearing_removes_only_our_rules_for_the_interface() {
+        let tag = exit_tag("tsun0", &range());
+        let other = exit_tag("tsun1", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -m comment --comment {tag} -j MASQUERADE"),
+            )
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -m comment --comment {other} -j MASQUERADE"),
+            )
+            .with_rule("filter", "-A FORWARD -j ACCEPT");
+        clear_offer(&|args| fake.run(args), "tsun0");
+        let deleted = fake.verbs("-D");
+        assert_eq!(deleted.len(), 1, "{deleted:?}");
+        assert!(deleted[0].contains(&tag));
+    }
+
+    #[test]
+    fn a_missing_iptables_is_a_failure_that_says_what_to_do() {
+        let mut fake = Fake::new(false);
+        fake.missing = true;
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        let RuleOutcome::Failed(reason) = &outcome[0].1 else {
+            panic!("{outcome:?}");
+        };
+        assert!(reason.contains("iptables not found"), "{reason}");
+        assert!(reason.contains("10.13.37.0/24"), "{reason}");
+        // Nothing wanted and nothing to run is not a failure of anything.
+        assert!(apply_offer(&|args| fake.run(args), "tsun0", &[]).is_empty());
+    }
+
+    #[test]
+    fn forwarding_is_read_per_interface_then_globally_and_never_written() {
+        let files = |per: Option<&'static str>, global: Option<&'static str>| {
+            move |path: &str| match path {
+                "/proc/sys/net/ipv4/conf/tsun0/forwarding" => per.map(str::to_string),
+                "/proc/sys/net/ipv4/ip_forward" => global.map(str::to_string),
+                _ => None,
+            }
+        };
+        assert_eq!(
+            forwarding_state(files(Some("1\n"), Some("0\n")), "tsun0"),
+            Some(true)
+        );
+        assert_eq!(
+            forwarding_state(files(Some("0\n"), Some("1\n")), "tsun0"),
+            Some(false)
+        );
+        assert_eq!(
+            forwarding_state(files(None, Some("1\n")), "tsun0"),
+            Some(true)
+        );
+        assert_eq!(forwarding_state(files(None, None), "tsun0"), None);
+        assert_eq!(forwarding_state(files(Some("junk"), None), "tsun0"), None);
+        // A name that would leave the directory is not looked up.
+        assert_eq!(forwarding_state(files(Some("1"), None), "../../x"), None);
+    }
+
+    #[test]
+    fn the_effective_uid_is_the_second_field() {
+        let status = "Name:\ttsunagi\nUid:\t1001\t1002\t1001\t1001\nGid:\t5\t5\t5\t5\n";
+        assert_eq!(parse_euid(status), Some(1002));
+        assert_eq!(parse_euid("Name:\tx\n"), None);
+    }
+
+    #[test]
+    fn a_login_session_is_refused_and_a_service_is_not() {
+        let reason = client_refusal(Some("/run/user/1000")).unwrap();
+        assert!(reason.contains("system service"), "{reason}");
+        assert!(client_refusal(None).is_none());
+        assert!(client_refusal(Some("")).is_none());
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/linux_exit_kernel.rs

@@ -0,0 +1,439 @@
+//! Privileged routing regression, isolated from the host in named namespaces.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use std::io::{BufRead, BufReader, Read, Write};
+use std::net::{SocketAddr, TcpListener, TcpStream, UdpSocket};
+use std::os::unix::process::CommandExt;
+use std::process::{Child, Command};
+use std::time::Duration;
+
+use super::linux::run_firewall;
+use super::linux_exit_mark::{clear, install, local_sockets, refresh};
+use crate::overlay::NetlinkProvisioner;
+
+const TEST: &str = "overlay::hostrules::linux_exit_kernel::strict_firewall_preserves_underlay_and_application_exit_routing";
+const ROLE: &str = "TSUNAGI_EXIT_KERNEL_TEST_ROLE";
+
+fn ip(arguments: &[&str]) -> String {
+    let result = Command::new("ip").args(arguments).output().unwrap();
+    assert!(
+        result.status.success(),
+        "ip {arguments:?}: {}",
+        String::from_utf8_lossy(&result.stderr)
+    );
+    String::from_utf8_lossy(&result.stdout).into_owned()
+}
+
+struct Namespace(String);
+impl Drop for Namespace {
+    fn drop(&mut self) {
+        let _ = Command::new("ip").args(["netns", "del", &self.0]).status();
+    }
+}
+struct Echo(Child);
+impl Drop for Echo {
+    fn drop(&mut self) {
+        let _ = self.0.kill();
+        let _ = self.0.wait();
+    }
+}
+
+fn probes() {
+    let should_pass = std::env::var("TSUNAGI_EXIT_KERNEL_EXPECT").unwrap() == "pass";
+    for address in ["198.51.100.2:42000", "[2001:db8:2::2]:42000"] {
+        let address: SocketAddr = address.parse().unwrap();
+        let socket = UdpSocket::bind(if address.is_ipv4() {
+            "0.0.0.0:0"
+        } else {
+            "[::]:0"
+        })
+        .unwrap();
+        socket
+            .set_read_timeout(Some(Duration::from_secs(1)))
+            .unwrap();
+        socket.send_to(b"rpf", address).unwrap();
+        let mut packet = [0u8; 3];
+        assert_eq!(
+            socket.recv(&mut packet).is_ok(),
+            should_pass,
+            "UDP {address}"
+        );
+        if should_pass {
+            assert_eq!(&packet, b"rpf");
+        }
+        let connected = TcpStream::connect_timeout(&address, Duration::from_secs(1));
+        assert_eq!(connected.is_ok(), should_pass, "TCP {address}");
+        if let Ok(mut stream) = connected {
+            stream
+                .set_read_timeout(Some(Duration::from_secs(1)))
+                .unwrap();
+            stream.write_all(b"rpf").unwrap();
+            stream.read_exact(&mut packet).unwrap();
+            assert_eq!(&packet, b"rpf");
+        }
+    }
+}
+
+fn run_probes(expect: &str) {
+    let result = Command::new(std::env::current_exe().unwrap())
+        .args(["--ignored", "--exact", TEST, "--nocapture"])
+        .env(ROLE, "probe")
+        .env("TSUNAGI_EXIT_KERNEL_EXPECT", expect)
+        .uid(990)
+        .gid(990)
+        .output()
+        .unwrap();
+    assert!(
+        result.status.success(),
+        "{}\n{}",
+        String::from_utf8_lossy(&result.stdout),
+        String::from_utf8_lossy(&result.stderr)
+    );
+}
+
+fn passive_listener() {
+    let listeners: Vec<_> = ["192.0.2.1:42010", "[2001:db8:1::1]:42010"]
+        .into_iter()
+        .map(|address| TcpListener::bind(address).unwrap())
+        .collect();
+    println!("PASSIVE_READY");
+    let tasks: Vec<_> = listeners
+        .into_iter()
+        .map(|listener| {
+            std::thread::spawn(move || {
+                let (mut socket, _) = listener.accept().unwrap();
+                socket
+                    .set_read_timeout(Some(Duration::from_secs(2)))
+                    .unwrap();
+                let mut bytes = [0; 3];
+                socket.read_exact(&mut bytes).unwrap();
+                socket.write_all(&bytes).unwrap();
+            })
+        })
+        .collect();
+    for task in tasks {
+        task.join().unwrap();
+    }
+}
+
+fn udp_listener() {
+    let sockets: Vec<_> = ["192.0.2.1:42011", "[2001:db8:1::1]:42011"]
+        .into_iter()
+        .map(|a| UdpSocket::bind(a).unwrap())
+        .collect();
+    println!("PASSIVE_READY");
+    let tasks: Vec<_> = sockets
+        .into_iter()
+        .map(|socket| {
+            std::thread::spawn(move || {
+                socket
+                    .set_read_timeout(Some(Duration::from_secs(5)))
+                    .unwrap();
+                let mut bytes = [0; 3];
+                let (len, peer) = socket.recv_from(&mut bytes).unwrap();
+                socket.send_to(&bytes[..len], peer).unwrap();
+            })
+        })
+        .collect();
+    for task in tasks {
+        task.join().unwrap();
+    }
+}
+
+fn spawn_listener(role: &str) -> Echo {
+    let process = Command::new(std::env::current_exe().unwrap())
+        .args(["--ignored", "--exact", TEST, "--nocapture"])
+        .env(ROLE, role)
+        .uid(990)
+        .gid(990)
+        .stdout(std::process::Stdio::piped())
+        .spawn()
+        .unwrap();
+    let mut child = Echo(process);
+    let mut output = BufReader::new(child.0.stdout.take().unwrap());
+    loop {
+        let mut line = String::new();
+        assert_ne!(
+            output.read_line(&mut line).unwrap(),
+            0,
+            "listener exited before readiness"
+        );
+        if line.contains("PASSIVE_READY") {
+            // Keep the reader open for libtest's final status after the socket
+            // worker finishes; closing it here makes success a broken pipe.
+            child.0.stdout = Some(output.into_inner());
+            return child;
+        }
+    }
+}
+
+fn client() {
+    for setting in ["rp_filter", "src_valid_mark"] {
+        std::fs::write(format!("/proc/sys/net/ipv4/conf/all/{setting}"), b"1").unwrap();
+    }
+    std::fs::write("/proc/sys/net/ipv4/conf/rpf-c/rp_filter", b"1").unwrap();
+    for family in [false, true] {
+        let args: Vec<_> = [
+            "-t",
+            "mangle",
+            "-A",
+            "PREROUTING",
+            "-m",
+            "rpfilter",
+            "--invert",
+            "--validmark",
+            "-j",
+            "DROP",
+        ]
+        .into_iter()
+        .map(str::to_string)
+        .collect();
+        let result = run_firewall(family, &args);
+        assert!(
+            matches!(result, super::linux::Ran::Exited { success: true, .. }),
+            "{result:?}"
+        );
+    }
+    let provisioner = NetlinkProvisioner::new().unwrap();
+    let route = provisioner.host_rules().route;
+    // Model Tailscale's earlier lookup, without running another VPN daemon.
+    ip(&["-4", "rule", "add", "priority", "5270", "table", "52"]);
+    ip(&[
+        "-4", "route", "add", "default", "dev", "tsun0", "table", "52",
+    ]);
+    ip(&["-6", "rule", "add", "priority", "5270", "table", "52"]);
+    ip(&[
+        "-6",
+        "route",
+        "add",
+        "unreachable",
+        "default",
+        "table",
+        "52",
+    ]);
+    // A foreign VPN must leave neighbour discovery and connected LAN paths
+    // usable; otherwise the fixture itself cannot resolve its uplink neighbour.
+    for prefix in ["fe80::/64", "2001:db8:1::/64"] {
+        ip(&["-6", "route", "add", "throw", prefix, "table", "52"]);
+    }
+    let mut passive = spawn_listener("listener");
+    route.set_exit_client("tsun0", 990).unwrap();
+    // uid bypass alone cannot protect reverse lookups, reproducing the bug.
+    run_probes("blocked");
+    install(&run_firewall, "tsun0", 990, true).unwrap();
+    run_probes("pass");
+    // Pair transports may bind fresh UDP ports after exit activation.
+    let mut dynamic_udp = spawn_listener("udp-listener");
+    refresh(
+        &run_firewall,
+        "tsun0",
+        true,
+        &local_sockets(990, true).unwrap(),
+    )
+    .unwrap();
+    let server = std::env::var("TSUNAGI_EXIT_KERNEL_SERVER").unwrap();
+    let incoming = Command::new("ip").args(["netns", "exec", &server, "python3", "-c", r#"import socket
+for family,source,target in [(socket.AF_INET,'198.51.100.2','192.0.2.1'),(socket.AF_INET6,'2001:db8:2::2','2001:db8:1::1')]:
+    for kind,port in [(socket.SOCK_STREAM,42010),(socket.SOCK_DGRAM,42011)]:
+        s=socket.socket(family,kind); s.settimeout(2); s.bind((source,0)); s.connect((target,port)); s.sendall(b'rpf'); assert s.recv(3)==b'rpf'; s.close()
+"#]).output().unwrap();
+    assert!(
+        incoming.status.success(),
+        "first inbound TCP/UDP must be protected: {}",
+        String::from_utf8_lossy(&incoming.stderr)
+    );
+    assert!(passive.0.wait().unwrap().success());
+    assert!(dynamic_udp.0.wait().unwrap().success());
+    // Repeated apply remains usable, and does not redirect application sockets.
+    install(&run_firewall, "tsun0", 990, true).unwrap();
+    run_probes("pass");
+    assert!(ip(&["-4", "route", "get", "198.51.100.2", "uid", "991"]).contains("dev tsun0"));
+    let blocked = Command::new("ip")
+        .args(["-6", "route", "get", "2001:db8:2::2", "uid", "991"])
+        .output()
+        .unwrap();
+    assert!(
+        !blocked.status.success(),
+        "application IPv6 must stay blocked"
+    );
+    route.clear_exit_client().unwrap();
+    clear(&run_firewall, "tsun0", true).unwrap();
+    for family in ["-4", "-6"] {
+        assert!(
+            ip(&[family, "rule", "show"]).contains("5270:"),
+            "foreign routing rules survive cleanup"
+        );
+        ip(&[family, "rule", "del", "priority", "5270"]);
+    }
+    assert!(ip(&["-4", "route", "get", "198.51.100.2", "uid", "991"]).contains("dev rpf-c"));
+    assert!(ip(&["-6", "route", "get", "2001:db8:2::2", "uid", "991"]).contains("dev rpf-c"));
+    for family in [false, true] {
+        let result = run_firewall(family, &["-t".into(), "mangle".into(), "-S".into()]);
+        if let super::linux::Ran::Exited { stdout, .. } = result {
+            assert!(!stdout.contains("tsunagi-exit-underlay:"));
+            assert!(!stdout.contains("tsunagi-exit-sockets:"));
+            assert!(
+                stdout.contains("rpfilter"),
+                "strict host filtering is retained"
+            );
+        }
+    }
+}
+
+#[test]
+#[ignore = "needs root, iproute2, iptables/ip6tables and Python; changes only temporary namespaces"]
+fn strict_firewall_preserves_underlay_and_application_exit_routing() {
+    match std::env::var(ROLE).as_deref() {
+        Ok("probe") => {
+            probes();
+            return;
+        }
+        Ok("client") => {
+            client();
+            return;
+        }
+        Ok("listener") => {
+            passive_listener();
+            return;
+        }
+        Ok("udp-listener") => {
+            udp_listener();
+            return;
+        }
+        _ => {}
+    }
+    let suffix = std::process::id();
+    let client_name = format!("tsng-rpf-c-{suffix}");
+    let server_name = format!("tsng-rpf-s-{suffix}");
+    ip(&["netns", "add", &client_name]);
+    let _client = Namespace(client_name.clone());
+    ip(&["netns", "add", &server_name]);
+    let _server = Namespace(server_name.clone());
+    ip(&[
+        "link",
+        "add",
+        "rpf-c",
+        "netns",
+        &client_name,
+        "type",
+        "veth",
+        "peer",
+        "name",
+        "rpf-s",
+        "netns",
+        &server_name,
+    ]);
+    for (namespace, interface, v4, v6) in [
+        (&client_name, "rpf-c", "192.0.2.1/24", "2001:db8:1::1/64"),
+        (&server_name, "rpf-s", "192.0.2.2/24", "2001:db8:1::2/64"),
+    ] {
+        ip(&["-n", namespace, "link", "set", "lo", "up"]);
+        ip(&["-n", namespace, "addr", "add", v4, "dev", interface]);
+        ip(&[
+            "-n", namespace, "-6", "addr", "add", v6, "dev", interface, "nodad",
+        ]);
+        ip(&["-n", namespace, "link", "set", interface, "up"]);
+    }
+    ip(&[
+        "-n",
+        &server_name,
+        "addr",
+        "add",
+        "198.51.100.2/32",
+        "dev",
+        "lo",
+    ]);
+    ip(&[
+        "-n",
+        &server_name,
+        "-6",
+        "addr",
+        "add",
+        "2001:db8:2::2/128",
+        "dev",
+        "lo",
+        "nodad",
+    ]);
+    ip(&[
+        "-n",
+        &client_name,
+        "route",
+        "add",
+        "default",
+        "via",
+        "192.0.2.2",
+    ]);
+    ip(&[
+        "-n",
+        &client_name,
+        "-6",
+        "route",
+        "add",
+        "default",
+        "via",
+        "2001:db8:1::2",
+    ]);
+    ip(&["-n", &client_name, "link", "add", "tsun0", "type", "dummy"]);
+    ip(&[
+        "-n",
+        &client_name,
+        "addr",
+        "add",
+        "10.13.37.1/24",
+        "dev",
+        "tsun0",
+    ]);
+    ip(&["-n", &client_name, "link", "set", "tsun0", "up"]);
+    let echo_code = r#"import socket,selectors,sys
+selector=selectors.DefaultSelector()
+for family in [socket.AF_INET,socket.AF_INET6]:
+    for kind in [socket.SOCK_DGRAM,socket.SOCK_STREAM]:
+        sock=socket.socket(family,kind)
+        if family==socket.AF_INET6: sock.setsockopt(socket.IPPROTO_IPV6,socket.IPV6_V6ONLY,1)
+        sock.bind(('198.51.100.2' if family==socket.AF_INET else '2001:db8:2::2',42000))
+        if kind==socket.SOCK_STREAM: sock.listen()
+        selector.register(sock,selectors.EVENT_READ,kind)
+print('ready',flush=True)
+while True:
+    for key,_ in selector.select():
+        if key.data==socket.SOCK_DGRAM:
+            data,address=key.fileobj.recvfrom(128); key.fileobj.sendto(data,address)
+        else:
+            conn,_=key.fileobj.accept(); conn.settimeout(2)
+            try: conn.sendall(conn.recv(128))
+            finally: conn.close()
+"#;
+    let child = Command::new("ip")
+        .args([
+            "netns",
+            "exec",
+            &server_name,
+            "python3",
+            "-u",
+            "-c",
+            echo_code,
+        ])
+        .stdout(std::process::Stdio::piped())
+        .spawn()
+        .unwrap();
+    let mut echo = Echo(child);
+    let mut ready = String::new();
+    BufReader::new(echo.0.stdout.take().unwrap())
+        .read_line(&mut ready)
+        .unwrap();
+    assert_eq!(ready.trim(), "ready");
+    let result = Command::new("ip")
+        .args(["netns", "exec", &client_name])
+        .arg(std::env::current_exe().unwrap())
+        .args(["--ignored", "--exact", TEST, "--nocapture"])
+        .env(ROLE, "client")
+        .env("TSUNAGI_EXIT_KERNEL_SERVER", &server_name)
+        .output()
+        .unwrap();
+    assert!(
+        result.status.success(),
+        "{}\n{}",
+        String::from_utf8_lossy(&result.stdout),
+        String::from_utf8_lossy(&result.stderr)
+    );
+}

crates/tsunagi-core/src/overlay/hostrules/linux_exit_mark.rs

@@ -0,0 +1,535 @@
+//! Own-bit connection marks for exit-client underlay traffic.
+//! Reply marks are restored before host reverse-path filtering, never accepted
+//! around it. No peer address becomes a host route or firewall argument.
+
+use super::linux::{Ran, failure, locked, split_rule_line};
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+
+/// Outside Tailscale's 0xff0000 mask and ordinary low-bit WireGuard marks.
+pub(crate) const UNDERLAY_MARK: u32 = 0x4000_0000;
+pub(crate) const UNDERLAY_MASK: u32 = UNDERLAY_MARK;
+const MARK: &str = "0x40000000/0x40000000";
+const MASK: &str = "0x40000000";
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
+pub(super) struct Listener {
+    address: IpAddr,
+    port: u16,
+    tcp: bool,
+}
+
+fn parse_listeners(listing: &str, uid: u32, v6: bool, tcp: bool) -> Result<Vec<Listener>, String> {
+    let mut result = Vec::new();
+    for line in listing.lines().skip(1) {
+        let fields: Vec<_> = line.split_whitespace().collect();
+        if (tcp && fields.get(3) != Some(&"0A"))
+            || fields.get(7).and_then(|s| s.parse::<u32>().ok()) != Some(uid)
+        {
+            continue;
+        }
+        let (address, port) = fields
+            .get(1)
+            .and_then(|s| s.split_once(':'))
+            .ok_or("malformed local socket")?;
+        let port = u16::from_str_radix(port, 16).map_err(|e| e.to_string())?;
+        let address = if v6 {
+            let mut bytes = [0; 16];
+            for (word, chunk) in bytes.as_chunks_mut::<4>().0.iter_mut().enumerate() {
+                let hex = address
+                    .get(word * 8..word * 8 + 8)
+                    .ok_or("malformed local IPv6 listener")?;
+                chunk.copy_from_slice(
+                    &u32::from_str_radix(hex, 16)
+                        .map_err(|e| e.to_string())?
+                        .to_ne_bytes(),
+                );
+            }
+            IpAddr::V6(Ipv6Addr::from(bytes))
+        } else {
+            IpAddr::V4(Ipv4Addr::from(
+                u32::from_str_radix(address, 16)
+                    .map_err(|e| e.to_string())?
+                    .to_ne_bytes(),
+            ))
+        };
+        if result.len() == 128 {
+            return Err("too many local sockets for exit protection".into());
+        }
+        if port != 0 {
+            result.push(Listener { address, port, tcp });
+        }
+    }
+    Ok(result)
+}
+
+pub(super) fn local_sockets(uid: u32, v6: bool) -> Result<Vec<Listener>, String> {
+    let mut listeners = Vec::new();
+    for (protocol, tcp) in [("tcp", true), ("udp", false)] {
+        for family in [false, true].into_iter().filter(|f| !f || v6) {
+            let path = format!("/proc/net/{protocol}{}", if family { "6" } else { "" });
+            listeners.extend(parse_listeners(
+                &std::fs::read_to_string(path).map_err(|e| e.to_string())?,
+                uid,
+                family,
+                tcp,
+            )?);
+        }
+    }
+    listeners.sort();
+    listeners.dedup();
+    Ok(listeners)
+}
+
+fn listener_rules(interface: &str, listener: &Listener) -> Vec<Vec<String>> {
+    let mut base = words(&[
+        "!",
+        "-i",
+        interface,
+        "-p",
+        if listener.tcp { "tcp" } else { "udp" },
+        "--dport",
+        &listener.port.to_string(),
+    ]);
+    if !listener.address.is_unspecified() {
+        base.extend(words(&["-d", &listener.address.to_string()]));
+    }
+    // Match an actual local socket, not just a destination port. Listening
+    // sockets need protection on the first packet, before any outbound connmark.
+    base.extend(words(&[
+        "-m",
+        "socket",
+        "--nowildcard",
+        "-m",
+        "comment",
+        "--comment",
+        &listener_tag(interface),
+    ]));
+    let save = base
+        .iter()
+        .cloned()
+        .chain(words(&[
+            "-m",
+            "mark",
+            "--mark",
+            MARK,
+            "-j",
+            "CONNMARK",
+            "--save-mark",
+            "--nfmask",
+            MASK,
+            "--ctmask",
+            MASK,
+        ]))
+        .collect();
+    let mark = base
+        .into_iter()
+        .chain(words(&["-j", "MARK", "--set-xmark", MARK]))
+        .collect();
+    vec![save, mark]
+}
+pub(super) type RunFamily<'a> = &'a dyn Fn(bool, &[String]) -> Ran;
+
+fn tag(interface: &str) -> String {
+    format!("tsunagi-exit-underlay:{interface}")
+}
+
+fn listener_tag(interface: &str) -> String {
+    format!("tsunagi-exit-sockets:{interface}")
+}
+
+fn words(values: &[&str]) -> Vec<String> {
+    values.iter().map(|s| (*s).into()).collect()
+}
+
+fn rules(interface: &str, uid: u32) -> Vec<(&'static str, Vec<String>)> {
+    let tag = tag(interface);
+    let uid = uid.to_string();
+    // Insert at position one in this order: final OUTPUT order is MARK, SAVE.
+    vec![
+        (
+            "PREROUTING",
+            words(&[
+                "!",
+                "-i",
+                interface,
+                "-m",
+                "connmark",
+                "--mark",
+                MARK,
+                "-m",
+                "comment",
+                "--comment",
+                &tag,
+                "-j",
+                "CONNMARK",
+                "--restore-mark",
+                "--nfmask",
+                MASK,
+                "--ctmask",
+                MASK,
+            ]),
+        ),
+        (
+            "OUTPUT",
+            words(&[
+                "-m",
+                "owner",
+                "--uid-owner",
+                &uid,
+                "-m",
+                "mark",
+                "--mark",
+                MARK,
+                "-m",
+                "comment",
+                "--comment",
+                &tag,
+                "-j",
+                "CONNMARK",
+                "--save-mark",
+                "--nfmask",
+                MASK,
+                "--ctmask",
+                MASK,
+            ]),
+        ),
+        (
+            "OUTPUT",
+            words(&[
+                "-m",
+                "owner",
+                "--uid-owner",
+                &uid,
+                "-m",
+                "comment",
+                "--comment",
+                &tag,
+                "-j",
+                "MARK",
+                "--set-xmark",
+                MARK,
+            ]),
+        ),
+    ]
+}
+
+fn execute(run: RunFamily<'_>, v6: bool, args: Vec<String>) -> Result<String, String> {
+    match run(v6, &locked(args)) {
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => Ok(stdout),
+        Ran::Exited { stderr, .. } => Err(failure(&stderr)),
+        Ran::Missing => Err(format!(
+            "{} is required for exit-node connection marks",
+            if v6 { "ip6tables" } else { "iptables" }
+        )),
+    }
+}
+
+/// Removes only exact interface tags, retaining other agents and host rules.
+pub(super) fn clear(run: RunFamily<'_>, interface: &str, v6: bool) -> Result<(), String> {
+    clear_tags(run, interface, v6, false)
+}
+
+fn clear_tags(
+    run: RunFamily<'_>,
+    interface: &str,
+    v6: bool,
+    sockets_only: bool,
+) -> Result<(), String> {
+    let wanted = tag(interface);
+    let sockets = listener_tag(interface);
+    for family in [false, true].into_iter().filter(|family| !family || v6) {
+        let chains: &[&str] = if sockets_only {
+            &["PREROUTING"]
+        } else {
+            &["OUTPUT", "PREROUTING"]
+        };
+        for chain in chains {
+            let listing = execute(run, family, words(&["-t", "mangle", "-S", chain]))?;
+            for line in listing.lines() {
+                let tokens = split_rule_line(line);
+                let [append, listed_chain, rest @ ..] = tokens.as_slice() else {
+                    continue;
+                };
+                if append == "-A"
+                    && listed_chain == *chain
+                    && rest.windows(2).any(|pair| {
+                        pair[0] == "--comment"
+                            && (pair[1] == sockets || (!sockets_only && pair[1] == wanted))
+                    })
+                {
+                    let args = words(&["-t", "mangle", "-D", chain])
+                        .into_iter()
+                        .chain(rest.iter().cloned())
+                        .collect();
+                    execute(run, family, args)?;
+                }
+            }
+        }
+    }
+    Ok(())
+}
+
+/// Establishes protection before routing may redirect application traffic.
+pub(super) fn install(
+    run: RunFamily<'_>,
+    interface: &str,
+    uid: u32,
+    v6: bool,
+) -> Result<Vec<Listener>, String> {
+    let listeners = local_sockets(uid, v6)?;
+    install_with_listeners(run, interface, uid, v6, &listeners)?;
+    Ok(listeners)
+}
+
+fn install_with_listeners(
+    run: RunFamily<'_>,
+    interface: &str,
+    uid: u32,
+    v6: bool,
+    listeners: &[Listener],
+) -> Result<(), String> {
+    clear(run, interface, v6)?;
+    let result = (|| {
+        for family in [false, true].into_iter().filter(|family| !family || v6) {
+            for (chain, spec) in rules(interface, uid) {
+                let args = words(&["-t", "mangle", "-I", chain, "1"])
+                    .into_iter()
+                    .chain(spec)
+                    .collect();
+                execute(run, family, args)?;
+            }
+            for listener in listeners.iter().filter(|s| s.address.is_ipv6() == family) {
+                for spec in listener_rules(interface, listener) {
+                    execute(
+                        run,
+                        family,
+                        words(&["-t", "mangle", "-I", "PREROUTING", "1"])
+                            .into_iter()
+                            .chain(spec)
+                            .collect(),
+                    )?;
+                }
+            }
+        }
+        Ok(())
+    })();
+    if result.is_err()
+        && let Err(cleanup) = clear(run, interface, v6)
+    {
+        return Err(format!(
+            "{}; cannot roll back connection marks: {cleanup}",
+            result.err().unwrap_or_default()
+        ));
+    }
+    result
+}
+
+/// Updates passive protection without interrupting established connmark rules.
+pub(super) fn refresh(
+    run: RunFamily<'_>,
+    interface: &str,
+    v6: bool,
+    listeners: &[Listener],
+) -> Result<(), String> {
+    clear_tags(run, interface, v6, true)?;
+    for family in [false, true].into_iter().filter(|f| !f || v6) {
+        for listener in listeners.iter().filter(|s| s.address.is_ipv6() == family) {
+            for spec in listener_rules(interface, listener) {
+                execute(
+                    run,
+                    family,
+                    words(&["-t", "mangle", "-I", "PREROUTING", "1"])
+                        .into_iter()
+                        .chain(spec)
+                        .collect(),
+                )?;
+            }
+        }
+    }
+    Ok(())
+}
+
+/// Kernel strict RPF must include the restored mark, independently of iptables.
+pub(super) fn check_kernel_rpf(
+    all_rpf: u32,
+    all_valid: u32,
+    interfaces: &[(String, u32, u32)],
+) -> Result<(), String> {
+    for (interface, rpf, valid) in interfaces {
+        if all_rpf.max(*rpf) == 1 && all_valid.max(*valid) == 0 {
+            return Err(format!(
+                "strict rp_filter on `{interface}` ignores connection marks; enable net.ipv4.conf.all.src_valid_mark=1 before using an exit node (rp_filter stays enabled)"
+            ));
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use std::cell::RefCell;
+
+    #[test]
+    fn both_families_restore_before_firewall_and_preserve_foreign_mark_bits() {
+        let calls = RefCell::new(Vec::new());
+        let run = |v6, args: &[String]| {
+            calls.borrow_mut().push((v6, args.to_vec()));
+            Ran::Exited {
+                success: true,
+                stdout: String::new(),
+                stderr: String::new(),
+            }
+        };
+        install_with_listeners(&run, "tsun0", 990, true, &[]).unwrap();
+        let calls = calls.borrow();
+        for family in [false, true] {
+            let inserts: Vec<_> = calls
+                .iter()
+                .filter(|(v6, a)| *v6 == family && a.iter().any(|s| s == "-I"))
+                .map(|(_, a)| a.join(" "))
+                .collect();
+            assert_eq!(inserts.len(), 3);
+            assert!(inserts[0].contains("-I PREROUTING 1 ! -i tsun0 -m connmark"));
+            assert!(inserts[0].contains("--restore-mark --nfmask 0x40000000 --ctmask 0x40000000"));
+            assert!(inserts[1].contains("--uid-owner 990"));
+            assert!(inserts[1].contains("--save-mark"));
+            assert!(inserts[2].contains("--set-xmark 0x40000000/0x40000000"));
+            assert!(
+                !inserts
+                    .iter()
+                    .any(|s| s.contains("-j ACCEPT") || s.contains("-j RETURN"))
+            );
+        }
+        assert_eq!(UNDERLAY_MASK & 0x00ff_0000, 0);
+    }
+
+    #[test]
+    fn cleanup_retains_other_interfaces_and_rolls_back_partial_ipv6_installation() {
+        let calls = RefCell::new(Vec::new());
+        let run = |v6, args: &[String]| {
+            calls.borrow_mut().push((v6, args.to_vec()));
+            let listing = "-A OUTPUT -m comment --comment tsunagi-exit-underlay:tsun0 -j MARK --set-xmark 0x40000000/0x40000000\n-A OUTPUT -m comment --comment tsunagi-exit-underlay:tsun1 -j MARK --set-xmark 0x40000000/0x40000000\n-A OUTPUT -m comment --comment unrelated -j RETURN\n";
+            Ran::Exited {
+                success: !(v6 && args.iter().any(|s| s == "-I")),
+                stdout: if args.iter().any(|s| s == "-S") {
+                    listing.into()
+                } else {
+                    String::new()
+                },
+                stderr: "injected IPv6 failure".into(),
+            }
+        };
+        assert!(install_with_listeners(&run, "tsun0", 990, true, &[]).is_err());
+        let calls = calls.borrow();
+        let deletions: Vec<_> = calls
+            .iter()
+            .filter(|(_, a)| a.iter().any(|s| s == "-D"))
+            .collect();
+        assert!(
+            deletions.len() >= 4,
+            "both initial and rollback cleanup run"
+        );
+        assert!(
+            deletions
+                .iter()
+                .all(|(_, a)| a.iter().any(|s| s == "tsunagi-exit-underlay:tsun0"))
+        );
+    }
+
+    #[test]
+    fn kernel_strict_validation_requires_mark_awareness_without_weakening_rpf() {
+        let interfaces = vec![("eth0".into(), 1, 0)];
+        assert!(check_kernel_rpf(0, 0, &interfaces).is_err());
+        assert!(check_kernel_rpf(0, 1, &interfaces).is_ok());
+        assert!(check_kernel_rpf(2, 0, &interfaces).is_ok());
+        assert!(check_kernel_rpf(0, 0, &[("eth0".into(), 1, 1)]).is_ok());
+    }
+
+    #[test]
+    fn passive_tcp_protection_uses_only_locally_owned_listeners() {
+        let listing = "header\n0: 00000000:01BB 00000000:0000 0A 0:0 0:0 0 990 0 1\n1: 0100007F:1111 00000000:0000 0A 0:0 0:0 0 1000 0 2\n2: 0100007F:2222 00000000:0000 01 0:0 0:0 0 990 0 3\n";
+        let listeners = parse_listeners(listing, 990, false, true).unwrap();
+        assert_eq!(
+            listeners,
+            vec![Listener {
+                address: IpAddr::V4(Ipv4Addr::UNSPECIFIED),
+                port: 443,
+                tcp: true
+            }]
+        );
+        let rules = listener_rules("tsun0", &listeners[0]);
+        assert!(
+            rules[0]
+                .join(" ")
+                .contains("-p tcp --dport 443 -m socket --nowildcard")
+        );
+        assert!(rules[0].join(" ").contains("--save-mark"));
+        assert!(rules[1].join(" ").contains("--set-xmark"));
+    }
+
+    #[test]
+    fn listener_parser_checks_family_encoding_and_bounds() {
+        let loopback = Ipv6Addr::LOCALHOST.octets();
+        let hex: String = loopback
+            .as_chunks::<4>()
+            .0
+            .iter()
+            .map(|bytes| format!("{:08X}", u32::from_ne_bytes(*bytes)))
+            .collect();
+        let listing = format!("header\n0: {hex}:01BB 0:0 0A 0:0 0:0 0 990 0 1\n");
+        assert_eq!(
+            parse_listeners(&listing, 990, true, true).unwrap(),
+            vec![Listener {
+                address: IpAddr::V6(Ipv6Addr::LOCALHOST),
+                port: 443,
+                tcp: true
+            }]
+        );
+        assert!(
+            parse_listeners(
+                "header\n0: BAD:01BB 0:0 0A 0:0 0:0 0 990 0 1",
+                990,
+                true,
+                true
+            )
+            .is_err()
+        );
+        let huge = format!(
+            "header\n{}",
+            "0: 00000000:01BB 0:0 0A 0:0 0:0 0 990 0 1\n".repeat(129)
+        );
+        assert!(parse_listeners(&huge, 990, false, true).is_err());
+    }
+
+    #[test]
+    fn udp_refresh_preserves_general_connmark_protection_and_other_interfaces() {
+        let listing = "header\n0: 00000000:ABA0 0:0 07 0:0 0:0 0 990 0 1\n1: 00000000:0000 0:0 07 0:0 0:0 0 990 0 2\n";
+        let sockets = parse_listeners(listing, 990, false, false).unwrap();
+        assert_eq!(sockets.len(), 1);
+        assert!(!sockets[0].tcp);
+        let calls = RefCell::new(Vec::new());
+        let run = |v6, args: &[String]| {
+            calls.borrow_mut().push((v6, args.join(" ")));
+            Ran::Exited { success: true, stdout: "-A PREROUTING -m comment --comment tsunagi-exit-underlay:tsun0 -j CONNMARK --restore-mark\n-A PREROUTING -m comment --comment tsunagi-exit-sockets:tsun0 -j MARK\n-A PREROUTING -m comment --comment tsunagi-exit-sockets:tsun1 -j MARK\n".into(), stderr: String::new() }
+        };
+        refresh(&run, "tsun0", false, &sockets).unwrap();
+        let calls = calls.borrow();
+        assert!(
+            calls
+                .iter()
+                .any(|(_, a)| a.contains("-I PREROUTING 1") && a.contains("-p udp"))
+        );
+        assert!(
+            calls
+                .iter()
+                .filter(|(_, a)| a.contains("-D"))
+                .all(|(_, a)| a.contains("tsunagi-exit-sockets:tsun0"))
+        );
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/macos.rs

@@ -0,0 +1,233 @@
+//! Broadcast host rules on macOS: a `route` entry; the firewall is deferred.
+//!
+//! The limited-broadcast route is installed with `/sbin/route`, so a game's
+//! LAN discovery packet leaves through the overlay interface. It belongs to
+//! the interface and is removed when broadcast is turned off, the network
+//! leaves, or the agent stops — and, because a utun is torn down with the
+//! process, it also goes away on its own if the agent is killed.
+//!
+//! The inbound-UDP firewall allowance is **not** implemented here. On macOS
+//! that means a `pf` anchor, which is a good deal more machinery; until it is
+//! written the firewall half reports itself as missing with what to allow by
+//! hand, and the route is still installed. Failure is contained: neither half
+//! can stop the agent, and `status` shows exactly which is in place.
+//!
+//! `route` is a program, not a system call, so it is run by absolute path from
+//! `/sbin` rather than off `PATH`, and its arguments are built from values
+//! this agent computed — the interface name the kernel assigned, nothing a
+//! peer said.
+
+use crate::BoxFuture;
+
+use std::sync::{Arc, Mutex};
+
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules};
+use super::macos_exit::{apply_plan, clear_all, run_tool};
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome};
+
+/// The limited-broadcast destination routed through the overlay interface.
+const LIMITED_BROADCAST: &str = "255.255.255.255";
+
+/// The `route` arguments that direct the limited broadcast out an interface.
+fn route_add_args(interface: &str) -> Vec<String> {
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "add".into(),
+        "-host".into(),
+        LIMITED_BROADCAST.into(),
+        "-interface".into(),
+        interface.into(),
+    ]
+}
+
+/// The `route` arguments that remove it again.
+fn route_del_args(interface: &str) -> Vec<String> {
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "delete".into(),
+        "-host".into(),
+        LIMITED_BROADCAST.into(),
+        "-interface".into(),
+        interface.into(),
+    ]
+}
+
+/// What to do by hand while the `pf` allowance is unimplemented.
+fn firewall_todo(plan: &BroadcastRulesPlan) -> String {
+    format!(
+        "the inbound-UDP firewall allowance is not implemented on macOS yet; if the host \
+         firewall drops discovery replies, allow inbound UDP from {} on `{}` with a pf rule \
+         yourself",
+        plan.range, plan.interface
+    )
+}
+
+/// The macOS implementation: `route` for the route, no firewall yet, and
+/// `pf` plus `route` for exit nodes (see [`super::macos_exit`]).
+#[derive(Debug, Default)]
+pub struct MacosHostRules {
+    /// The reference `pfctl -E` handed out, held while exit-node rules need
+    /// `pf` on and released when they are removed.
+    pub(super) pf_token: Arc<Mutex<Option<String>>>,
+}
+
+impl MacosHostRules {
+    /// Creates the host rules.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Installs the route, replacing one this agent added earlier so the call
+    /// can be repeated. The delete is best effort: a route that is not there
+    /// is the state the add then reaches anyway.
+    async fn set_route(&self, interface: String) -> RuleOutcome {
+        let _ = run(route_del_args(&interface)).await;
+        match run(route_add_args(&interface)).await {
+            Ok(()) => RuleOutcome::Applied,
+            Err(reason) => RuleOutcome::Failed(reason),
+        }
+    }
+}
+
+impl ExitHostRules for MacosHostRules {
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let job_plan = plan.clone();
+            let token = Arc::clone(&self.pf_token);
+            tokio::task::spawn_blocking(move || apply_plan(&run_tool, &token, &job_plan))
+                .await
+                .unwrap_or_else(|err| {
+                    let failed = RuleOutcome::Failed(format!("the task failed: {err}"));
+                    ExitHostReport {
+                        offer: plan
+                            .offer
+                            .iter()
+                            .map(|range| (*range, failed.clone()))
+                            .collect(),
+                        forwarding: None,
+                        client: plan.client.then_some(failed),
+                    }
+                })
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let token = Arc::clone(&self.pf_token);
+            let _ = tokio::task::spawn_blocking(move || clear_all(&run_tool, &token, &name)).await;
+        })
+    }
+}
+
+impl BroadcastHostRules for MacosHostRules {
+    fn name(&self) -> &str {
+        "route"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = self.set_route(plan.interface.clone()).await;
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall: RuleOutcome::Failed(firewall_todo(plan)),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let Err(reason) = run(route_del_args(interface)).await {
+                tracing::debug!(interface = %interface, %reason, "cannot remove the broadcast route");
+            }
+        })
+    }
+}
+
+/// Runs `/sbin/route`, turning a non-zero exit into the message it printed.
+async fn run(args: Vec<String>) -> Result<(), String> {
+    let display = format!("route {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new("/sbin/route")
+            .args(&args)
+            .output()
+    })
+    .await
+    .map_err(|err| format!("could not run {display}: {err}"))?
+    .map_err(|err| format!("could not run {display}: {err}"))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = stderr.trim();
+    if text.is_empty() {
+        Err(format!("{display} failed with {}", output.status))
+    } else {
+        Err(format!("{display} failed: {text}"))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::net::Ipv4Addr;
+
+    use super::*;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "utun5".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn the_route_is_the_limited_broadcast_out_the_interface() {
+        assert_eq!(
+            route_add_args("utun5"),
+            vec![
+                "-q",
+                "-n",
+                "add",
+                "-host",
+                "255.255.255.255",
+                "-interface",
+                "utun5"
+            ]
+        );
+        assert_eq!(
+            route_del_args("utun5"),
+            vec![
+                "-q",
+                "-n",
+                "delete",
+                "-host",
+                "255.255.255.255",
+                "-interface",
+                "utun5"
+            ]
+        );
+    }
+
+    #[test]
+    fn the_firewall_todo_names_the_range_and_interface() {
+        let todo = firewall_todo(&plan());
+        assert!(todo.contains("10.13.37.0/24"), "{todo}");
+        assert!(todo.contains("utun5"), "{todo}");
+        assert!(todo.contains("pf"), "{todo}");
+        // Deferred, so the firewall half is always reported as not in place;
+        // the route runs for real and so is left to an on-host check. Running
+        // `route` here would need root and would change the host routing
+        // table, which the default test suite must never do.
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/macos_exit.rs

@@ -0,0 +1,915 @@
+//! Exit-node host rules on macOS: a `pf` anchor and `route` entries.
+//!
+//! Compiled on every platform so the rule text and the parsers are tested
+//! everywhere; only `MacosHostRules` runs them, and only on macOS.
+//!
+//! # Offering
+//!
+//! Per offered range, in the anchor `com.apple/tsunagi-exit-<if>`:
+//!
+//! ```text
+//! nat on <egress> inet from <range> to any -> (<egress>)
+//! pass in quick on <if> inet from <range> to ! <range>
+//! ```
+//!
+//! The anchor lives under `com.apple/` on purpose: the stock `/etc/pf.conf`
+//! evaluates `nat-anchor "com.apple/*"` and `anchor "com.apple/*"`, so the
+//! rules take effect without editing the user's ruleset. `pf` cannot say
+//! "every interface but one" in a translation address, so the egress is the
+//! interface of the default route **when the rules are applied**; after the
+//! default route moves (Wi-Fi to Ethernet) the rules are re-applied by
+//! switching the exit node off and on, or restarting the agent.
+//!
+//! The kernel's `net.inet.ip.forwarding` is only read, never written: it
+//! changes how the whole host behaves, which is the owner's decision.
+//!
+//! # Using
+//!
+//! macOS has no `uidrange` policy routing, and the agent is root, so the
+//! Linux scheme does not carry over. Instead:
+//!
+//! ```text
+//! route add -net 0.0.0.0/1   -interface <if>
+//! route add -net 128.0.0.0/1 -interface <if>
+//! pass out quick on <if> route-to (<egress> <gateway>) inet proto { tcp udp }
+//!     from any to ! <overlay ranges> user <agent uid>
+//! nat on <egress> inet from (<if>) to any -> (<egress>)
+//! ```
+//!
+//! The two halves outrank the default route without replacing it, and every
+//! more specific route (LAN, another VPN) keeps working. The `route-to` rule
+//! sends the *agent's own* traffic — QUIC, relay, DHT, which would otherwise
+//! loop into the tunnel it carries — back out the physical interface, and the
+//! translation gives it that interface's address, since an unbound socket
+//! picked the overlay address when the route pointed at the overlay. Like the
+//! Linux exemption this is by user id, so anything else running as root also
+//! bypasses the tunnel.
+//!
+//! IPv6 is blocked while an exit node is in use, as on Linux: `::/1` and
+//! `8000::/1` are rejecting routes, so applications fall back to IPv4 at once
+//! instead of leaking.
+//!
+//! Every object is derived from local state and the interface the agent owns;
+//! none of it comes from a peer.
+
+use std::net::Ipv4Addr;
+use std::sync::Mutex;
+
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+use super::exit::{ExitHostPlan, ExitHostReport};
+
+/// The programs this module runs, always by absolute path.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(super) enum Tool {
+    Route,
+    Pfctl,
+    Sysctl,
+    Id,
+}
+
+impl Tool {
+    fn path(self) -> &'static str {
+        match self {
+            Self::Route => "/sbin/route",
+            Self::Pfctl => "/sbin/pfctl",
+            Self::Sysctl => "/usr/sbin/sysctl",
+            Self::Id => "/usr/bin/id",
+        }
+    }
+}
+
+/// What running a tool came to.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) enum Ran {
+    /// The program is not there.
+    Missing,
+    /// It ran.
+    Exited {
+        success: bool,
+        stdout: String,
+        stderr: String,
+    },
+}
+
+/// Runs a tool with optional standard input; injectable so the logic is
+/// tested without touching the host.
+pub(super) type Run<'a> = &'a dyn Fn(Tool, &[String], Option<&str>) -> Ran;
+
+/// Runs a real tool. Blocking: callers use `spawn_blocking`.
+pub(super) fn run_tool(tool: Tool, args: &[String], stdin: Option<&str>) -> Ran {
+    use std::io::Write;
+    use std::process::{Command, Stdio};
+
+    let mut command = Command::new(tool.path());
+    command
+        .args(args)
+        .stdin(if stdin.is_some() {
+            Stdio::piped()
+        } else {
+            Stdio::null()
+        })
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped());
+    let mut child = match command.spawn() {
+        Ok(child) => child,
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ran::Missing,
+        Err(err) => {
+            return Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: format!("could not run {}: {err}", tool.path()),
+            };
+        }
+    };
+    if let (Some(text), Some(mut pipe)) = (stdin, child.stdin.take()) {
+        // A write error shows up as the tool's own failure below.
+        let _ = pipe.write_all(text.as_bytes());
+    }
+    match child.wait_with_output() {
+        Ok(output) => Ran::Exited {
+            success: output.status.success(),
+            stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
+            stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
+        },
+        Err(err) => Ran::Exited {
+            success: false,
+            stdout: String::new(),
+            stderr: format!("could not run {}: {err}", tool.path()),
+        },
+    }
+}
+
+fn owned(words: &[&str]) -> Vec<String> {
+    words.iter().map(|word| word.to_string()).collect()
+}
+
+/// A name that is safe in a `pf` rule, an anchor path and a table name.
+fn safe_name(name: &str) -> bool {
+    !name.is_empty() && name.len() <= 32 && name.chars().all(|c| c.is_ascii_alphanumeric())
+}
+
+/// The anchor holding every rule of one interface.
+fn anchor_name(interface: &str) -> Result<String, String> {
+    if safe_name(interface) {
+        Ok(format!("com.apple/tsunagi-exit-{interface}"))
+    } else {
+        Err(format!(
+            "`{interface}` is not an interface name this agent can write a pf rule for"
+        ))
+    }
+}
+
+/// The default route as `route -n get default` reports it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) struct DefaultRoute {
+    pub interface: String,
+    pub gateway: Option<Ipv4Addr>,
+}
+
+/// Reads `interface:` and `gateway:` out of `route -n get default`.
+fn parse_default_route(text: &str) -> Option<DefaultRoute> {
+    let field = |name: &str| {
+        text.lines().find_map(|line| {
+            let (key, value) = line.split_once(':')?;
+            (key.trim() == name).then(|| value.trim().to_string())
+        })
+    };
+    let interface = field("interface").filter(|name| safe_name(name))?;
+    let gateway = field("gateway").and_then(|value| value.parse().ok());
+    Some(DefaultRoute { interface, gateway })
+}
+
+fn failure(stderr: &str, stdout: &str) -> String {
+    let text = if stderr.trim().is_empty() {
+        stdout.trim()
+    } else {
+        stderr.trim()
+    };
+    if text.is_empty() {
+        "the command failed without saying why".to_string()
+    } else {
+        text.to_string()
+    }
+}
+
+/// Runs a tool and turns a failure into a sentence a person can act on.
+fn run_ok(
+    run: Run<'_>,
+    tool: Tool,
+    args: &[String],
+    stdin: Option<&str>,
+) -> Result<String, String> {
+    match run(tool, args, stdin) {
+        Ran::Missing => Err(format!("{} not found", tool.path())),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => Ok(stdout),
+        Ran::Exited { stderr, stdout, .. } => Err(format!(
+            "{} {} failed: {}",
+            tool.path(),
+            args.join(" "),
+            failure(&stderr, &stdout)
+        )),
+    }
+}
+
+/// The route traffic leaves by now, which is the egress for translation and
+/// for the agent's own sockets. Read *before* the overlay routes are added.
+fn default_route(run: Run<'_>, overlay: &str) -> Result<DefaultRoute, String> {
+    let out = run_ok(run, Tool::Route, &owned(&["-n", "get", "default"]), None)
+        .map_err(|why| format!("cannot find the default route: {why}"))?;
+    let route = parse_default_route(&out)
+        .ok_or_else(|| "the host has no default route to send traffic out of".to_string())?;
+    if route.interface == overlay {
+        return Err(format!(
+            "the default route already goes through `{overlay}`"
+        ));
+    }
+    Ok(route)
+}
+
+/// Whether the kernel forwards IPv4. Read only.
+fn forwarding_state(run: Run<'_>) -> Option<bool> {
+    let out = run_ok(
+        run,
+        Tool::Sysctl,
+        &owned(&["-n", "net.inet.ip.forwarding"]),
+        None,
+    )
+    .ok()?;
+    match out.trim() {
+        "1" => Some(true),
+        "0" => Some(false),
+        _ => None,
+    }
+}
+
+/// The effective user id of this process, from `id -u`.
+fn agent_uid(run: Run<'_>) -> Result<u32, String> {
+    run_ok(run, Tool::Id, &owned(&["-u"]), None)
+        .ok()
+        .and_then(|out| out.trim().parse().ok())
+        .ok_or_else(|| "cannot tell which user this agent runs as".to_string())
+}
+
+/// What the using side adds to the rules.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ClientRules<'a> {
+    uid: u32,
+    /// Every overlay range: the agent's own traffic *to* those still goes
+    /// through the interface.
+    overlay: &'a [Ipv4Range],
+}
+
+/// The text of the anchor: translation first, filtering after, which is the
+/// order `pf` insists on.
+fn pf_rules(
+    interface: &str,
+    egress: &DefaultRoute,
+    offer: &[Ipv4Range],
+    client: Option<&ClientRules<'_>>,
+) -> String {
+    let out = &egress.interface;
+    let mut nat = Vec::new();
+    let mut filter = Vec::new();
+    for range in offer {
+        nat.push(format!("nat on {out} inet from {range} to any -> ({out})"));
+        filter.push(format!(
+            "pass in quick on {interface} inet from {range} to ! {range}"
+        ));
+    }
+    let mut tables = Vec::new();
+    if let Some(client) = client {
+        nat.push(format!(
+            "nat on {out} inet from ({interface}) to any -> ({out})"
+        ));
+        let target = if client.overlay.is_empty() {
+            "any".to_string()
+        } else {
+            let ranges: Vec<String> = client.overlay.iter().map(|r| r.to_string()).collect();
+            // A table, because `! { a, b }` expands to one rule per element
+            // in older `pf`, which matches everything.
+            tables.push(format!(
+                "table <tsunagi_overlay_{interface}> const {{ {} }}",
+                ranges.join(", ")
+            ));
+            format!("! <tsunagi_overlay_{interface}>")
+        };
+        let via = match egress.gateway {
+            Some(gateway) => format!("({out} {gateway})"),
+            None => out.clone(),
+        };
+        filter.push(format!(
+            "pass out quick on {interface} route-to {via} inet proto {{ tcp udp }} \
+             from any to {target} user {}",
+            client.uid
+        ));
+    }
+    tables
+        .into_iter()
+        .chain(nat)
+        .chain(filter)
+        .collect::<Vec<_>>()
+        .join("\n")
+        + "\n"
+}
+
+/// Turns `pf` on if nobody has, holding the reference `pfctl` hands out so it
+/// can be released when the agent is done. One reference at most.
+fn acquire_pf(run: Run<'_>, slot: &Mutex<Option<String>>) -> Result<(), String> {
+    if lock(slot).is_some() {
+        return Ok(());
+    }
+    match run(Tool::Pfctl, &owned(&["-E"]), None) {
+        Ran::Missing => Err(format!("{} not found", Tool::Pfctl.path())),
+        Ran::Exited {
+            success,
+            stdout,
+            stderr,
+        } => {
+            // `pfctl -E` reports `Token : <n>` on stderr, and also says so
+            // when pf is already enabled.
+            let token = stdout.lines().chain(stderr.lines()).find_map(parse_token);
+            if !success && token.is_none() {
+                return Err(format!("cannot enable pf: {}", failure(&stderr, &stdout)));
+            }
+            *lock(slot) = token;
+            Ok(())
+        }
+    }
+}
+
+fn parse_token(line: &str) -> Option<String> {
+    let (key, value) = line.split_once(':')?;
+    let value = value.trim();
+    (key.trim().eq_ignore_ascii_case("token")
+        && !value.is_empty()
+        && value.chars().all(|c| c.is_ascii_digit()))
+    .then(|| value.to_string())
+}
+
+fn release_pf(run: Run<'_>, slot: &Mutex<Option<String>>) {
+    if let Some(token) = lock(slot).take()
+        && let Err(why) = run_ok(run, Tool::Pfctl, &["-X".to_string(), token], None)
+    {
+        tracing::debug!(%why, "cannot release the pf reference");
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+fn flush_args(anchor: &str) -> Vec<String> {
+    owned(&["-q", "-a", anchor, "-F", "all"])
+}
+
+fn load_args(anchor: &str) -> Vec<String> {
+    owned(&["-q", "-a", anchor, "-f", "-"])
+}
+
+/// The two halves that out-rank the default route.
+const HALVES: [&str; 2] = ["0.0.0.0/1", "128.0.0.0/1"];
+/// The IPv6 halves, made unreachable.
+const HALVES_V6: [&str; 2] = ["::/1", "8000::/1"];
+
+fn route_add_args(half: &str, interface: &str) -> Vec<String> {
+    owned(&["-q", "-n", "add", "-net", half, "-interface", interface])
+}
+
+fn route_del_args(half: &str, interface: &str) -> Vec<String> {
+    owned(&["-q", "-n", "delete", "-net", half, "-interface", interface])
+}
+
+fn route_add_v6_args(half: &str) -> Vec<String> {
+    owned(&["-q", "-n", "add", "-inet6", "-net", half, "::1", "-reject"])
+}
+
+fn route_del_v6_args(half: &str) -> Vec<String> {
+    owned(&["-q", "-n", "delete", "-inet6", "-net", half, "::1"])
+}
+
+/// Takes away the overlay default routes. Best effort: what is not there is
+/// the state wanted.
+fn clear_routes(run: Run<'_>, interface: &str) {
+    for half in HALVES {
+        let _ = run(Tool::Route, &route_del_args(half, interface), None);
+    }
+    for half in HALVES_V6 {
+        let _ = run(Tool::Route, &route_del_v6_args(half), None);
+    }
+}
+
+fn set_routes(run: Run<'_>, interface: &str) -> Result<(), String> {
+    clear_routes(run, interface);
+    for half in HALVES {
+        if let Err(why) = run_ok(run, Tool::Route, &route_add_args(half, interface), None) {
+            clear_routes(run, interface);
+            return Err(why);
+        }
+    }
+    for half in HALVES_V6 {
+        // A host without IPv6 has nothing to block; not a failure.
+        if let Err(why) = run_ok(run, Tool::Route, &route_add_v6_args(half), None) {
+            tracing::debug!(%why, "cannot block IPv6 for the exit node");
+        }
+    }
+    Ok(())
+}
+
+/// Makes the host match the plan.
+pub(super) fn apply_plan(
+    run: Run<'_>,
+    token: &Mutex<Option<String>>,
+    plan: &ExitHostPlan,
+) -> ExitHostReport {
+    let forwarding = if plan.offer.is_empty() {
+        None
+    } else {
+        forwarding_state(run)
+    };
+    let outcome = apply_rules(run, token, plan);
+    let verdict = |ok: bool, why: &str| {
+        if ok {
+            RuleOutcome::Applied
+        } else {
+            RuleOutcome::Failed(why.to_string())
+        }
+    };
+    match outcome {
+        Ok(()) => ExitHostReport {
+            offer: plan
+                .offer
+                .iter()
+                .map(|range| (*range, RuleOutcome::Applied))
+                .collect(),
+            forwarding,
+            client: plan.client.then_some(RuleOutcome::Applied),
+        },
+        Err(why) => ExitHostReport {
+            offer: plan
+                .offer
+                .iter()
+                .map(|range| (*range, verdict(false, &why)))
+                .collect(),
+            forwarding,
+            client: plan.client.then(|| verdict(false, &why)),
+        },
+    }
+}
+
+fn apply_rules(
+    run: Run<'_>,
+    token: &Mutex<Option<String>>,
+    plan: &ExitHostPlan,
+) -> Result<(), String> {
+    let interface = plan.interface.as_str();
+    let anchor = anchor_name(interface)?;
+    let egress = default_route(run, interface)?;
+    let uid = if plan.client {
+        Some(agent_uid(run)?)
+    } else {
+        None
+    };
+    let client = uid.map(|uid| ClientRules {
+        uid,
+        overlay: &plan.overlay,
+    });
+    let rules = pf_rules(interface, &egress, &plan.offer, client.as_ref());
+
+    acquire_pf(run, token)?;
+    run_ok(run, Tool::Pfctl, &load_args(&anchor), Some(&rules))
+        .map_err(|why| format!("cannot load the pf rules: {why}"))?;
+    // The routes come after the rules that keep the agent's own traffic off
+    // them: the other order would loop for as long as it takes.
+    if plan.client {
+        set_routes(run, interface)?;
+    } else {
+        clear_routes(run, interface);
+    }
+    Ok(())
+}
+
+/// Removes everything this agent put there for the interface.
+pub(super) fn clear_all(run: Run<'_>, token: &Mutex<Option<String>>, interface: &str) {
+    clear_routes(run, interface);
+    if let Ok(anchor) = anchor_name(interface) {
+        // Flushing an anchor nothing was ever loaded into is not an error
+        // worth reporting.
+        if let Err(why) = run_ok(run, Tool::Pfctl, &flush_args(&anchor), None) {
+            tracing::debug!(%why, "cannot flush the exit node pf anchor");
+        }
+    }
+    release_pf(run, token);
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    const DEFAULT_ROUTE: &str = "   route to: default\n\
+destination: default\n\
+       mask: default\n\
+    gateway: 192.168.1.1\n\
+  interface: en0\n\
+      flags: <UP,GATEWAY,DONE,STATIC,PRCLONING,GLOBAL>\n";
+
+    fn egress() -> DefaultRoute {
+        DefaultRoute {
+            interface: "en0".into(),
+            gateway: Some(Ipv4Addr::new(192, 168, 1, 1)),
+        }
+    }
+
+    /// A scripted host that records every call.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        stdins: RefCell<Vec<String>>,
+        default_route: Option<String>,
+        forwarding: &'static str,
+        pf_missing: bool,
+        pf_load_fails: bool,
+        route_add_fails: bool,
+        pf_enable_output: &'static str,
+    }
+
+    impl Fake {
+        fn new() -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                stdins: RefCell::new(Vec::new()),
+                default_route: Some(DEFAULT_ROUTE.to_string()),
+                forwarding: "1\n",
+                pf_missing: false,
+                pf_load_fails: false,
+                route_add_fails: false,
+                pf_enable_output: "pf enabled\nToken : 4242\n",
+            }
+        }
+
+        fn run(&self, tool: Tool, args: &[String], stdin: Option<&str>) -> Ran {
+            let line = format!("{:?} {}", tool, args.join(" "));
+            self.calls.borrow_mut().push(line);
+            if let Some(text) = stdin {
+                self.stdins.borrow_mut().push(text.to_string());
+            }
+            let ok = |stdout: &str, stderr: &str| Ran::Exited {
+                success: true,
+                stdout: stdout.to_string(),
+                stderr: stderr.to_string(),
+            };
+            let fail = |why: &str| Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: why.to_string(),
+            };
+            match tool {
+                Tool::Id => ok("0\n", ""),
+                Tool::Sysctl => ok(self.forwarding, ""),
+                Tool::Route if args.first().map(String::as_str) == Some("-n") => {
+                    match &self.default_route {
+                        Some(text) => ok(text, ""),
+                        None => fail("route: writing to routing socket: not in table"),
+                    }
+                }
+                Tool::Route if args.contains(&"add".to_string()) && self.route_add_fails => {
+                    fail("route: writing to routing socket: File exists")
+                }
+                Tool::Route => ok("", ""),
+                Tool::Pfctl if self.pf_missing => Ran::Missing,
+                Tool::Pfctl if args == ["-E"] => ok("", self.pf_enable_output),
+                Tool::Pfctl if args.contains(&"-f".to_string()) && self.pf_load_fails => {
+                    fail("pfctl: Syntax error")
+                }
+                Tool::Pfctl => ok("", ""),
+            }
+        }
+
+        fn calls(&self) -> Vec<String> {
+            self.calls.borrow().clone()
+        }
+
+        fn position(&self, needle: &str) -> Option<usize> {
+            self.calls.borrow().iter().position(|c| c.contains(needle))
+        }
+
+        fn count(&self, needle: &str) -> usize {
+            self.calls
+                .borrow()
+                .iter()
+                .filter(|c| c.contains(needle))
+                .count()
+        }
+    }
+
+    fn plan(offer: bool, client: bool) -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "utun5".into(),
+            offer: if offer { vec![range()] } else { Vec::new() },
+            client,
+            overlay: vec![range()],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn the_default_route_is_read_from_route_output() {
+        assert_eq!(parse_default_route(DEFAULT_ROUTE), Some(egress()));
+        // A point-to-point default has no gateway line to use.
+        let ppp = "destination: default\n  interface: ppp0\n";
+        assert_eq!(
+            parse_default_route(ppp),
+            Some(DefaultRoute {
+                interface: "ppp0".into(),
+                gateway: None
+            })
+        );
+        assert_eq!(parse_default_route("route: not in table"), None);
+        // An interface name that is not plain letters and digits never ends
+        // up in a rule.
+        assert_eq!(parse_default_route("interface: en0; pass all\n"), None);
+    }
+
+    #[test]
+    fn the_anchor_is_under_com_apple_and_only_for_plain_names() {
+        assert_eq!(
+            anchor_name("utun5").unwrap(),
+            "com.apple/tsunagi-exit-utun5"
+        );
+        assert!(anchor_name("").is_err());
+        assert!(anchor_name("utun5 pass all").is_err());
+        assert!(anchor_name("../x").is_err());
+    }
+
+    #[test]
+    fn offering_translates_the_range_and_admits_it_in() {
+        let text = pf_rules("utun5", &egress(), &[range()], None);
+        assert_eq!(
+            text,
+            "nat on en0 inet from 10.13.37.0/24 to any -> (en0)\n\
+             pass in quick on utun5 inet from 10.13.37.0/24 to ! 10.13.37.0/24\n"
+        );
+        // Translation precedes filtering, as pf requires.
+        assert!(text.find("nat on").unwrap() < text.find("pass in").unwrap());
+    }
+
+    #[test]
+    fn using_sends_the_agents_own_traffic_back_out_the_physical_interface() {
+        let other: Ipv4Range = "10.99.0.0/16".parse().unwrap();
+        let overlay = [range(), other];
+        let client = ClientRules {
+            uid: 0,
+            overlay: &overlay,
+        };
+        let text = pf_rules("utun5", &egress(), &[], Some(&client));
+        assert!(
+            text.starts_with(
+                "table <tsunagi_overlay_utun5> const { 10.13.37.0/24, 10.99.0.0/16 }\n"
+            ),
+            "{text}"
+        );
+        assert!(
+            text.contains("nat on en0 inet from (utun5) to any -> (en0)\n"),
+            "{text}"
+        );
+        assert!(
+            text.contains(
+                "pass out quick on utun5 route-to (en0 192.168.1.1) inet proto { tcp udp } \
+                 from any to ! <tsunagi_overlay_utun5> user 0\n"
+            ),
+            "{text}"
+        );
+        // Only the agent's user, and only what is not the overlay.
+        assert!(!text.contains("from any to any"));
+
+        let direct = DefaultRoute {
+            interface: "ppp0".into(),
+            gateway: None,
+        };
+        let text = pf_rules(
+            "utun5",
+            &direct,
+            &[],
+            Some(&ClientRules {
+                uid: 501,
+                overlay: &[],
+            }),
+        );
+        assert!(text.contains("route-to ppp0 inet"), "{text}");
+        assert!(text.contains("to any user 501"), "{text}");
+        assert!(!text.contains("table"), "{text}");
+    }
+
+    #[test]
+    fn the_overlay_default_routes_are_halves_and_ipv6_is_rejected() {
+        assert_eq!(
+            route_add_args("0.0.0.0/1", "utun5").join(" "),
+            "-q -n add -net 0.0.0.0/1 -interface utun5"
+        );
+        assert_eq!(
+            route_del_args("128.0.0.0/1", "utun5").join(" "),
+            "-q -n delete -net 128.0.0.0/1 -interface utun5"
+        );
+        assert_eq!(
+            route_add_v6_args("8000::/1").join(" "),
+            "-q -n add -inet6 -net 8000::/1 ::1 -reject"
+        );
+        assert_eq!(
+            route_del_v6_args("::/1").join(" "),
+            "-q -n delete -inet6 -net ::/1 ::1"
+        );
+    }
+
+    #[test]
+    fn the_pf_token_is_found_in_either_stream() {
+        assert_eq!(parse_token("Token : 1234"), Some("1234".into()));
+        assert_eq!(parse_token("token: 9"), Some("9".into()));
+        assert_eq!(parse_token("pf enabled"), None);
+        assert_eq!(parse_token("Token : x; rm"), None);
+    }
+
+    #[test]
+    fn offering_loads_the_anchor_and_reads_but_never_writes_forwarding() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.offer, vec![(range(), RuleOutcome::Applied)]);
+        assert_eq!(report.forwarding, Some(true));
+        assert_eq!(report.client, None);
+        let calls = fake.calls();
+        assert!(
+            calls
+                .iter()
+                .any(|c| c == "Pfctl -q -a com.apple/tsunagi-exit-utun5 -f -"),
+            "{calls:?}"
+        );
+        assert_eq!(
+            fake.stdins.borrow()[0],
+            "nat on en0 inet from 10.13.37.0/24 to any -> (en0)\n\
+             pass in quick on utun5 inet from 10.13.37.0/24 to ! 10.13.37.0/24\n"
+        );
+        // sysctl is only ever asked, never told.
+        assert!(
+            calls
+                .iter()
+                .filter(|c| c.starts_with("Sysctl"))
+                .all(|c| c == "Sysctl -n net.inet.ip.forwarding"),
+            "{calls:?}"
+        );
+        // No client routes were added for an offer alone.
+        assert_eq!(fake.count(" add "), 0);
+        assert_eq!(token.lock().unwrap().as_deref(), Some("4242"));
+    }
+
+    #[test]
+    fn forwarding_off_is_reported_and_unreadable_is_none() {
+        let mut fake = Fake::new();
+        fake.forwarding = "0\n";
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.forwarding, Some(false));
+        fake.forwarding = "junk";
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.forwarding, None);
+    }
+
+    #[test]
+    fn the_client_rules_are_loaded_before_the_routes_that_need_them() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(false, true));
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        assert!(report.offer.is_empty());
+        let load = fake.position("-f -").unwrap();
+        let add = fake.position(" add -net 0.0.0.0/1").unwrap();
+        assert!(load < add, "{:?}", fake.calls());
+        assert!(fake.position("add -net 128.0.0.0/1").is_some());
+        assert!(fake.position("-inet6 -net ::/1 ::1 -reject").is_some());
+        assert!(fake.stdins.borrow()[0].contains("user 0"));
+        // The default route was read before any overlay route existed.
+        assert!(fake.position("-n get default").unwrap() < add);
+    }
+
+    #[test]
+    fn no_routes_are_added_when_the_rules_did_not_load() {
+        let mut fake = Fake::new();
+        fake.pf_load_fails = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, true));
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("cannot load the pf rules"), "{why}");
+        assert!(why.contains("Syntax error"), "{why}");
+        assert!(matches!(report.offer[0].1, RuleOutcome::Failed(_)));
+        // Nothing sent the host's traffic into a tunnel it could not exempt
+        // the agent from.
+        assert_eq!(fake.count(" add "), 0);
+    }
+
+    #[test]
+    fn a_failing_route_is_a_failure_and_leaves_nothing_behind() {
+        let mut fake = Fake::new();
+        fake.route_add_fails = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(false, true));
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("File exists"), "{why}");
+        // The cleanup after the failed add removed whatever did land.
+        let last_add = fake
+            .calls()
+            .iter()
+            .rposition(|c| c.contains(" add "))
+            .unwrap();
+        assert!(
+            fake.calls()[last_add..]
+                .iter()
+                .any(|c| c.contains("delete"))
+        );
+    }
+
+    #[test]
+    fn a_missing_pfctl_or_default_route_says_what_is_wrong() {
+        let mut fake = Fake::new();
+        fake.pf_missing = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("/sbin/pfctl not found"), "{why}");
+
+        let mut fake = Fake::new();
+        fake.default_route = None;
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("default route"), "{why}");
+        assert_eq!(fake.count("-f -"), 0);
+    }
+
+    #[test]
+    fn a_default_route_through_the_overlay_is_refused() {
+        let mut fake = Fake::new();
+        fake.default_route = Some("gateway: 10.13.37.1\ninterface: utun5\n".into());
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("already goes through"), "{why}");
+    }
+
+    #[test]
+    fn pf_is_enabled_once_and_released_once_on_clear() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let run = |t, a: &[String], s: Option<&str>| fake.run(t, a, s);
+        apply_plan(&run, &token, &plan(true, false));
+        apply_plan(&run, &token, &plan(true, true));
+        assert_eq!(fake.count("Pfctl -E"), 1, "{:?}", fake.calls());
+
+        clear_all(&run, &token, "utun5");
+        assert_eq!(fake.count("Pfctl -X 4242"), 1);
+        assert!(
+            fake.position("-a com.apple/tsunagi-exit-utun5 -F all")
+                .is_some()
+        );
+        assert!(
+            fake.position("delete -net 0.0.0.0/1 -interface utun5")
+                .is_some()
+        );
+        assert_eq!(*token.lock().unwrap(), None);
+        // A second clear has nothing to release and still succeeds.
+        clear_all(&run, &token, "utun5");
+        assert_eq!(fake.count("Pfctl -X"), 1);
+    }
+
+    #[test]
+    fn dropping_the_client_removes_its_routes() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let run = |t, a: &[String], s: Option<&str>| fake.run(t, a, s);
+        apply_plan(&run, &token, &plan(true, false));
+        assert!(fake.position("delete -net 0.0.0.0/1").is_some());
+        assert_eq!(fake.count(" add "), 0);
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/mod.rs

@@ -0,0 +1,369 @@
+//! Host rules that let a game's LAN discovery reach the overlay interface.
+//!
+//! Broadcast fanout happens at local TUN ingress (see [`super::broadcast`]),
+//! so a limited broadcast only reaches Tsunagi if the operating system routes
+//! it through the overlay interface, and a discovery *reply* only reaches the
+//! game if the host firewall lets it in. Left to the user that is two manual
+//! steps per host that nobody remembers, so when a network's broadcast is on
+//! the agent installs exactly these two things for the interface it owns:
+//!
+//! * a `255.255.255.255/32` route through the overlay interface, with the
+//!   overlay address as its preferred source, so the game's discovery packet
+//!   leaves with a source address the router recognises;
+//! * an inbound firewall allowance for **UDP only**, from the overlay range
+//!   only, on that interface only. Replies to a broadcast come from a
+//!   different address than the one the request was sent to, so a stateful
+//!   firewall never matches them to the request. Opening every protocol
+//!   would expose services such as SSH to every member of the network, which
+//!   is not what a game needs.
+//!
+//! # Why this is allowed here
+//!
+//! The architecture says plugins never touch routing or firewall settings,
+//! and nothing a peer announces becomes an OS setting. This is neither: it is
+//! the system level, which owns the interface, configuring that interface from
+//! values it computed itself (the interface name is derived locally, the
+//! address and range come from signed state it already verified). Every object
+//! is tagged `tsunagi:<interface>`, reversible, and removed when broadcast is
+//! turned off, the network leaves, or the agent stops. A crash can leave a
+//! firewall rule behind; the next apply for the same interface removes it.
+//!
+//! # Failure is contained
+//!
+//! Neither step can stop the agent. Each reports its own outcome, so a missing
+//! firewall tool still leaves the route in place and says what is missing.
+//!
+//! # Why a trait
+//!
+//! Linux, Windows and the tests do this very differently, while *what* to
+//! install is the same everywhere and lives in [`BroadcastRulesPlan`] and
+//! [`rules_summary`], tested on every platform.
+
+use std::net::Ipv4Addr;
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+mod exit;
+pub use exit::{
+    ExitHostPlan, ExitHostReport, ExitHostRules, MockExitRules, exit_tag, exit_tag_prefix,
+    offer_detail,
+};
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux_exit;
+#[cfg(all(test, feature = "tun-device", target_os = "linux"))]
+mod linux_exit_kernel;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub(in crate::overlay) mod linux_exit_mark;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use linux::LinuxHostRules;
+
+// Compiled everywhere so the scripts it builds are tested everywhere; only
+// the type that runs them is exported on Windows.
+mod windows;
+#[cfg_attr(not(target_os = "windows"), allow(dead_code))]
+mod windows_exit;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows::WindowsHostRules;
+
+// Compiled everywhere so its rule text and parsers are tested everywhere.
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos;
+#[cfg_attr(not(target_os = "macos"), allow(dead_code))]
+mod macos_exit;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos::MacosHostRules;
+
+/// What the host should be told for one interface.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastRulesPlan {
+    /// The overlay interface. Derived locally from the network id, never from
+    /// anything a peer said.
+    pub interface: String,
+    /// This agent's overlay address, the route's preferred source.
+    pub source: Ipv4Addr,
+    /// The overlay range inbound UDP is admitted from.
+    pub range: Ipv4Range,
+}
+
+impl BroadcastRulesPlan {
+    /// The tag every object carries, so it can be found again and nothing
+    /// else is ever touched.
+    pub fn tag(&self) -> String {
+        tag_for(&self.interface)
+    }
+}
+
+/// The tag for an interface's objects.
+pub fn tag_for(interface: &str) -> String {
+    format!("tsunagi:{interface}")
+}
+
+/// What happened to one of the two host objects.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum RuleOutcome {
+    /// It is in place.
+    Applied,
+    /// It is not, with a reason a person can act on.
+    Failed(String),
+}
+
+impl RuleOutcome {
+    /// Whether the object is in place.
+    pub fn is_applied(&self) -> bool {
+        matches!(self, Self::Applied)
+    }
+}
+
+/// What applying a plan achieved.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastRulesReport {
+    /// What was asked for.
+    pub plan: BroadcastRulesPlan,
+    /// The `255.255.255.255` route.
+    pub route: RuleOutcome,
+    /// The inbound firewall allowance.
+    pub firewall: RuleOutcome,
+}
+
+impl BroadcastRulesReport {
+    /// Whether both objects are in place.
+    pub fn is_ok(&self) -> bool {
+        self.route.is_applied() && self.firewall.is_applied()
+    }
+
+    /// One line for `status`.
+    pub fn summary(&self) -> String {
+        rules_summary(&self.route, &self.firewall)
+    }
+}
+
+/// The status line for a pair of outcomes.
+pub fn rules_summary(route: &RuleOutcome, firewall: &RuleOutcome) -> String {
+    match (route, firewall) {
+        (RuleOutcome::Applied, RuleOutcome::Applied) => "ok (route + firewall)".to_string(),
+        (route, firewall) => {
+            let part = |what: &str, outcome: &RuleOutcome| match outcome {
+                RuleOutcome::Applied => format!("{what} ok"),
+                RuleOutcome::Failed(reason) => format!("{what} failed: {reason}"),
+            };
+            format!(
+                "incomplete; {}; {}",
+                part("route", route),
+                part("firewall", firewall)
+            )
+        }
+    }
+}
+
+/// Installs and removes the host rules for an overlay interface.
+///
+/// Both calls are idempotent. Neither can fail as a whole: a step that cannot
+/// be done is reported in the [`BroadcastRulesReport`], because the caller's
+/// only useful response is to say so.
+pub trait BroadcastHostRules: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Makes the host match the plan, replacing whatever this agent put there
+    /// for the same interface before.
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport>;
+
+    /// Removes what this agent put there for the interface. Succeeds when
+    /// nothing is there: it runs on the shutdown path.
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()>;
+
+    /// The exit-node rules that go with these, when the platform has them.
+    ///
+    /// One object per host, because both halves need the same privileged
+    /// helpers; the default is that this platform has none, which is said
+    /// where it matters rather than pretended.
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        None
+    }
+}
+
+/// The platform has no implementation, which is said rather than pretended.
+#[derive(Debug, Default)]
+pub struct UnsupportedHostRules;
+
+impl BroadcastHostRules for UnsupportedHostRules {
+    fn name(&self) -> &str {
+        "unsupported"
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let reason = "host rules are not implemented on this platform; route \
+                          255.255.255.255 through the overlay interface and allow inbound \
+                          UDP from the overlay range yourself"
+                .to_string();
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route: RuleOutcome::Failed(reason.clone()),
+                firewall: RuleOutcome::Failed(reason),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, _interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// What a [`MockHostRules`] currently holds, per interface.
+pub type MockRulesState = std::collections::BTreeMap<String, BroadcastRulesPlan>;
+
+/// An in-memory host, for tests: nothing is touched.
+#[derive(Debug, Clone, Default)]
+pub struct MockHostRules {
+    exit: MockExitRules,
+    state: Arc<Mutex<MockRulesState>>,
+    calls: Arc<Mutex<Vec<String>>>,
+    firewall_failure: Arc<Mutex<Option<String>>>,
+}
+
+impl MockHostRules {
+    /// An empty host.
+    pub fn new() -> Self {
+        Self {
+            exit: MockExitRules::new(),
+            ..Self::default()
+        }
+    }
+
+    /// The exit-node half of this host, for inspecting and steering tests.
+    pub fn exit(&self) -> &MockExitRules {
+        &self.exit
+    }
+
+    /// Makes the firewall step fail with this reason from now on.
+    pub fn fail_firewall(&self, reason: impl Into<String>) {
+        *lock(&self.firewall_failure) = Some(reason.into());
+    }
+
+    /// What is installed now.
+    pub fn installed(&self) -> MockRulesState {
+        lock(&self.state).clone()
+    }
+
+    /// Every `apply:<interface>` / `clear:<interface>` call, in order.
+    pub fn calls(&self) -> Vec<String> {
+        lock(&self.calls).clone()
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+impl BroadcastHostRules for MockHostRules {
+    fn name(&self) -> &str {
+        "mock"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(&self.exit)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("apply:{}", plan.interface));
+            lock(&self.state).insert(plan.interface.clone(), plan.clone());
+            let firewall = match lock(&self.firewall_failure).clone() {
+                Some(reason) => RuleOutcome::Failed(reason),
+                None => RuleOutcome::Applied,
+            };
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route: RuleOutcome::Applied,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("clear:{interface}"));
+            lock(&self.state).remove(interface);
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn every_object_is_tagged_with_its_interface() {
+        assert_eq!(plan().tag(), "tsunagi:tsun0");
+        assert_eq!(tag_for("tsun7"), "tsunagi:tsun7");
+    }
+
+    #[test]
+    fn the_summary_says_which_half_is_missing() {
+        assert_eq!(
+            rules_summary(&RuleOutcome::Applied, &RuleOutcome::Applied),
+            "ok (route + firewall)"
+        );
+        let line = rules_summary(
+            &RuleOutcome::Applied,
+            &RuleOutcome::Failed("iptables not found".into()),
+        );
+        assert!(line.contains("route ok"), "{line}");
+        assert!(
+            line.contains("firewall failed: iptables not found"),
+            "{line}"
+        );
+        let report = BroadcastRulesReport {
+            plan: plan(),
+            route: RuleOutcome::Failed("denied".into()),
+            firewall: RuleOutcome::Applied,
+        };
+        assert!(!report.is_ok());
+        assert!(report.summary().contains("route failed: denied"));
+    }
+
+    #[tokio::test]
+    async fn the_mock_applies_replaces_and_clears_idempotently() {
+        let rules = MockHostRules::new();
+        let first = rules.apply(&plan()).await;
+        assert!(first.is_ok());
+        assert_eq!(rules.installed().len(), 1);
+
+        let mut moved = plan();
+        moved.source = Ipv4Addr::new(10, 13, 37, 7);
+        rules.apply(&moved).await;
+        assert_eq!(rules.installed()["tsun0"].source, moved.source);
+
+        rules.clear("tsun0").await;
+        rules.clear("tsun0").await;
+        assert!(rules.installed().is_empty());
+        assert_eq!(rules.calls().len(), 4);
+    }
+
+    #[tokio::test]
+    async fn an_unsupported_platform_reports_both_steps_failed() {
+        let report = UnsupportedHostRules.apply(&plan()).await;
+        assert!(!report.route.is_applied());
+        assert!(!report.firewall.is_applied());
+        UnsupportedHostRules.clear("tsun0").await;
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/windows.rs

@@ -0,0 +1,341 @@
+//! Broadcast host rules on Windows: a route and a Defender Firewall rule.
+//!
+//! Both are written with the `NetTCPIP` and `NetSecurity` PowerShell modules,
+//! run by absolute path from `%SystemRoot%` like the DNS publisher does: the
+//! in-process alternatives are `unsafe` IP Helper calls this crate forbids.
+//!
+//! * The route is `255.255.255.255/32` on the overlay interface, on-link, with
+//!   metric 1, written to the active store only so it cannot outlive a reboot.
+//!   Windows picks the interface's own address as the source of a packet sent
+//!   along it; with several networks on one interface that is whichever
+//!   address Windows treats as primary, which this cannot choose.
+//! * The firewall rule allows inbound UDP on that interface alias from the
+//!   overlay range only, in every profile. Wintun adapters are usually
+//!   classified as a public network, where inbound UDP is otherwise dropped.
+//!   Its display name is the tag `tsunagi:<interface>`, which is how it is
+//!   found again, replaced and removed.
+//!
+//! Every value in a script is one this agent produced: the interface name is
+//! derived from the network id and the range comes from signed state. They are
+//! still single-quoted and escaped, so nothing could be read as PowerShell
+//! rather than as data. Changing routes and firewall rules needs an elevated
+//! process, which is reported as its own cause.
+//!
+//! This module is compiled everywhere so its script building is tested
+//! everywhere. It could not be run against a real Windows host while it was
+//! written.
+
+#![cfg_attr(not(target_os = "windows"), allow(dead_code))]
+
+use crate::BoxFuture;
+
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules};
+use super::windows_exit;
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome, tag_for};
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// The step that makes the route match the plan.
+///
+/// A leftover route (from a crashed run, or for an old interface index) is
+/// removed first so the result is exactly one route.
+fn route_script(plan: &BroadcastRulesPlan) -> String {
+    let alias = ps_literal(&plan.interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         Get-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias $alias \
+         -PolicyStore ActiveStore -ErrorAction SilentlyContinue | \
+         Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias $alias \
+         -NextHop '0.0.0.0' -RouteMetric 1 -PolicyStore ActiveStore | Out-Null\n"
+    )
+}
+
+/// The step that makes the firewall hold exactly one rule for the interface.
+fn firewall_script(plan: &BroadcastRulesPlan) -> String {
+    let alias = ps_literal(&plan.interface);
+    let tag = ps_literal(&plan.tag());
+    let range = ps_literal(&plan.range.to_string());
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         $tag = {tag}\n\
+         Get-NetFirewallRule -DisplayName $tag -ErrorAction SilentlyContinue | \
+         Remove-NetFirewallRule -ErrorAction SilentlyContinue\n\
+         New-NetFirewallRule -DisplayName $tag -Direction Inbound -Action Allow -Protocol UDP \
+         -InterfaceAlias $alias -RemoteAddress {range} -Profile Any | Out-Null\n"
+    )
+}
+
+/// The script that removes both, ignoring what is not there.
+fn clear_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    let tag = ps_literal(&tag_for(interface));
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         Get-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias {alias} \
+         -PolicyStore ActiveStore | Remove-NetRoute -Confirm:$false\n\
+         Get-NetFirewallRule -DisplayName {tag} | Remove-NetFirewallRule\n\
+         exit 0\n"
+    )
+}
+
+/// Turns what PowerShell said into a cause a person can act on.
+fn classify(what: &str, text: &str) -> String {
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("access is denied")
+        || lower.contains("requires elevation")
+        || lower.contains("run as administrator")
+        || lower.contains("permissiondenied")
+    {
+        format!("{what}: Windows refused; run tsng from an elevated (administrator) process")
+    } else if lower.contains("is not recognized") || lower.contains("commandnotfound") {
+        format!("{what}: the NetTCPIP/NetSecurity PowerShell modules are not available: {text}")
+    } else if text.is_empty() {
+        format!("{what}: failed without saying why")
+    } else {
+        format!("{what}: {text}")
+    }
+}
+
+/// The absolute path to Windows PowerShell, so nothing on `PATH` can stand in
+/// for it.
+fn powershell_path() -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root)
+        .join("System32")
+        .join("WindowsPowerShell")
+        .join("v1.0")
+        .join("powershell.exe")
+}
+
+/// Runs a script and returns what it said on failure.
+fn run_powershell(script: &str) -> Result<(), String> {
+    run_powershell_output(script).map(|_| ())
+}
+
+/// Runs a script and returns what it printed, or what it said on failure.
+fn run_powershell_output(script: &str) -> Result<String, String> {
+    let output = std::process::Command::new(powershell_path())
+        .args([
+            "-NoProfile",
+            "-NonInteractive",
+            "-ExecutionPolicy",
+            "Bypass",
+            "-Command",
+            script,
+        ])
+        .output()
+        .map_err(|err| format!("could not run PowerShell: {err}"))?;
+    if output.status.success() {
+        return Ok(String::from_utf8_lossy(&output.stdout).into_owned());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = if stderr.trim().is_empty() {
+        String::from_utf8_lossy(&output.stdout).trim().to_string()
+    } else {
+        stderr.trim().to_string()
+    };
+    Err(text)
+}
+
+async fn step(what: &'static str, script: String) -> RuleOutcome {
+    match tokio::task::spawn_blocking(move || run_powershell(&script)).await {
+        Ok(Ok(())) => RuleOutcome::Applied,
+        Ok(Err(text)) => RuleOutcome::Failed(classify(what, &text)),
+        Err(err) => RuleOutcome::Failed(format!("{what}: the task failed: {err}")),
+    }
+}
+
+/// The Windows implementation, through PowerShell.
+#[derive(Debug, Default)]
+pub struct WindowsHostRules;
+
+impl WindowsHostRules {
+    /// Creates it. Nothing is run until [`BroadcastHostRules::apply`].
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+/// The IPv4 addresses a host name resolves to; empty when it does not.
+fn resolve_host(host: &str) -> Vec<std::net::Ipv4Addr> {
+    use std::net::{IpAddr, ToSocketAddrs};
+    (host, 443)
+        .to_socket_addrs()
+        .map(|addrs| {
+            addrs
+                .filter_map(|addr| match addr.ip() {
+                    IpAddr::V4(ip) => Some(ip),
+                    IpAddr::V6(_) => None,
+                })
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+impl ExitHostRules for WindowsHostRules {
+    fn needs_bypass(&self) -> bool {
+        true
+    }
+
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let job = plan.clone();
+            tokio::task::spawn_blocking(move || {
+                windows_exit::apply_plan(
+                    &|script| run_powershell_output(script),
+                    &resolve_host,
+                    &job,
+                )
+            })
+            .await
+            .unwrap_or_else(|err| {
+                let failed = RuleOutcome::Failed(format!("the task failed: {err}"));
+                ExitHostReport {
+                    offer: plan
+                        .offer
+                        .iter()
+                        .map(|range| (*range, failed.clone()))
+                        .collect(),
+                    forwarding: None,
+                    client: plan.client.then_some(failed),
+                }
+            })
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let _ = tokio::task::spawn_blocking(move || {
+                windows_exit::clear_all(&|script| run_powershell_output(script), &name)
+            })
+            .await;
+        })
+    }
+}
+
+impl BroadcastHostRules for WindowsHostRules {
+    fn name(&self) -> &str {
+        "windows-netsecurity"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = step("route", route_script(plan)).await;
+            let firewall = step("firewall", firewall_script(plan)).await;
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let RuleOutcome::Failed(reason) = step("clear", clear_script(interface)).await {
+                tracing::warn!(
+                    interface,
+                    "cannot remove the broadcast host rules: {reason}"
+                );
+            }
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::Ipv4Addr;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        assert_eq!(ps_literal("a'b"), "'a''b'");
+        let mut hostile = plan();
+        hostile.interface = "x'; Remove-Item C:\\ #".into();
+        let script = route_script(&hostile);
+        assert!(
+            script.contains("$alias = 'x''; Remove-Item C:\\ #'"),
+            "{script}"
+        );
+        assert!(firewall_script(&hostile).contains("'x''; Remove-Item C:\\ #'"));
+    }
+
+    #[test]
+    fn the_route_is_on_link_metric_one_in_the_active_store_and_replaces_its_own() {
+        let script = route_script(&plan());
+        assert!(script.contains("'255.255.255.255/32'"));
+        assert!(script.contains("-NextHop '0.0.0.0'"));
+        assert!(script.contains("-RouteMetric 1"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        assert!(
+            script.find("Remove-NetRoute").unwrap() < script.find("New-NetRoute").unwrap(),
+            "a stale route goes before the new one:\n{script}"
+        );
+    }
+
+    #[test]
+    fn the_firewall_rule_is_udp_only_from_the_overlay_range_on_this_interface() {
+        let script = firewall_script(&plan());
+        assert!(script.contains("-Protocol UDP"));
+        assert!(script.contains("-Direction Inbound -Action Allow"));
+        assert!(script.contains("-InterfaceAlias $alias"));
+        assert!(script.contains("-RemoteAddress '10.13.37.0/24'"));
+        assert!(script.contains("$tag = 'tsunagi:tsun0'"));
+        assert!(!script.contains("-Protocol Any") && !script.contains("-Protocol TCP"));
+        assert!(
+            script.find("Remove-NetFirewallRule").unwrap()
+                < script.find("New-NetFirewallRule").unwrap(),
+            "{script}"
+        );
+    }
+
+    #[test]
+    fn clearing_touches_only_our_route_and_our_tagged_rule() {
+        let script = clear_script("tsun0");
+        assert!(script.contains("-InterfaceAlias 'tsun0'"));
+        assert!(script.contains("Get-NetFirewallRule -DisplayName 'tsunagi:tsun0'"));
+        assert!(script.contains("exit 0"));
+    }
+
+    #[test]
+    fn a_refusal_names_elevation_and_a_missing_module_names_the_module() {
+        assert!(classify("route", "Access is denied").contains("elevated"));
+        assert!(
+            classify(
+                "firewall",
+                "The term 'New-NetFirewallRule' is not recognized"
+            )
+            .contains("modules are not available")
+        );
+        assert_eq!(classify("route", "boom"), "route: boom");
+        assert!(classify("route", "").contains("without saying why"));
+    }
+
+    #[test]
+    fn powershell_is_found_by_absolute_path() {
+        let path = powershell_path();
+        assert!(path.ends_with("powershell.exe"));
+        assert!(path.to_string_lossy().contains("System32"), "{path:?}");
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/windows_exit.rs

@@ -0,0 +1,603 @@
+//! Exit-node host rules on Windows: WinNAT through PowerShell.
+//!
+//! Compiled on every platform so the scripts and the plan logic are tested
+//! everywhere; only `WindowsHostRules` runs them, and only on Windows. It
+//! could not be run against a real Windows host while it was written.
+//!
+//! # Offering
+//!
+//! One NAT object per offered range, named with the usual tag
+//! `tsunagi-exit:<if>:<range>`:
+//!
+//! ```text
+//! New-NetNat -Name <tag> -InternalIPInterfaceAddressPrefix <range>
+//! ```
+//!
+//! WinNAT translates that range out of whichever interface the traffic leaves
+//! by, which is the Windows counterpart of the Linux masquerade. Windows
+//! forwards per interface, and the egress is not known here, so forwarding is
+//! only *read* for the overlay interface; turning it on for the interfaces
+//! involved is the owner's decision (`Set-NetIPInterface -Forwarding Enabled`).
+//! Some Windows versions allow a single NAT per host: a second offered range
+//! then reports its own failure.
+//!
+//! # Using
+//!
+//! Windows has no per-user or per-process routing, so the Linux scheme (an
+//! exemption for the agent's user) does not carry over. Instead the agent's
+//! own underlay traffic is exempted by *destination*:
+//!
+//! ```text
+//! New-NetRoute 0.0.0.0/1   on the overlay interface, metric 1
+//! New-NetRoute 128.0.0.0/1 on the overlay interface, metric 1
+//! New-NetRoute <ip>/32 via the physical default gateway, metric 7373
+//! ```
+//!
+//! The halves out-rank the default route without replacing it, so every more
+//! specific route (LAN, another VPN) keeps working. A `/32` per address the
+//! agent talks to directly (the verified direct paths of its connections, and
+//! the relays it uses, resolved when the rules are applied) keeps that traffic
+//! on the physical network instead of looping into the tunnel it carries. The
+//! metric 7373 marks those routes as ours, which is how stale ones are found
+//! and removed. Everything is written to the active store, so a reboot clears
+//! it. IPv6 is blackholed with `::/1` and `8000::/1` on the overlay interface,
+//! which has no IPv6: applications fall back to IPv4 after a delay, where
+//! Linux and macOS answer at once. The bypass follows the agent's paths as
+//! they change; a path that appears after the last apply is reached through
+//! the tunnel until the next one.
+//!
+//! The bypass routes are written before the halves, and removed after them.
+
+use super::RuleOutcome;
+use std::net::Ipv4Addr;
+
+use super::exit::{ExitHostPlan, ExitHostReport, exit_tag, exit_tag_prefix};
+use crate::state::Ipv4Range;
+
+/// The route metric that marks a bypass route as this agent's.
+const BYPASS_METRIC: u32 = 7373;
+
+/// Runs a PowerShell script and returns what it printed.
+pub(super) type RunScript<'a> = &'a dyn Fn(&str) -> Result<String, String>;
+
+/// Resolves a host name to its IPv4 addresses.
+pub(super) type Resolve<'a> = &'a dyn Fn(&str) -> Vec<Ipv4Addr>;
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// Creates the NAT of one range, replacing one this agent made before.
+fn nat_script(interface: &str, range: &Ipv4Range) -> String {
+    let name = ps_literal(&exit_tag(interface, range));
+    let prefix = ps_literal(&range.to_string());
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $name = {name}\n\
+         Get-NetNat -Name $name -ErrorAction SilentlyContinue | \
+         Remove-NetNat -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetNat -Name $name -InternalIPInterfaceAddressPrefix {prefix} | Out-Null\n"
+    )
+}
+
+/// Removes this interface's NAT objects except those named in `keep`.
+fn prune_script(interface: &str, keep: &[Ipv4Range]) -> String {
+    let prefix = ps_literal(&exit_tag_prefix(interface));
+    let keep: Vec<String> = keep
+        .iter()
+        .map(|range| ps_literal(&exit_tag(interface, range)))
+        .collect();
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         $prefix = {prefix}\n\
+         $keep = @({})\n\
+         Get-NetNat | Where-Object {{ $_.Name.StartsWith($prefix) -and \
+         ($keep -notcontains $_.Name) }} | Remove-NetNat -Confirm:$false\n\
+         exit 0\n",
+        keep.join(", ")
+    )
+}
+
+/// Prints `Enabled` or `Disabled` for the interface's IPv4 forwarding.
+fn forwarding_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         (Get-NetIPInterface -InterfaceAlias {alias} -AddressFamily IPv4).Forwarding\n"
+    )
+}
+
+/// Prints `<interface index> <next hop>` of the best default route that is
+/// not through the overlay interface.
+fn gateway_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         $best = Get-NetRoute -DestinationPrefix '0.0.0.0/0' -AddressFamily IPv4 \
+         -PolicyStore ActiveStore | Where-Object {{ $_.InterfaceAlias -ne $alias -and \
+         $_.NextHop -ne '0.0.0.0' }} | Sort-Object {{ $_.RouteMetric + \
+         (Get-NetIPInterface -InterfaceIndex $_.InterfaceIndex -AddressFamily IPv4).InterfaceMetric }} | \
+         Select-Object -First 1\n\
+         if (-not $best) {{ throw 'the host has no default gateway outside the overlay' }}\n\
+         \"$($best.InterfaceIndex) $($best.NextHop)\"\n"
+    )
+}
+
+/// The default gateway as the script above printed it.
+fn parse_gateway(text: &str) -> Option<(u32, Ipv4Addr)> {
+    let mut words = text.split_whitespace();
+    let index = words.next()?.parse().ok()?;
+    let hop = words.next()?.parse().ok()?;
+    words.next().is_none().then_some((index, hop))
+}
+
+/// Replaces the bypass routes with one per address, through the gateway.
+fn bypass_script(index: u32, gateway: Ipv4Addr, addresses: &[Ipv4Addr]) -> String {
+    let list: Vec<String> = addresses
+        .iter()
+        .map(|address| ps_literal(&address.to_string()))
+        .collect();
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         Get-NetRoute -AddressFamily IPv4 -PolicyStore ActiveStore -ErrorAction SilentlyContinue | \
+         Where-Object {{ $_.RouteMetric -eq {BYPASS_METRIC} -and $_.DestinationPrefix.EndsWith('/32') }} | \
+         Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         foreach ($ip in @({})) {{\n\
+         New-NetRoute -DestinationPrefix \"$ip/32\" -InterfaceIndex {index} -NextHop '{gateway}' \
+         -RouteMetric {BYPASS_METRIC} -PolicyStore ActiveStore -ErrorAction SilentlyContinue | Out-Null\n\
+         }}\n",
+        list.join(", ")
+    )
+}
+
+/// Sends everything else through the overlay interface.
+fn halves_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         foreach ($p in '0.0.0.0/1', '128.0.0.0/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore \
+         -ErrorAction SilentlyContinue | Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -NextHop '0.0.0.0' \
+         -RouteMetric 1 -PolicyStore ActiveStore | Out-Null\n\
+         }}\n\
+         foreach ($p in '::/1', '8000::/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore \
+         -ErrorAction SilentlyContinue | Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -NextHop '::' \
+         -RouteMetric 1 -PolicyStore ActiveStore -ErrorAction SilentlyContinue | Out-Null\n\
+         }}\n"
+    )
+}
+
+/// Removes the halves and every bypass route. What is not there is the state
+/// wanted.
+fn remove_client_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         $alias = {alias}\n\
+         foreach ($p in '0.0.0.0/1', '128.0.0.0/1', '::/1', '8000::/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore | \
+         Remove-NetRoute -Confirm:$false\n\
+         }}\n\
+         Get-NetRoute -AddressFamily IPv4 -PolicyStore ActiveStore | \
+         Where-Object {{ $_.RouteMetric -eq {BYPASS_METRIC} -and $_.DestinationPrefix.EndsWith('/32') }} | \
+         Remove-NetRoute -Confirm:$false\n\
+         exit 0\n"
+    )
+}
+
+/// Every address to keep direct: the ones given and what the names resolve
+/// to, sorted and without duplicates.
+fn bypass_addresses(
+    resolve: Resolve<'_>,
+    addresses: &[Ipv4Addr],
+    hosts: &[String],
+) -> Vec<Ipv4Addr> {
+    let mut all: std::collections::BTreeSet<Ipv4Addr> = addresses.iter().copied().collect();
+    for host in hosts {
+        all.extend(resolve(host));
+    }
+    all.into_iter()
+        .filter(|a| !a.is_loopback() && !a.is_unspecified() && !a.is_multicast())
+        .collect()
+}
+
+fn apply_client(
+    run: RunScript<'_>,
+    resolve: Resolve<'_>,
+    plan: &ExitHostPlan,
+) -> Result<(), String> {
+    let interface = plan.interface.as_str();
+    let gateway = run(&gateway_script(interface))?;
+    let (index, hop) = parse_gateway(&gateway)
+        .ok_or_else(|| format!("cannot read the default gateway from `{}`", gateway.trim()))?;
+    let addresses = bypass_addresses(resolve, &plan.bypass, &plan.bypass_hosts);
+    run(&bypass_script(index, hop, &addresses))?;
+    if let Err(why) = run(&halves_script(interface)) {
+        // Do not leave a half-installed set: bypass routes alone are harmless
+        // but the next apply starts clean.
+        let _ = run(&remove_client_script(interface));
+        return Err(why);
+    }
+    Ok(())
+}
+
+fn parse_forwarding(text: &str) -> Option<bool> {
+    match text.trim().to_ascii_lowercase().as_str() {
+        "enabled" | "1" => Some(true),
+        "disabled" | "0" => Some(false),
+        _ => None,
+    }
+}
+
+/// Makes the host match the plan.
+pub(super) fn apply_plan(
+    run: RunScript<'_>,
+    resolve: Resolve<'_>,
+    plan: &ExitHostPlan,
+) -> ExitHostReport {
+    let interface = plan.interface.as_str();
+    // What an earlier plan left and this one no longer wants.
+    let pruned = run(&prune_script(interface, &plan.offer));
+    let offer = plan
+        .offer
+        .iter()
+        .map(|range| {
+            let outcome = match &pruned {
+                Err(why) => RuleOutcome::Failed(why.clone()),
+                Ok(_) => match run(&nat_script(interface, range)) {
+                    Ok(_) => RuleOutcome::Applied,
+                    Err(why) => RuleOutcome::Failed(why),
+                },
+            };
+            (*range, outcome)
+        })
+        .collect();
+    let client = if plan.client {
+        Some(match apply_client(run, resolve, plan) {
+            Ok(()) => RuleOutcome::Applied,
+            Err(why) => RuleOutcome::Failed(why),
+        })
+    } else {
+        // Whatever an earlier plan left must not stay in force.
+        let _ = run(&remove_client_script(interface));
+        None
+    };
+    let forwarding = if plan.offer.is_empty() {
+        None
+    } else {
+        run(&forwarding_script(interface))
+            .ok()
+            .and_then(|text| parse_forwarding(&text))
+    };
+    ExitHostReport {
+        offer,
+        forwarding,
+        client,
+    }
+}
+
+/// Removes every NAT object of this agent's for the interface.
+pub(super) fn clear_all(run: RunScript<'_>, interface: &str) {
+    if let Err(why) = run(&prune_script(interface, &[])) {
+        tracing::warn!(interface, "cannot remove the exit node NAT: {why}");
+    }
+    if let Err(why) = run(&remove_client_script(interface)) {
+        tracing::warn!(interface, "cannot remove the exit node routes: {why}");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    fn plan(client: bool) -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: vec![range()],
+            client,
+            overlay: vec![range()],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn the_nat_is_named_with_the_tag_and_replaces_its_own() {
+        let script = nat_script("tsun0", &range());
+        assert!(
+            script.contains("$name = 'tsunagi-exit:tsun0:10.13.37.0/24'"),
+            "{script}"
+        );
+        assert!(script.contains("-InternalIPInterfaceAddressPrefix '10.13.37.0/24'"));
+        assert!(
+            script.find("Remove-NetNat").unwrap() < script.find("New-NetNat").unwrap(),
+            "{script}"
+        );
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        let script = nat_script("x'; Remove-Item C:\\ #", &range());
+        assert!(script.contains("'tsunagi-exit:x''; Remove-Item C:\\ #:10.13.37.0/24'"));
+        assert!(prune_script("x'y", &[]).contains("'tsunagi-exit:x''y:'"));
+        assert!(forwarding_script("x'y").contains("'x''y'"));
+    }
+
+    #[test]
+    fn pruning_keeps_wanted_ranges_and_only_selects_this_interfaces_nats() {
+        let script = prune_script("tsun0", &[range()]);
+        assert!(script.contains("$prefix = 'tsunagi-exit:tsun0:'"));
+        assert!(script.contains("$keep = @('tsunagi-exit:tsun0:10.13.37.0/24')"));
+        assert!(script.contains("StartsWith($prefix)"));
+        assert!(script.contains("exit 0"));
+        assert!(prune_script("tsun0", &[]).contains("$keep = @()"));
+    }
+
+    #[test]
+    fn forwarding_text_is_read_never_written() {
+        assert_eq!(parse_forwarding("Enabled\r\n"), Some(true));
+        assert_eq!(parse_forwarding("Disabled"), Some(false));
+        assert_eq!(parse_forwarding("what"), None);
+        let script = forwarding_script("tsun0");
+        assert!(script.contains("Get-NetIPInterface"));
+        assert!(!script.contains("Set-NetIPInterface"));
+    }
+
+    #[test]
+    fn offering_prunes_then_creates_and_reads_forwarding() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(if script.contains("Get-NetIPInterface") {
+                "Disabled\n".to_string()
+            } else {
+                String::new()
+            })
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        assert_eq!(report.offer, vec![(range(), RuleOutcome::Applied)]);
+        assert_eq!(report.forwarding, Some(false));
+        assert_eq!(report.client, None);
+        let calls = calls.borrow();
+        assert!(calls[0].contains("StartsWith"), "{calls:?}");
+        assert!(calls[1].contains("New-NetNat"), "{calls:?}");
+        // Without a client, the routes of an earlier plan are taken away.
+        assert!(calls[2].contains("RouteMetric -eq 7373"), "{calls:?}");
+        assert!(calls[3].contains("Get-NetIPInterface"), "{calls:?}");
+    }
+
+    #[test]
+    fn a_failing_nat_is_that_ranges_failure() {
+        let run = |script: &str| -> Result<String, String> {
+            if script.contains("New-NetNat") {
+                Err("A NAT with this prefix already exists".into())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("already exists"), "{why}");
+    }
+
+    fn no_names(_: &str) -> Vec<Ipv4Addr> {
+        Vec::new()
+    }
+
+    fn client_plan() -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: Vec::new(),
+            client: true,
+            overlay: vec![range()],
+            bypass: vec![Ipv4Addr::new(138, 201, 61, 182)],
+            bypass_hosts: vec!["relay.example".into()],
+        }
+    }
+
+    #[test]
+    fn the_gateway_is_read_from_the_script_output() {
+        assert_eq!(
+            parse_gateway("12 192.168.1.1\r\n"),
+            Some((12, Ipv4Addr::new(192, 168, 1, 1)))
+        );
+        assert_eq!(parse_gateway(""), None);
+        assert_eq!(parse_gateway("12"), None);
+        assert_eq!(parse_gateway("12 gw"), None);
+        assert_eq!(parse_gateway("12 1.1.1.1 extra"), None);
+        let script = gateway_script("tsun0");
+        assert!(script.contains("-ne $alias"));
+        assert!(script.contains("-ne '0.0.0.0'"));
+    }
+
+    #[test]
+    fn bypass_routes_are_marked_host_routes_through_the_physical_gateway() {
+        let script = bypass_script(
+            12,
+            Ipv4Addr::new(192, 168, 1, 1),
+            &[Ipv4Addr::new(1, 2, 3, 4), Ipv4Addr::new(5, 6, 7, 8)],
+        );
+        assert!(script.contains("@('1.2.3.4', '5.6.7.8')"), "{script}");
+        assert!(script.contains("-InterfaceIndex 12 -NextHop '192.168.1.1'"));
+        assert!(script.contains("-RouteMetric 7373"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        // Stale marked routes go first, and only /32 ones with our metric.
+        assert!(script.find("Remove-NetRoute").unwrap() < script.find("New-NetRoute").unwrap());
+        assert!(script.contains("RouteMetric -eq 7373"));
+        assert!(script.contains("EndsWith('/32')"));
+    }
+
+    #[test]
+    fn the_halves_go_through_the_overlay_and_ipv6_is_blackholed() {
+        let script = halves_script("tsun0");
+        assert!(script.contains("'0.0.0.0/1', '128.0.0.0/1'"));
+        assert!(script.contains("-InterfaceAlias $alias -NextHop '0.0.0.0'"));
+        assert!(script.contains("'::/1', '8000::/1'"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        assert!(
+            !script.contains("0.0.0.0/0"),
+            "the default route is left alone"
+        );
+        assert!(halves_script("x'y").contains("'x''y'"));
+    }
+
+    #[test]
+    fn addresses_are_resolved_merged_and_cleaned() {
+        let resolve = |host: &str| -> Vec<Ipv4Addr> {
+            match host {
+                "relay.example" => vec![Ipv4Addr::new(9, 9, 9, 9), Ipv4Addr::new(1, 2, 3, 4)],
+                _ => Vec::new(),
+            }
+        };
+        let all = bypass_addresses(
+            &resolve,
+            &[
+                Ipv4Addr::new(1, 2, 3, 4),
+                Ipv4Addr::LOCALHOST,
+                Ipv4Addr::UNSPECIFIED,
+            ],
+            &["relay.example".into(), "gone.example".into()],
+        );
+        assert_eq!(
+            all,
+            vec![Ipv4Addr::new(1, 2, 3, 4), Ipv4Addr::new(9, 9, 9, 9)]
+        );
+    }
+
+    #[test]
+    fn the_bypass_is_written_before_the_halves_that_need_it() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(if script.contains("Select-Object -First 1") {
+                "12 192.168.1.1\n".to_string()
+            } else {
+                String::new()
+            })
+        };
+        let resolve = |_: &str| vec![Ipv4Addr::new(9, 9, 9, 9)];
+        let report = apply_plan(&run, &resolve, &client_plan());
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        let calls = calls.borrow();
+        let at = |needle: &str| calls.iter().position(|c| c.contains(needle)).unwrap();
+        assert!(at("Select-Object -First 1") < at("-RouteMetric 7373"));
+        assert!(at("-RouteMetric 7373") < at("'0.0.0.0/1', '128.0.0.0/1'"));
+        let bypass = &calls[at("-RouteMetric 7373")];
+        assert!(
+            bypass.contains("'9.9.9.9'") && bypass.contains("'138.201.61.182'"),
+            "{bypass}"
+        );
+    }
+
+    #[test]
+    fn nothing_goes_through_the_overlay_when_the_bypass_cannot_be_written() {
+        let run = |script: &str| -> Result<String, String> {
+            if script.contains("Select-Object -First 1") {
+                Ok("12 192.168.1.1".to_string())
+            } else if script.contains("-RouteMetric 7373") && script.contains("foreach ($ip") {
+                Err("Access is denied".to_string())
+            } else {
+                assert!(!script.contains("-NextHop '0.0.0.0'"), "{script}");
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        assert_eq!(
+            report.client,
+            Some(RuleOutcome::Failed("Access is denied".into()))
+        );
+    }
+
+    #[test]
+    fn a_missing_gateway_is_a_failure_that_installs_no_route() {
+        let run = |script: &str| -> Result<String, String> {
+            assert!(!script.contains("New-NetRoute"), "{script}");
+            if script.contains("Select-Object -First 1") {
+                Err("the host has no default gateway outside the overlay".into())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("no default gateway"), "{why}");
+    }
+
+    #[test]
+    fn a_failed_halves_step_removes_what_was_written() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            if script.contains("Select-Object -First 1") {
+                Ok("12 192.168.1.1".to_string())
+            } else if script.contains("'128.0.0.0/1'") && script.contains("New-NetRoute") {
+                Err("boom".to_string())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        assert!(matches!(report.client, Some(RuleOutcome::Failed(_))));
+        assert!(calls.borrow().last().unwrap().contains("Remove-NetRoute"));
+    }
+
+    #[test]
+    fn dropping_the_client_removes_its_routes_and_clearing_does_too() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(String::new())
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        assert_eq!(report.client, None);
+        assert!(
+            calls
+                .borrow()
+                .iter()
+                .any(|c| c.contains("RouteMetric -eq 7373"))
+        );
+        let script = remove_client_script("tsun0");
+        assert!(script.contains("'0.0.0.0/1', '128.0.0.0/1', '::/1', '8000::/1'"));
+        assert!(script.contains("exit 0"));
+        // Only routes on our interface, or marked ours, are touched.
+        assert!(script.contains("-InterfaceAlias $alias"));
+        calls.borrow_mut().clear();
+        clear_all(&run, "tsun0");
+        assert!(
+            calls
+                .borrow()
+                .iter()
+                .any(|c| c.contains("RouteMetric -eq 7373"))
+        );
+    }
+
+    #[test]
+    fn clearing_prunes_everything_for_the_interface() {
+        let seen = RefCell::new(Vec::<String>::new());
+        let run = |script: &str| -> Result<String, String> {
+            seen.borrow_mut().push(script.to_string());
+            Ok(String::new())
+        };
+        clear_all(&run, "tsun0");
+        let seen = seen.borrow();
+        assert!(seen[0].contains("$keep = @()"));
+        assert!(seen[0].contains("'tsunagi-exit:tsun0:'"));
+    }
+}

crates/tsunagi-core/src/overlay/mod.rs

@@ -0,0 +1,80 @@
+//! The overlay interface: the one network interface an agent owns.
+//!
+//! One agent, one interface. It belongs to the system level rather than to
+//! any protocol, and that is the whole point: a packet leaving it is routed
+//! to whichever peer owns its destination address, over whichever protocol
+//! currently has a link to that peer. Several protocols can be live at once
+//! without arguing over who holds the address, because neither of them
+//! holds it — the agent does.
+//!
+//! What lives here:
+//!
+//! * [`provision`] creates the interface and configures it, and removes it
+//!   again. Platform mechanics behind one decision function.
+//! * [`hostrules`] installs the route and firewall allowance that let a
+//!   game's LAN discovery reach the interface, while broadcast is on.
+//! * [`tun`] is the packet interface itself, real or in memory.
+//! * [`packet`] reads just enough of an IP header to route by it.
+//! * [`config`] names interfaces and describes addresses.
+//!
+//! What does not live here: encryption, peer discovery, and how a packet
+//! actually reaches another machine. Those belong to a protocol plugin,
+//! which this level hands packets to and takes packets from.
+
+/// Why the overlay interface cannot be brought into the state asked for.
+///
+/// Its own error rather than the plugin one it used to borrow: this level
+/// owns the interface now, and a plugin failing is a different event from
+/// the interface failing.
+#[derive(Debug, thiserror::Error)]
+#[non_exhaustive]
+pub enum OverlayError {
+    /// The interface cannot be created or configured right now, with the
+    /// reason and, where there is one, what to do about it.
+    #[error("overlay unavailable: {0}")]
+    Unavailable(String),
+    /// Anything else.
+    #[error("{0}")]
+    Other(String),
+}
+
+pub mod broadcast;
+pub mod config;
+pub mod exit;
+pub mod hostrules;
+pub mod interface;
+pub mod packet;
+pub mod provision;
+pub mod router;
+pub mod tun;
+
+pub use config::{Cidr, DEFAULT_INTERFACE_PREFIX, MAX_INTERFACE_NAME_LEN, interface_name};
+pub use exit::{ExitRulesReport, RuleSetReport};
+pub use hostrules::{
+    BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, ExitHostPlan, ExitHostReport,
+    ExitHostRules, MockExitRules, MockHostRules, RuleOutcome, UnsupportedHostRules,
+};
+pub use interface::{Counters, Interface, PacketCarrier, Rejected};
+pub use packet::IpHeader;
+pub use provision::{
+    Changes, InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, ManagedTunFactory,
+    MockHost, MockProvisioner, Privilege, Provisioned, UnsupportedProvisioner, plan_changes,
+    probe_net_admin,
+};
+pub use router::{ExitPolicy, NetworkRoutes, Route, RouteError, RoutingTable};
+pub use tun::{MemoryTun, MemoryTunFactory, TunDevice, TunFactory, TunRequest, address_is_local};
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use hostrules::LinuxHostRules;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use hostrules::MacosHostRules;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use hostrules::WindowsHostRules;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use provision::NetlinkProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use provision::UtunProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use provision::WintunProvisioner;

crates/tsunagi-core/src/overlay/packet.rs

@@ -65,6 +65,22 @@ impl IpHeader {
             IpHeader::V4 { .. } => None,
         }
     }
+
+    /// The destination, whichever family it is.
+    pub fn destination(&self) -> std::net::IpAddr {
+        match self {
+            IpHeader::V4 { destination, .. } => std::net::IpAddr::V4(*destination),
+            IpHeader::V6 { destination, .. } => std::net::IpAddr::V6(*destination),
+        }
+    }
+
+    /// The source, whichever family it is.
+    pub fn source(&self) -> std::net::IpAddr {
+        match self {
+            IpHeader::V4 { source, .. } => std::net::IpAddr::V4(*source),
+            IpHeader::V6 { source, .. } => std::net::IpAddr::V6(*source),
+        }
+    }
 }
 
 #[cfg(test)]

crates/tsunagi-core/src/overlay/provision/factory.rs

@@ -0,0 +1,271 @@
+//! The adapter between the plugin's view of a packet interface and the
+//! host-management view.
+//!
+//! The plugin asks a [`TunFactory`] for a device and knows nothing else. This
+//! factory answers by reconciling the host — creating the interface, fixing
+//! up whatever an earlier run left behind, assigning the addresses — and
+//! handing back the device that came out of it.
+
+use std::sync::Arc;
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::hostrules::BroadcastHostRules;
+use super::super::tun::{TunDevice, TunFactory, TunRequest};
+use super::{InterfacePlan, InterfaceProvisioner};
+
+/// Turns a [`TunRequest`] into the plan for a host interface.
+fn plan_for(request: &TunRequest) -> Result<InterfacePlan, OverlayError> {
+    // Every address, because the plan is exhaustive: the provisioner adds
+    // what is missing and removes what is not in it. Passing one network's
+    // address would take every other network's off the host.
+    Ok(InterfacePlan::new(
+        request.name.clone(),
+        request.mtu,
+        request.addresses.clone(),
+    ))
+}
+
+/// A [`TunFactory`] backed by an [`InterfaceProvisioner`].
+#[derive(Debug)]
+pub struct ManagedTunFactory {
+    provisioner: Arc<dyn InterfaceProvisioner>,
+    host_rules: Option<Arc<dyn BroadcastHostRules>>,
+}
+
+impl ManagedTunFactory {
+    /// Wraps a provisioner.
+    pub fn new(provisioner: Arc<dyn InterfaceProvisioner>) -> Self {
+        Self {
+            provisioner,
+            host_rules: None,
+        }
+    }
+
+    /// Lets the agent install the broadcast route and firewall allowance for
+    /// the interface this factory manages.
+    pub fn with_host_rules(mut self, rules: Arc<dyn BroadcastHostRules>) -> Self {
+        self.host_rules = Some(rules);
+        self
+    }
+
+    /// The provisioner underneath.
+    pub fn provisioner(&self) -> &Arc<dyn InterfaceProvisioner> {
+        &self.provisioner
+    }
+}
+
+impl TunFactory for ManagedTunFactory {
+    fn name(&self) -> &str {
+        self.provisioner.name()
+    }
+
+    /// It creates the interface on the host and holds it open.
+    fn on_host(&self) -> bool {
+        true
+    }
+
+    fn host_rules(&self) -> Option<Arc<dyn BroadcastHostRules>> {
+        self.host_rules.clone()
+    }
+
+    fn create<'a>(
+        &'a self,
+        request: TunRequest,
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>> {
+        Box::pin(async move {
+            let plan = plan_for(&request)?;
+            let provisioned = self.provisioner.reconcile(&plan).await?;
+            tracing::info!(
+                interface = %plan.name,
+                changes = %provisioned.changes.summary(),
+                "overlay interface reconciled"
+            );
+            provisioned.device.ok_or_else(|| {
+                // Reaching here would mean the interface already existed and
+                // was held open by us, which cannot be true on the path that
+                // creates a device.
+                OverlayError::Other(format!(
+                    "interface `{}` was reconciled but no device came back",
+                    plan.name
+                ))
+            })
+        })
+    }
+
+    fn reconfigure<'a>(&'a self, request: TunRequest) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            let plan = plan_for(&request)?;
+            let provisioned = self.provisioner.reconcile(&plan).await?;
+            if !provisioned.changes.is_empty() {
+                tracing::info!(
+                    interface = %plan.name,
+                    changes = %provisioned.changes.summary(),
+                    "overlay interface updated"
+                );
+            }
+            Ok(())
+        })
+    }
+
+    fn destroy<'a>(&'a self, name: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            match self.provisioner.remove(name).await {
+                Ok(()) => tracing::info!(interface = %name, "overlay interface removed"),
+                Err(err) => {
+                    tracing::warn!(interface = %name, %err, "cannot remove the overlay interface")
+                }
+            }
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::net::Ipv4Addr;
+
+    use super::super::super::config::Cidr;
+    use super::super::{LinkKind, MockHost, MockProvisioner};
+    use super::*;
+
+    fn request(v4: Option<Ipv4Addr>) -> TunRequest {
+        addressed(v4.into_iter().collect())
+    }
+
+    fn addressed(v4: Vec<Ipv4Addr>) -> TunRequest {
+        TunRequest {
+            name: "tsunfactory".into(),
+            addresses: v4
+                .into_iter()
+                .map(|address| Cidr::new(address.into(), 24).unwrap())
+                .collect(),
+            mtu: 1280,
+        }
+    }
+
+    #[tokio::test]
+    async fn creating_a_device_provisions_the_host_and_returns_it() {
+        let provisioner = Arc::new(MockProvisioner::default());
+        let factory = ManagedTunFactory::new(provisioner.clone());
+
+        let device = factory
+            .create(request(Some(Ipv4Addr::new(10, 13, 37, 69))))
+            .await
+            .unwrap();
+        assert_eq!(device.name(), "tsunfactory");
+        assert_eq!(device.mtu(), 1280);
+
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        assert_eq!(state.kind, LinkKind::Tun);
+        assert_eq!(state.mtu, 1280);
+        assert_eq!(state.addresses.len(), 1, "the overlay address is assigned");
+    }
+
+    #[tokio::test]
+    async fn a_reallocated_address_is_applied_to_the_live_interface() {
+        let provisioner = Arc::new(MockProvisioner::default());
+        let factory = ManagedTunFactory::new(provisioner.clone());
+        factory
+            .create(request(Some(Ipv4Addr::new(10, 13, 37, 69))))
+            .await
+            .unwrap();
+
+        factory
+            .reconfigure(request(Some(Ipv4Addr::new(10, 13, 37, 70))))
+            .await
+            .unwrap();
+
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        assert!(state.attached, "the interface was not recreated");
+        let addresses: Vec<String> = state
+            .addresses
+            .iter()
+            .map(|entry| entry.to_string())
+            .collect();
+        assert!(
+            addresses.contains(&"10.13.37.70/24".to_string()),
+            "{addresses:?}"
+        );
+        assert!(
+            !addresses.contains(&"10.13.37.69/24".to_string()),
+            "{addresses:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn destroying_takes_the_interface_off_the_host() {
+        let provisioner = Arc::new(MockProvisioner::default());
+        let factory = ManagedTunFactory::new(provisioner.clone());
+        factory.create(request(None)).await.unwrap();
+
+        factory.destroy("tsunfactory").await;
+        assert!(provisioner.host().names().is_empty());
+    }
+
+    #[tokio::test]
+    async fn a_host_that_cannot_be_provisioned_fails_the_create() {
+        let host = MockHost::new();
+        host.insert(
+            "tsunfactory",
+            super::super::InterfaceState {
+                kind: LinkKind::Foreign("bridge".into()),
+                attached: true,
+                up: true,
+                mtu: 1500,
+                addresses: Vec::new(),
+            },
+        );
+        let factory = ManagedTunFactory::new(Arc::new(MockProvisioner::new(host)));
+        let err = factory.create(request(None)).await.unwrap_err();
+        assert!(err.to_string().contains("bridge"), "{err}");
+    }
+
+    #[tokio::test]
+    async fn every_network_s_address_reaches_the_host_not_just_the_first() {
+        // One agent has one interface and a network apiece on it. Carrying
+        // only the first address is carrying only the first network: the
+        // rest have addresses the operating system has never heard of, and
+        // their traffic goes nowhere while the status says all is well.
+        let provisioner = Arc::new(MockProvisioner::default());
+        let factory = ManagedTunFactory::new(provisioner.clone());
+        factory
+            .create(addressed(vec![
+                Ipv4Addr::new(10, 13, 37, 69),
+                Ipv4Addr::new(10, 156, 200, 116),
+            ]))
+            .await
+            .unwrap();
+
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        let addresses: Vec<String> = state
+            .addresses
+            .iter()
+            .map(|entry| entry.to_string())
+            .collect();
+        assert!(
+            addresses.contains(&"10.13.37.69/24".to_string()),
+            "{addresses:?}"
+        );
+        assert!(
+            addresses.contains(&"10.156.200.116/24".to_string()),
+            "{addresses:?}"
+        );
+
+        // And leaving one network takes its address off, because the plan
+        // is what the interface should carry and nothing else.
+        factory
+            .reconfigure(addressed(vec![Ipv4Addr::new(10, 13, 37, 69)]))
+            .await
+            .unwrap();
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        let addresses: Vec<String> = state
+            .addresses
+            .iter()
+            .map(|entry| entry.to_string())
+            .collect();
+        assert_eq!(addresses, vec!["10.13.37.69/24".to_string()]);
+    }
+}

crates/tsunagi-core/src/overlay/provision/macos.rs

@@ -0,0 +1,531 @@
+//! Managing the overlay interface on macOS, through utun and `ifconfig`/`route`.
+//!
+//! The shape is the same as the Linux and Windows provisioners: observe what
+//! is there, [`plan_changes`](super::plan_changes) to decide what to do, apply
+//! the difference. Only the mechanisms differ.
+//!
+//! # Creating the interface
+//!
+//! The interface is a `utun` device, created through the `tun` dependency's
+//! safe wrapper — [`open_tun`](super::super::tun) — over the in-kernel
+//! `com.apple.net.utun_control` control socket. No kernel extension and no
+//! Network Extension entitlement is involved; what it needs is root. The
+//! device is not made persistent, so closing the last file descriptor removes
+//! the interface — it goes away when the agent does, however the agent ends.
+//!
+//! # The kernel names the interface, not us
+//!
+//! Unlike Linux and Windows, a utun interface cannot be given an arbitrary
+//! name: it is always `utunN`, and the unit number is assigned by the kernel.
+//! So the name derived from the network id (`tsun…`) is only ever a *request*;
+//! the real name is read back off the device after it is opened, and
+//! everything downstream — the host rules, the resolver publication — uses the
+//! real one. The agent adopts it as soon as the device is created (see
+//! [`Interface`](crate::overlay::Interface)).
+//!
+//! # Configuring it
+//!
+//! Addresses, MTU and routes are applied with `ifconfig` and `route`. On Linux
+//! the same work is done in process over a netlink socket; here there is no
+//! equivalent this crate may call, because the in-process route is raw `ioctl`
+//! and `PF_ROUTE` and that is `unsafe`. The tools are invoked by absolute path
+//! from `/sbin`, so nothing on `PATH` can stand in for them, and every value
+//! handed to them is one this agent derived — an interface name chosen by the
+//! kernel, an address it allocated — never anything a peer said.
+//!
+//! # Observing it
+//!
+//! The interface table is read through `netdev`, a safe wrapper over
+//! `getifaddrs`. An interface is matched by its real `utunN` name, which only
+//! matters on reconfiguration: the first create always makes a fresh device.
+
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::config::Cidr;
+use super::super::tun::{TunDevice, TunRequest, open_tun};
+use super::privilege::{Privilege, probe_net_admin};
+use super::{
+    InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, Provisioned, plan_changes,
+};
+
+/// Manages the overlay interface with utun and `ifconfig`/`route`.
+#[derive(Debug)]
+pub struct UtunProvisioner {
+    /// Interfaces this process created, by their real `utunN` name, so a
+    /// reconfiguration adjusts them rather than trying to create them again.
+    ours: Mutex<Vec<String>>,
+    /// Devices kept alive for as long as the interface should exist. Dropping
+    /// one closes the last descriptor, which removes the utin interface.
+    held: Mutex<Vec<(String, Arc<dyn TunDevice>)>>,
+}
+
+impl UtunProvisioner {
+    /// Creates the provisioner, after checking the platform supports one.
+    ///
+    /// Whether the process is actually root is not checked here — reading the
+    /// effective uid needs a libc call this crate forbids — so a missing
+    /// privilege surfaces at utun creation with a precise message, the same
+    /// way the Linux and Windows paths treat the open itself as the honest
+    /// answer.
+    pub fn new() -> Result<Self, OverlayError> {
+        match probe_net_admin() {
+            Privilege::Available => {}
+            Privilege::Missing(reason) => return Err(OverlayError::Unavailable(reason)),
+            Privilege::Unsupported => {
+                return Err(OverlayError::Unavailable(
+                    "interface management is not compiled in".to_string(),
+                ));
+            }
+        }
+        Ok(Self {
+            ours: Mutex::new(Vec::new()),
+            held: Mutex::new(Vec::new()),
+        })
+    }
+
+    fn is_ours(&self, name: &str) -> bool {
+        lock(&self.ours).iter().any(|owned| owned == name)
+    }
+
+    /// Opens the utun device, which is what creates it. The kernel assigns the
+    /// name, so the returned device reports a `utunN` that is not `plan.name`.
+    fn create_device(&self, plan: &InterfacePlan) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        let request = TunRequest::bare(plan.name.clone(), plan.mtu);
+        open_tun(&request)
+    }
+}
+
+impl InterfaceProvisioner for UtunProvisioner {
+    fn name(&self) -> &str {
+        "utun"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move {
+            // A reconfiguration: the plan carries the real `utunN` name the
+            // agent adopted after the first create, so the interface is there
+            // and only the addresses and MTU need bringing into line.
+            if self.is_ours(&plan.name) {
+                let current = observe(&plan.name).await;
+                let changes = plan_changes(&current, plan, true)?;
+                configure(&plan.name, &changes).await?;
+                return Ok(Provisioned {
+                    changes,
+                    device: None,
+                });
+            }
+
+            // The first create. The kernel assigns the unit, so the device is
+            // opened first and its real name used for everything after.
+            let device = self.create_device(plan)?;
+            let real = device.name().to_string();
+
+            // A fresh interface starts with nothing on it, so the whole plan is
+            // applied. `plan_changes` against an absent interface is exactly
+            // that decision; its `create_link` flag is already satisfied by the
+            // open above, so `configure` ignores it and only applies the rest.
+            let changes = plan_changes(&InterfaceState::absent(), plan, false)?;
+
+            // A failure here leaves an interface that exists but cannot carry
+            // traffic. Drop the device so the utun closes, and report, rather
+            // than leave a trap behind.
+            if let Err(err) = configure(&real, &changes).await {
+                drop(device);
+                return Err(err);
+            }
+
+            lock(&self.ours).push(real.clone());
+            lock(&self.held).push((real, Arc::clone(&device)));
+            Ok(Provisioned {
+                changes,
+                device: Some(device),
+            })
+        })
+    }
+
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            // Dropping the device closes the last descriptor, which removes the
+            // utun interface and the routes that were bound to it. There is
+            // nothing to flush: a utun never outlives the process that holds it.
+            lock(&self.held).retain(|(held, _)| held != name);
+            lock(&self.ours).retain(|owned| owned != name);
+            Ok(())
+        })
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+/// `fe80::/10` and `169.254.0.0/16`, which macOS assigns on its own.
+fn is_link_local(addr: IpAddr) -> bool {
+    match addr {
+        IpAddr::V4(addr) => addr.is_link_local(),
+        IpAddr::V6(addr) => (addr.segments()[0] & 0xffc0) == 0xfe80,
+    }
+}
+
+/// What the interface named `name` currently looks like.
+///
+/// Never fails: a name nothing is using reads as [`InterfaceState::absent`],
+/// which is what the plan then acts on.
+async fn observe(name: &str) -> InterfaceState {
+    let name = name.to_string();
+    let found = tokio::task::spawn_blocking(move || {
+        netdev::get_interfaces()
+            .into_iter()
+            .find(|iface| iface.name == name)
+    })
+    .await
+    .ok()
+    .flatten();
+
+    let Some(iface) = found else {
+        return InterfaceState::absent();
+    };
+
+    let mut addresses: Vec<Cidr> = Vec::new();
+    for net in &iface.ipv4 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V4(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    for net in &iface.ipv6 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V6(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    addresses.sort();
+
+    InterfaceState {
+        // Only ever reached for an interface this agent created (see
+        // `reconcile`), so it is ours and a TUN by construction.
+        kind: LinkKind::Tun,
+        attached: iface.is_up(),
+        up: iface.is_up(),
+        mtu: iface.mtu.unwrap_or(0),
+        addresses,
+    }
+}
+
+/// Applies MTU, link state, addresses and the on-link routes they imply.
+async fn configure(name: &str, changes: &super::Changes) -> Result<(), OverlayError> {
+    if let Some(mtu) = changes.set_mtu {
+        set_mtu(name, mtu).await?;
+    }
+    for cidr in &changes.remove {
+        // Best effort: an address that is already gone is the outcome wanted.
+        if let Err(err) = address_del(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not remove address");
+        }
+        if let Err(err) = route_del(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not remove route");
+        }
+    }
+    for cidr in &changes.add {
+        address_add(name, *cidr).await?;
+        // The subnet route makes the whole overlay range reach the interface;
+        // a point-to-point utun does not install one from the address alone. A
+        // host address (/32 or /128) needs none. Best effort: a route that is
+        // already there is the outcome wanted.
+        if let Err(err) = route_add(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not add route");
+        }
+    }
+    if changes.bring_up {
+        // Best effort: a utun with a descriptor open is up already, and a
+        // failure to nudge it is not a reason to fail the whole reconcile.
+        if let Err(err) = set_up(name).await {
+            tracing::debug!(interface = %name, %err, "could not enable the interface");
+        }
+    }
+    Ok(())
+}
+
+/// The network address of a CIDR, with the host bits cleared.
+fn network_of(cidr: Cidr) -> IpAddr {
+    match cidr.addr {
+        IpAddr::V4(addr) => {
+            let bits = u32::from(addr);
+            let mask = if cidr.prefix_len == 0 {
+                0
+            } else {
+                u32::MAX << (32 - u32::from(cidr.prefix_len))
+            };
+            IpAddr::V4(Ipv4Addr::from(bits & mask))
+        }
+        IpAddr::V6(addr) => {
+            let bits = u128::from(addr);
+            let mask = if cidr.prefix_len == 0 {
+                0
+            } else {
+                u128::MAX << (128 - u32::from(cidr.prefix_len))
+            };
+            IpAddr::V6(Ipv6Addr::from(bits & mask))
+        }
+    }
+}
+
+/// Whether a CIDR is a single host address, which needs no subnet route.
+fn is_host_route(cidr: Cidr) -> bool {
+    match cidr.addr {
+        IpAddr::V4(_) => cidr.prefix_len >= 32,
+        IpAddr::V6(_) => cidr.prefix_len >= 128,
+    }
+}
+
+/// The `ifconfig` arguments that add an address to a utun interface.
+///
+/// A utun is point-to-point, so the IPv4 form names the address as its own
+/// peer; `alias` lets several coexist on the one interface.
+fn address_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            name.into(),
+            "inet".into(),
+            format!("{addr}/{}", cidr.prefix_len),
+            addr.to_string(),
+            "alias".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            name.into(),
+            "inet6".into(),
+            format!("{addr}/{}", cidr.prefix_len),
+            "alias".into(),
+        ],
+    }
+}
+
+/// The `ifconfig` arguments that remove an address from a utun interface.
+fn address_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "inet",
+        IpAddr::V6(_) => "inet6",
+    };
+    vec![
+        name.into(),
+        family.into(),
+        cidr.addr.to_string(),
+        "-alias".into(),
+    ]
+}
+
+/// The `route` arguments that point a CIDR's subnet at the interface.
+fn route_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "-inet",
+        IpAddr::V6(_) => "-inet6",
+    };
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "add".into(),
+        family.into(),
+        "-net".into(),
+        format!("{}/{}", network_of(cidr), cidr.prefix_len),
+        "-interface".into(),
+        name.into(),
+    ]
+}
+
+/// The `route` arguments that remove that subnet route.
+fn route_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "-inet",
+        IpAddr::V6(_) => "-inet6",
+    };
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "delete".into(),
+        family.into(),
+        "-net".into(),
+        format!("{}/{}", network_of(cidr), cidr.prefix_len),
+        "-interface".into(),
+        name.into(),
+    ]
+}
+
+async fn address_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    ifconfig(address_add_args(name, cidr)).await
+}
+
+async fn address_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    ifconfig(address_del_args(name, cidr)).await
+}
+
+async fn route_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    if is_host_route(cidr) {
+        return Ok(());
+    }
+    route(route_add_args(name, cidr)).await
+}
+
+async fn route_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    if is_host_route(cidr) {
+        return Ok(());
+    }
+    route(route_del_args(name, cidr)).await
+}
+
+async fn set_mtu(name: &str, mtu: u32) -> Result<(), OverlayError> {
+    ifconfig(vec![name.into(), "mtu".into(), mtu.to_string()]).await
+}
+
+async fn set_up(name: &str) -> Result<(), OverlayError> {
+    ifconfig(vec![name.into(), "up".into()]).await
+}
+
+/// Runs `/sbin/ifconfig`, turning a non-zero exit into an error with its text.
+async fn ifconfig(args: Vec<String>) -> Result<(), OverlayError> {
+    run("/sbin/ifconfig", args).await
+}
+
+/// Runs `/sbin/route`, turning a non-zero exit into an error with its text.
+async fn route(args: Vec<String>) -> Result<(), OverlayError> {
+    run("/sbin/route", args).await
+}
+
+/// Runs a system tool by absolute path, so nothing on `PATH` can stand in for
+/// it, and turns a non-zero exit into an error carrying what it said.
+async fn run(program: &'static str, args: Vec<String>) -> Result<(), OverlayError> {
+    let display = format!("{program} {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program).args(&args).output()
+    })
+    .await
+    .map_err(|err| OverlayError::Other(format!("could not run {display}: {err}")))?
+    .map_err(|err| OverlayError::Unavailable(format!("could not run {display}: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let message = {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        let text = if stderr.trim().is_empty() {
+            String::from_utf8_lossy(&output.stdout).trim().to_string()
+        } else {
+            stderr.trim().to_string()
+        };
+        if text.is_empty() {
+            format!("exit code {}", output.status)
+        } else {
+            text
+        }
+    };
+    Err(OverlayError::Unavailable(format!(
+        "{display} failed: {message}"
+    )))
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn v4(last: u8, prefix: u8) -> Cidr {
+        Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)), prefix).unwrap()
+    }
+
+    fn v6(last: u16, prefix: u8) -> Cidr {
+        Cidr::new(
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, last)),
+            prefix,
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn a_v4_address_is_added_point_to_point_with_an_alias() {
+        let args = address_add_args("utun5", v4(69, 24));
+        assert_eq!(
+            args,
+            vec!["utun5", "inet", "10.13.37.69/24", "10.13.37.69", "alias"]
+        );
+    }
+
+    #[test]
+    fn a_v6_address_is_added_with_a_prefix_and_an_alias() {
+        let args = address_add_args("utun5", v6(1, 128));
+        assert_eq!(args, vec!["utun5", "inet6", "fd00::1/128", "alias"]);
+    }
+
+    #[test]
+    fn deleting_an_address_names_the_family_and_removes_the_alias() {
+        let args = address_del_args("utun5", v4(69, 24));
+        assert_eq!(args, vec!["utun5", "inet", "10.13.37.69", "-alias"]);
+        let args = address_del_args("utun5", v6(1, 128));
+        assert_eq!(args, vec!["utun5", "inet6", "fd00::1", "-alias"]);
+    }
+
+    #[test]
+    fn a_subnet_route_points_the_masked_network_at_the_interface() {
+        let args = route_add_args("utun5", v4(69, 24));
+        assert_eq!(
+            args,
+            vec![
+                "-q",
+                "-n",
+                "add",
+                "-inet",
+                "-net",
+                "10.13.37.0/24",
+                "-interface",
+                "utun5"
+            ]
+        );
+    }
+
+    #[test]
+    fn a_host_address_installs_no_subnet_route() {
+        assert!(is_host_route(v4(69, 32)));
+        assert!(is_host_route(v6(1, 128)));
+        assert!(!is_host_route(v4(69, 24)));
+        assert!(!is_host_route(v6(1, 64)));
+    }
+
+    #[test]
+    fn the_network_address_clears_the_host_bits() {
+        assert_eq!(
+            network_of(v4(69, 24)),
+            IpAddr::V4(Ipv4Addr::new(10, 13, 37, 0))
+        );
+        assert_eq!(
+            network_of(Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 69)), 16).unwrap()),
+            IpAddr::V4(Ipv4Addr::new(10, 13, 0, 0))
+        );
+        assert_eq!(
+            network_of(v6(0x1234, 64)),
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, 0))
+        );
+    }
+
+    #[test]
+    fn link_local_addresses_are_not_ours_to_manage() {
+        assert!(is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfe80, 0, 0, 0, 0, 0, 0, 1
+        ))));
+        assert!(is_link_local(IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
+        assert!(!is_link_local(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 1))));
+        assert!(!is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfd00, 0, 0, 0, 0, 0, 0, 1
+        ))));
+    }
+}

crates/tsunagi-core/src/overlay/provision/mock.rs

@@ -0,0 +1,336 @@
+//! An in-memory host, so provisioning is tested without touching this one.
+//!
+//! [`MockHost`] is a pretend `/sys/class/net`: a test can seed it with the
+//! leftovers of a crashed run, or with somebody else's bridge, then check
+//! what the provisioner did about it. It is also what the platforms that have
+//! no provisioner yet are wired to in their own tests.
+
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::config::Cidr;
+use super::super::tun::{MemoryTun, MemoryTunFactory, TunFactory, TunRequest};
+use super::{
+    Changes, InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, Provisioned,
+    plan_changes,
+};
+
+/// A pretend host with interfaces on it.
+#[derive(Debug, Clone, Default)]
+pub struct MockHost {
+    links: Arc<Mutex<HashMap<String, InterfaceState>>>,
+}
+
+impl MockHost {
+    /// An empty host.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    fn lock(&self) -> std::sync::MutexGuard<'_, HashMap<String, InterfaceState>> {
+        match self.links.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    /// Puts an interface on the host.
+    pub fn insert(&self, name: impl Into<String>, state: InterfaceState) {
+        self.lock().insert(name.into(), state);
+    }
+
+    /// Seeds the leftovers of a run that died: a TUN nobody holds open, still
+    /// carrying whatever addresses it had.
+    pub fn insert_stale_tun(&self, name: impl Into<String>, addresses: Vec<Cidr>) {
+        self.insert(
+            name,
+            InterfaceState {
+                kind: LinkKind::Tun,
+                attached: false,
+                up: true,
+                mtu: 1280,
+                addresses,
+            },
+        );
+    }
+
+    /// The state of an interface, if it exists.
+    pub fn get(&self, name: &str) -> Option<InterfaceState> {
+        self.lock().get(name).cloned()
+    }
+
+    /// The names currently on the host.
+    pub fn names(&self) -> Vec<String> {
+        let mut names: Vec<String> = self.lock().keys().cloned().collect();
+        names.sort();
+        names
+    }
+}
+
+/// Applies plans to a [`MockHost`].
+#[derive(Debug, Clone)]
+pub struct MockProvisioner {
+    host: MockHost,
+    ours: Arc<Mutex<Vec<String>>>,
+    /// The devices handed out, so a test can drive packets through them.
+    devices: MemoryTunFactory,
+    /// Set to fail every call, to exercise the error path.
+    failure: Option<String>,
+}
+
+impl Default for MockProvisioner {
+    fn default() -> Self {
+        Self::new(MockHost::new())
+    }
+}
+
+impl MockProvisioner {
+    /// A provisioner over a host.
+    pub fn new(host: MockHost) -> Self {
+        Self {
+            host,
+            ours: Arc::new(Mutex::new(Vec::new())),
+            devices: MemoryTunFactory::new(),
+            failure: None,
+        }
+    }
+
+    /// A provisioner that refuses everything, with this reason.
+    pub fn failing(reason: impl Into<String>) -> Self {
+        Self {
+            host: MockHost::new(),
+            ours: Arc::new(Mutex::new(Vec::new())),
+            devices: MemoryTunFactory::new(),
+            failure: Some(reason.into()),
+        }
+    }
+
+    /// The host it applies to.
+    pub fn host(&self) -> &MockHost {
+        &self.host
+    }
+
+    /// The device created for an interface name, if any.
+    pub fn device(&self, name: &str) -> Option<Arc<MemoryTun>> {
+        self.devices.device(name)
+    }
+
+    fn owned(&self) -> std::sync::MutexGuard<'_, Vec<String>> {
+        match self.ours.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    fn apply(&self, plan: &InterfacePlan) -> Result<Changes, OverlayError> {
+        if let Some(reason) = &self.failure {
+            return Err(OverlayError::Unavailable(reason.clone()));
+        }
+
+        let ours = self.owned().iter().any(|name| name == &plan.name);
+        let current = self
+            .host
+            .get(&plan.name)
+            .unwrap_or_else(InterfaceState::absent);
+        let changes = plan_changes(&current, plan, ours)?;
+
+        let mut state = current;
+        if changes.delete_link {
+            self.host.lock().remove(&plan.name);
+            state = InterfaceState::absent();
+        }
+        if changes.create_link {
+            state = InterfaceState {
+                kind: LinkKind::Tun,
+                // Creating it means holding it open, so it has carrier.
+                attached: true,
+                up: false,
+                mtu: 1500,
+                addresses: Vec::new(),
+            };
+            self.owned().push(plan.name.clone());
+        }
+        if let Some(mtu) = changes.set_mtu {
+            state.mtu = mtu;
+        }
+        if changes.bring_up {
+            state.up = true;
+        }
+        state
+            .addresses
+            .retain(|addr| !changes.remove.contains(addr));
+        state.addresses.extend(changes.add.iter().copied());
+        state.addresses.sort();
+        state.addresses.dedup();
+        self.host.insert(plan.name.clone(), state);
+
+        Ok(changes)
+    }
+}
+
+impl InterfaceProvisioner for MockProvisioner {
+    fn name(&self) -> &str {
+        "mock"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move {
+            let changes = self.apply(plan)?;
+            let device = if changes.create_link {
+                Some(
+                    self.devices
+                        .create(TunRequest::bare(plan.name.clone(), plan.mtu))
+                        .await?,
+                )
+            } else {
+                None
+            };
+            Ok(Provisioned { changes, device })
+        })
+    }
+
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            self.host.lock().remove(name);
+            self.owned().retain(|owned| owned != name);
+            Ok(())
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+
+    fn v6() -> Cidr {
+        Cidr {
+            addr: IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, 1)),
+            prefix_len: 64,
+        }
+    }
+
+    fn v4(last: u8) -> Cidr {
+        Cidr {
+            addr: IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)),
+            prefix_len: 24,
+        }
+    }
+
+    fn plan(addresses: Vec<Cidr>) -> InterfacePlan {
+        InterfacePlan::new("tsunmock", 1280, addresses)
+    }
+
+    #[tokio::test]
+    async fn reconciling_an_empty_host_creates_a_configured_interface() {
+        let provisioner = MockProvisioner::default();
+        let plan = plan(vec![v6(), v4(69)]);
+        let changes = provisioner.reconcile(&plan).await.unwrap().changes;
+        assert!(changes.create_link);
+
+        let state = provisioner.host().get("tsunmock").unwrap();
+        assert_eq!(state.kind, LinkKind::Tun);
+        assert!(state.up);
+        assert_eq!(state.mtu, 1280);
+        assert_eq!(state.addresses, vec![v4(69), v6()]);
+    }
+
+    #[tokio::test]
+    async fn reconciling_is_idempotent() {
+        let provisioner = MockProvisioner::default();
+        let plan = plan(vec![v6(), v4(69)]);
+        provisioner.reconcile(&plan).await.unwrap();
+        let second = provisioner.reconcile(&plan).await.unwrap().changes;
+        assert!(second.is_empty(), "{second:?}");
+    }
+
+    #[tokio::test]
+    async fn a_crashed_run_is_repaired_on_the_next_start() {
+        let host = MockHost::new();
+        // What the previous run left: the interface, with the address it had
+        // been allocated back then.
+        host.insert_stale_tun("tsunmock", vec![v4(178)]);
+        let provisioner = MockProvisioner::new(host);
+
+        let changes = provisioner
+            .reconcile(&plan(vec![v6(), v4(69)]))
+            .await
+            .unwrap()
+            .changes;
+        assert!(changes.delete_link && changes.create_link);
+
+        let state = provisioner.host().get("tsunmock").unwrap();
+        assert_eq!(
+            state.addresses,
+            vec![v4(69), v6()],
+            "the stale address is gone and the current one is there"
+        );
+        assert!(state.attached, "the new interface is held open by us");
+    }
+
+    #[tokio::test]
+    async fn a_changed_allocation_is_applied_without_recreating_the_interface() {
+        let provisioner = MockProvisioner::default();
+        provisioner
+            .reconcile(&plan(vec![v6(), v4(69)]))
+            .await
+            .unwrap();
+
+        // The overlay agreed on a different address for us while running.
+        let changes = provisioner
+            .reconcile(&plan(vec![v6(), v4(70)]))
+            .await
+            .unwrap()
+            .changes;
+        assert!(
+            !changes.delete_link && !changes.create_link,
+            "recreating would drop every tunnel"
+        );
+        assert_eq!(changes.add, vec![v4(70)]);
+        assert_eq!(changes.remove, vec![v4(69)]);
+        assert_eq!(
+            provisioner.host().get("tsunmock").unwrap().addresses,
+            vec![v4(70), v6()]
+        );
+    }
+
+    #[tokio::test]
+    async fn removing_takes_the_interface_off_the_host_and_is_idempotent() {
+        let provisioner = MockProvisioner::default();
+        provisioner.reconcile(&plan(vec![v6()])).await.unwrap();
+        assert_eq!(provisioner.host().names(), vec!["tsunmock".to_string()]);
+
+        provisioner.remove("tsunmock").await.unwrap();
+        assert!(provisioner.host().names().is_empty());
+        provisioner.remove("tsunmock").await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn a_foreign_interface_makes_reconciling_fail_and_changes_nothing() {
+        let host = MockHost::new();
+        host.insert(
+            "tsunmock",
+            InterfaceState {
+                kind: LinkKind::Foreign("bridge".into()),
+                attached: true,
+                up: true,
+                mtu: 1500,
+                addresses: vec![v4(1)],
+            },
+        );
+        let provisioner = MockProvisioner::new(host);
+        assert!(provisioner.reconcile(&plan(vec![v6()])).await.is_err());
+
+        let state = provisioner.host().get("tsunmock").unwrap();
+        assert_eq!(state.kind, LinkKind::Foreign("bridge".into()));
+        assert_eq!(state.addresses, vec![v4(1)], "left exactly as it was");
+    }
+}

crates/tsunagi-core/src/overlay/provision/mod.rs

@@ -0,0 +1,456 @@
+//! Bringing a host interface into the state the overlay needs.
+//!
+//! The agent used to print a list of `ip` commands and ask a human to run
+//! them. That is fragile in exactly the way hand-held setup always is: the
+//! interface does not survive a reboot, a changed address allocation needs
+//! another manual round, and a crashed run leaves a half-configured interface
+//! that the next run then trips over.
+//!
+//! So the agent does it itself, and the shape of that is a *reconciliation*:
+//! it is handed an [`InterfacePlan`] describing what the interface should look
+//! like, it observes what is actually there, and it applies the difference.
+//! Running it twice changes nothing the second time, and running it after a
+//! crash repairs whatever was left behind.
+//!
+//! # Why this is a trait
+//!
+//! Every platform does this differently — netlink on Linux, `SystemConfiguration`
+//! on macOS, the Windows IP Helper API — while the *decision* of what to change
+//! is the same everywhere. So the decision lives in [`plan_changes`], which is
+//! pure and tested on every platform, and only the execution is behind
+//! [`InterfaceProvisioner`].
+//!
+//! Five implementations:
+//!
+//! * `NetlinkProvisioner` on Linux, which needs `CAP_NET_ADMIN`.
+//! * `WintunProvisioner` on Windows, a Wintun adapter configured with `netsh`,
+//!   which needs an elevated process.
+//! * `UtunProvisioner` on macOS, a utun interface configured with `ifconfig`
+//!   and `route`, which needs root.
+//! * [`MockProvisioner`], an in-memory host used by the tests.
+//! * [`UnsupportedProvisioner`] elsewhere, which fails with an explanation
+//!   and a pointer at the manual route rather than pretending to work.
+//!
+//! # What it is not allowed to do
+//!
+//! Nothing here takes a name, an address or a command from the network. The
+//! interface name is derived from the network id, the addresses come from the
+//! local plugin, and an interface this agent did not create is never deleted
+//! or reconfigured — see [`plan_changes`].
+
+use std::collections::BTreeSet;
+use std::sync::Arc;
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::config::Cidr;
+use super::tun::TunDevice;
+
+mod mock;
+pub use mock::{MockHost, MockProvisioner};
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use linux::NetlinkProvisioner;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub(crate) use linux::RouteHandle;
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+mod windows;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows::WintunProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos::UtunProvisioner;
+
+mod privilege;
+pub use privilege::{Privilege, probe_net_admin};
+
+mod unsupported;
+pub use unsupported::UnsupportedProvisioner;
+
+mod factory;
+pub use factory::ManagedTunFactory;
+
+/// What an interface should look like.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct InterfacePlan {
+    /// Interface name. Derived from the network id, never from the network.
+    pub name: String,
+    /// Interface MTU.
+    pub mtu: u32,
+    /// Every address the interface should carry, and no others.
+    pub addresses: Vec<Cidr>,
+}
+
+impl InterfacePlan {
+    /// Builds a plan, normalising the address list.
+    pub fn new(name: impl Into<String>, mtu: u32, addresses: Vec<Cidr>) -> Self {
+        let unique: BTreeSet<Cidr> = addresses.into_iter().collect();
+        Self {
+            name: name.into(),
+            mtu,
+            addresses: unique.into_iter().collect(),
+        }
+    }
+}
+
+/// What kind of link is sitting on a name.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum LinkKind {
+    /// Nothing is there.
+    Absent,
+    /// A TUN interface.
+    Tun,
+    /// Something else entirely — a bridge, a physical device, a VPN from
+    /// another program. Never ours to touch.
+    Foreign(String),
+}
+
+/// What an interface currently looks like.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct InterfaceState {
+    /// What kind of link holds the name.
+    pub kind: LinkKind,
+    /// Whether a process is attached to the TUN, which the kernel reports as
+    /// carrier. A leftover interface from a crashed run has none.
+    pub attached: bool,
+    /// Whether the link is administratively up.
+    pub up: bool,
+    /// The current MTU.
+    pub mtu: u32,
+    /// The addresses currently assigned.
+    pub addresses: Vec<Cidr>,
+}
+
+impl InterfaceState {
+    /// The state of a name nothing is using.
+    pub fn absent() -> Self {
+        Self {
+            kind: LinkKind::Absent,
+            attached: false,
+            up: false,
+            mtu: 0,
+            addresses: Vec::new(),
+        }
+    }
+}
+
+/// The steps that turn an [`InterfaceState`] into an [`InterfacePlan`].
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct Changes {
+    /// Remove a stale interface left behind by an earlier run.
+    pub delete_link: bool,
+    /// Create the interface.
+    pub create_link: bool,
+    /// Set the MTU, when it is not already right.
+    pub set_mtu: Option<u32>,
+    /// Bring the link up.
+    pub bring_up: bool,
+    /// Addresses to add.
+    pub add: Vec<Cidr>,
+    /// Addresses to remove, because the plan no longer contains them.
+    pub remove: Vec<Cidr>,
+}
+
+impl Changes {
+    /// Whether anything at all needs doing.
+    pub fn is_empty(&self) -> bool {
+        !self.delete_link
+            && !self.create_link
+            && self.set_mtu.is_none()
+            && !self.bring_up
+            && self.add.is_empty()
+            && self.remove.is_empty()
+    }
+
+    /// A one-line summary for a log line.
+    pub fn summary(&self) -> String {
+        if self.is_empty() {
+            return "already as planned".to_string();
+        }
+        let mut parts = Vec::new();
+        if self.delete_link {
+            parts.push("remove a stale interface".to_string());
+        }
+        if self.create_link {
+            parts.push("create the interface".to_string());
+        }
+        if let Some(mtu) = self.set_mtu {
+            parts.push(format!("set mtu {mtu}"));
+        }
+        if self.bring_up {
+            parts.push("bring it up".to_string());
+        }
+        for address in &self.add {
+            parts.push(format!("add {address}"));
+        }
+        for address in &self.remove {
+            parts.push(format!("remove {address}"));
+        }
+        parts.join(", ")
+    }
+}
+
+/// Decides what to change, or refuses.
+///
+/// `ours` says whether this process created the interface in its current run.
+/// It is the whole reason this can be safe: an interface we made is adjusted
+/// in place, and an interface we did not make is only ever *replaced* when it
+/// is plainly abandoned.
+///
+/// The refusals matter more than the changes:
+///
+/// * A link that is not a TUN is never touched. Deriving the name from the
+///   network id makes a collision with a real device unlikely, not impossible,
+///   and destroying somebody's bridge because it happened to share a name is
+///   not a recoverable mistake.
+/// * A TUN with a process attached to it is never deleted. It is somebody
+///   else's working interface — most likely another agent on this host — and
+///   yanking it out from under them would break a running overlay.
+///
+/// What is left is a TUN with nothing attached, which is precisely the
+/// footprint of a run that died: those are removed and rebuilt.
+pub fn plan_changes(
+    current: &InterfaceState,
+    plan: &InterfacePlan,
+    ours: bool,
+) -> Result<Changes, OverlayError> {
+    let mut changes = Changes::default();
+
+    match &current.kind {
+        LinkKind::Foreign(kind) => {
+            return Err(OverlayError::Unavailable(format!(
+                "`{}` already exists and is a {kind} interface, not one of ours. \
+                 Refusing to touch it. Run with a different interface prefix.",
+                plan.name
+            )));
+        }
+        LinkKind::Tun if !ours && current.attached => {
+            return Err(OverlayError::Unavailable(format!(
+                "`{}` already exists and another process is attached to it. \
+                 That is most likely a second agent on this host in the same \
+                 network; give one of them a different interface prefix.",
+                plan.name
+            )));
+        }
+        LinkKind::Tun if !ours => {
+            // Abandoned: a TUN with no carrier is one nothing holds open. It
+            // is either a leftover from a run that died or an interface made
+            // by the old manual recipe. Either way it is replaced, which also
+            // discards whatever stale addresses it carried.
+            changes.delete_link = true;
+            changes.create_link = true;
+        }
+        LinkKind::Tun => {}
+        LinkKind::Absent => changes.create_link = true,
+    }
+
+    if changes.create_link {
+        // A fresh interface starts down, with the kernel default MTU and no
+        // addresses, so everything in the plan has to be applied.
+        changes.set_mtu = Some(plan.mtu);
+        changes.bring_up = true;
+        changes.add = plan.addresses.clone();
+        return Ok(changes);
+    }
+
+    if current.mtu != plan.mtu {
+        changes.set_mtu = Some(plan.mtu);
+    }
+    if !current.up {
+        changes.bring_up = true;
+    }
+
+    let wanted: BTreeSet<Cidr> = plan.addresses.iter().copied().collect();
+    let present: BTreeSet<Cidr> = current.addresses.iter().copied().collect();
+    changes.add = wanted.difference(&present).copied().collect();
+    changes.remove = present.difference(&wanted).copied().collect();
+
+    Ok(changes)
+}
+
+/// The result of a reconciliation.
+pub struct Provisioned {
+    /// What was changed to get here.
+    pub changes: Changes,
+    /// The packet interface, when this call is what created it.
+    ///
+    /// Creating the interface and configuring it are the same privileged act
+    /// and belong together, so the provisioner owns both. A reconciliation
+    /// that only adjusted an interface already in place returns `None`.
+    pub device: Option<Arc<dyn TunDevice>>,
+}
+
+impl std::fmt::Debug for Provisioned {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("Provisioned")
+            .field("changes", &self.changes)
+            .field("device", &self.device.as_ref().map(|device| device.name()))
+            .finish()
+    }
+}
+
+/// Applies an [`InterfacePlan`] to the host.
+///
+/// Implementations are expected to be idempotent: calling [`reconcile`] twice
+/// with the same plan changes nothing the second time.
+///
+/// [`reconcile`]: InterfaceProvisioner::reconcile
+pub trait InterfaceProvisioner: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Brings the interface in line with the plan, and reports what it did.
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>>;
+
+    /// Removes an interface this provisioner created.
+    ///
+    /// Removing one that is already gone succeeds: this runs on the shutdown
+    /// path, where the interface having vanished is the desired outcome.
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>>;
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+
+    fn v6(last: u16) -> Cidr {
+        Cidr {
+            addr: IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, last)),
+            prefix_len: 64,
+        }
+    }
+
+    fn v4(last: u8) -> Cidr {
+        Cidr {
+            addr: IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)),
+            prefix_len: 24,
+        }
+    }
+
+    fn plan() -> InterfacePlan {
+        InterfacePlan::new("tsuntest", 1280, vec![v6(1), v4(69)])
+    }
+
+    #[test]
+    fn an_absent_interface_is_created_and_fully_configured() {
+        let changes = plan_changes(&InterfaceState::absent(), &plan(), false).unwrap();
+        assert!(changes.create_link);
+        assert!(!changes.delete_link);
+        assert_eq!(changes.set_mtu, Some(1280));
+        assert!(changes.bring_up);
+        assert_eq!(changes.add, vec![v4(69), v6(1)]);
+        assert!(changes.remove.is_empty());
+    }
+
+    #[test]
+    fn an_abandoned_tun_from_a_crashed_run_is_replaced() {
+        // The footprint of a run that died: the interface is still there, it
+        // still has its old addresses, and nothing holds it open.
+        let current = InterfaceState {
+            kind: LinkKind::Tun,
+            attached: false,
+            up: true,
+            mtu: 1280,
+            addresses: vec![v4(178)],
+        };
+        let changes = plan_changes(&current, &plan(), false).unwrap();
+        assert!(changes.delete_link, "the stale interface goes");
+        assert!(changes.create_link);
+        // Replacing it discards the stale address, so it need not be removed
+        // one by one.
+        assert_eq!(changes.add, vec![v4(69), v6(1)]);
+        assert!(changes.remove.is_empty());
+    }
+
+    #[test]
+    fn a_foreign_interface_is_never_touched() {
+        let current = InterfaceState {
+            kind: LinkKind::Foreign("bridge".into()),
+            attached: true,
+            up: true,
+            mtu: 1500,
+            addresses: vec![v4(1)],
+        };
+        let err = plan_changes(&current, &plan(), false).unwrap_err();
+        let message = err.to_string();
+        assert!(message.contains("bridge"), "{message}");
+        assert!(message.contains("Refusing to touch it"), "{message}");
+    }
+
+    #[test]
+    fn a_tun_another_process_is_using_is_never_deleted() {
+        let current = InterfaceState {
+            kind: LinkKind::Tun,
+            attached: true,
+            up: true,
+            mtu: 1280,
+            addresses: vec![v6(1)],
+        };
+        let err = plan_changes(&current, &plan(), false).unwrap_err();
+        assert!(err.to_string().contains("another process"), "{err}");
+    }
+
+    #[test]
+    fn our_own_interface_is_adjusted_in_place() {
+        // The live case: our address allocation changed while running. The
+        // interface must not be recreated, or every tunnel on it would drop.
+        let current = InterfaceState {
+            kind: LinkKind::Tun,
+            attached: true,
+            up: true,
+            mtu: 1280,
+            addresses: vec![v6(1), v4(178)],
+        };
+        let changes = plan_changes(&current, &plan(), true).unwrap();
+        assert!(!changes.delete_link && !changes.create_link);
+        assert_eq!(changes.add, vec![v4(69)]);
+        assert_eq!(changes.remove, vec![v4(178)]);
+    }
+
+    #[test]
+    fn reconciling_an_interface_that_already_matches_changes_nothing() {
+        let current = InterfaceState {
+            kind: LinkKind::Tun,
+            attached: true,
+            up: true,
+            mtu: 1280,
+            addresses: vec![v4(69), v6(1)],
+        };
+        let changes = plan_changes(&current, &plan(), true).unwrap();
+        assert!(changes.is_empty(), "{changes:?}");
+        assert_eq!(changes.summary(), "already as planned");
+    }
+
+    #[test]
+    fn a_wrong_mtu_or_a_down_link_is_corrected() {
+        let current = InterfaceState {
+            kind: LinkKind::Tun,
+            attached: true,
+            up: false,
+            mtu: 1500,
+            addresses: vec![v4(69), v6(1)],
+        };
+        let changes = plan_changes(&current, &plan(), true).unwrap();
+        assert_eq!(changes.set_mtu, Some(1280));
+        assert!(changes.bring_up);
+        assert!(changes.add.is_empty() && changes.remove.is_empty());
+    }
+
+    #[test]
+    fn a_plan_deduplicates_and_orders_its_addresses() {
+        let plan = InterfacePlan::new("x", 1280, vec![v6(1), v4(69), v6(1)]);
+        assert_eq!(plan.addresses, vec![v4(69), v6(1)]);
+    }
+}

crates/tsunagi-core/src/overlay/provision/privilege.rs

@@ -0,0 +1,222 @@
+//! Holding `CAP_NET_ADMIN` for as short a time as possible.
+//!
+//! Creating a TUN interface and assigning addresses to it needs
+//! `CAP_NET_ADMIN`, and there is no way around that on Linux. What *is* in
+//! our control is how long the process can actually use it.
+//!
+//! Linux splits capabilities into sets. The *permitted* set is what a process
+//! may use; the *effective* set is what it may use **right now**. A process
+//! can lower a capability out of effective and raise it back later, but it
+//! can never add to permitted. So the agent keeps `CAP_NET_ADMIN` out of the
+//! effective set and raises it only around the handful of netlink calls that
+//! need it, which is a few milliseconds at startup and again whenever the
+//! address allocation changes.
+//!
+//! Grant it with:
+//!
+//! ```text
+//! sudo setcap cap_net_admin,cap_net_bind_service+p /usr/local/bin/tsng
+//! ```
+//!
+//! `+p` rather than `+ep`: with `+p` the capability is permitted but not
+//! effective at exec, which is exactly the resting state this module wants.
+//! `+ep` also works — [`NetAdmin::acquire`] lowers it on the way in.
+//!
+//! # Capabilities are per thread
+//!
+//! `capset` affects the calling thread only, so raising one inside an async
+//! block would be a bug the moment the task migrated to another worker. Every
+//! caller here runs on a single dedicated thread; see
+//! [`linux`](super::linux).
+
+/// Whether this process can configure interfaces.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum Privilege {
+    /// `CAP_NET_ADMIN` is available. The agent manages the interface itself.
+    Available,
+    /// It is not, with a description of what was found.
+    Missing(String),
+    /// This platform has no provisioner yet.
+    Unsupported,
+}
+
+impl Privilege {
+    /// Whether interfaces can be managed.
+    pub fn is_available(&self) -> bool {
+        matches!(self, Privilege::Available)
+    }
+
+    /// How to obtain it, for a diagnostic.
+    pub fn how_to_grant(program: &str) -> String {
+        format!(
+            "Grant it once with `sudo setcap cap_net_admin,cap_net_bind_service+p {program}` \
+             and the agent manages its own interface. Without it, run with \
+             `--no-tun`: the tunnels still form, they just do not reach the \
+             operating system."
+        )
+    }
+}
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use linux_impl::{NetAdmin, probe_net_admin};
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows_impl::probe_net_admin;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos_impl::probe_net_admin;
+
+#[cfg(not(any(
+    all(feature = "tun-device", target_os = "linux"),
+    all(feature = "tun-device", target_os = "windows"),
+    all(feature = "tun-device", target_os = "macos")
+)))]
+pub use other_impl::probe_net_admin;
+
+#[cfg(not(any(
+    all(feature = "tun-device", target_os = "linux"),
+    all(feature = "tun-device", target_os = "windows"),
+    all(feature = "tun-device", target_os = "macos")
+)))]
+mod other_impl {
+    use super::Privilege;
+
+    /// Whether this process can configure interfaces.
+    pub fn probe_net_admin() -> Privilege {
+        Privilege::Unsupported
+    }
+}
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+mod windows_impl {
+    use super::Privilege;
+
+    /// Whether this process can configure interfaces.
+    ///
+    /// Windows has no capability to hold and lower the way Linux does:
+    /// creating an adapter simply needs the process to be elevated. Reading
+    /// whether it *is* elevated means inspecting the process token through a
+    /// raw call this crate forbids, so the probe is deliberately optimistic —
+    /// it reports that the platform can manage interfaces — and the real
+    /// check is left to Wintun's adapter creation, which fails with a precise
+    /// message when the process is not elevated. This matches how the Linux
+    /// path treats the open itself as the honest answer.
+    pub fn probe_net_admin() -> Privilege {
+        Privilege::Available
+    }
+}
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos_impl {
+    use super::Privilege;
+
+    /// Whether this process can configure interfaces.
+    ///
+    /// macOS has no capability to hold and lower the way Linux does: creating
+    /// a utun interface simply needs the process to be root. Reading the
+    /// effective uid means a libc call this crate forbids, so the probe is
+    /// deliberately optimistic — it reports that the platform can manage
+    /// interfaces — and the real check is left to the utun open, which fails
+    /// with a precise "run with sudo" message when the process is not root.
+    /// This matches how the Linux path treats the open itself as the honest
+    /// answer.
+    pub fn probe_net_admin() -> Privilege {
+        Privilege::Available
+    }
+}
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux_impl {
+    use caps::{CapSet, Capability};
+
+    use super::Privilege;
+    use crate::overlay::OverlayError;
+
+    /// Whether this thread holds `CAP_NET_ADMIN` in its permitted set.
+    pub fn probe_net_admin() -> Privilege {
+        match caps::has_cap(None, CapSet::Permitted, Capability::CAP_NET_ADMIN) {
+            Ok(true) => Privilege::Available,
+            Ok(false) => Privilege::Missing(
+                "this process does not hold CAP_NET_ADMIN, so it cannot create \
+                 or configure a network interface"
+                    .to_string(),
+            ),
+            Err(err) => {
+                Privilege::Missing(format!("cannot read this process's capabilities: {err}"))
+            }
+        }
+    }
+
+    /// `CAP_NET_ADMIN`, raised for as long as this value is alive.
+    ///
+    /// Dropping it lowers the capability again, including on the error paths,
+    /// which is the point of it being a guard rather than a pair of calls.
+    #[derive(Debug)]
+    pub struct NetAdmin {
+        /// Whether this guard is the one that raised it, and so the one that
+        /// must lower it. Nested acquisition leaves the inner guard inert.
+        raised: bool,
+    }
+
+    impl NetAdmin {
+        /// Raises `CAP_NET_ADMIN` into the effective set.
+        pub fn acquire() -> Result<Self, OverlayError> {
+            let already = caps::has_cap(None, CapSet::Effective, Capability::CAP_NET_ADMIN)
+                .map_err(|err| {
+                    OverlayError::Unavailable(format!("cannot read capabilities: {err}"))
+                })?;
+            if already {
+                return Ok(Self { raised: false });
+            }
+            caps::raise(None, CapSet::Effective, Capability::CAP_NET_ADMIN).map_err(|err| {
+                OverlayError::Unavailable(format!(
+                    "cannot raise CAP_NET_ADMIN: {err}. {}",
+                    Privilege::how_to_grant("tsng")
+                ))
+            })?;
+            Ok(Self { raised: true })
+        }
+
+        /// Lowers `CAP_NET_ADMIN` out of the effective set of this thread.
+        ///
+        /// Called on the way in as well as on the way out, so that a binary
+        /// granted `cap_net_admin+ep` — which starts with it effective — still
+        /// spends almost all of its life unable to use it.
+        pub fn lower() {
+            let _ = caps::drop(None, CapSet::Effective, Capability::CAP_NET_ADMIN);
+        }
+    }
+
+    impl Drop for NetAdmin {
+        fn drop(&mut self) {
+            if self.raised {
+                Self::lower();
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[test]
+    fn the_grant_instructions_name_the_program() {
+        let text = Privilege::how_to_grant("/usr/local/bin/tsng");
+        assert!(text.contains("setcap cap_net_admin,cap_net_bind_service+p /usr/local/bin/tsng"));
+        assert!(text.contains("--no-tun"), "the fallback is offered too");
+    }
+
+    #[test]
+    fn probing_says_something_definite_about_this_host() {
+        // Whatever the answer is, it must be one of the three, and a missing
+        // capability must come with a reason rather than a bare `false`.
+        match probe_net_admin() {
+            Privilege::Available => {}
+            Privilege::Missing(reason) => assert!(!reason.is_empty()),
+            Privilege::Unsupported => {}
+        }
+    }
+}

crates/tsunagi-core/src/overlay/provision/unsupported.rs

@@ -0,0 +1,80 @@
+//! The provisioner for platforms that do not have one yet.
+//!
+//! macOS and Windows both need real work here — `utun` plus the
+//! `SystemConfiguration` framework on one, the IP Helper API and a Wintun
+//! adapter on the other — and neither is written. Rather than let the agent
+//! come up and fail obscurely at the first packet, this refuses at the point
+//! of provisioning and says what to do instead.
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::{InterfacePlan, InterfaceProvisioner, Provisioned};
+
+/// Refuses to provision, with an explanation.
+#[derive(Debug, Clone)]
+pub struct UnsupportedProvisioner {
+    platform: &'static str,
+}
+
+impl Default for UnsupportedProvisioner {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl UnsupportedProvisioner {
+    /// A provisioner naming the platform it is standing in for.
+    pub fn new() -> Self {
+        Self {
+            platform: std::env::consts::OS,
+        }
+    }
+
+    fn refusal(&self) -> OverlayError {
+        OverlayError::Unavailable(format!(
+            "managing the overlay interface is not implemented on {} yet. \
+             Run with `--no-tun` until it is: the tunnels still form, they just \
+             do not reach the operating system.",
+            self.platform
+        ))
+    }
+}
+
+impl InterfaceProvisioner for UnsupportedProvisioner {
+    fn name(&self) -> &str {
+        "unsupported"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        _plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move { Err(self.refusal()) })
+    }
+
+    fn remove<'a>(&'a self, _name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        // Nothing was ever created, so there is nothing to clean up and no
+        // reason to fail a shutdown path.
+        Box::pin(async move { Ok(()) })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn it_refuses_with_a_usable_pointer_and_still_cleans_up_quietly() {
+        let provisioner = UnsupportedProvisioner::new();
+        let plan = InterfacePlan::new("tsuntest", 1280, Vec::new());
+        let err = provisioner.reconcile(&plan).await.unwrap_err();
+        let message = err.to_string();
+        assert!(message.contains("--no-tun"), "{message}");
+        assert!(message.contains(std::env::consts::OS), "{message}");
+
+        provisioner.remove("tsuntest").await.unwrap();
+    }
+}

crates/tsunagi-core/src/overlay/provision/windows.rs

@@ -0,0 +1,513 @@
+//! Managing the overlay interface on Windows, through Wintun and `netsh`.
+//!
+//! The shape is the same as the Linux provisioner: observe what is there,
+//! [`plan_changes`](super::plan_changes) to decide what to do, apply the
+//! difference. Only the two mechanisms differ.
+//!
+//! # Creating the interface
+//!
+//! The interface is a Wintun adapter, created through the `tun` dependency's
+//! safe wrapper — [`open_tun`](super::super::tun) — because loading
+//! `wintun.dll` is an `unsafe` call and this crate forbids `unsafe`. The
+//! adapter is created, not made persistent: closing the last handle to it
+//! removes it, so it goes away when the agent does, however the agent ends.
+//! The reader and writer inside the device hold that handle, so the adapter
+//! lives exactly as long as the device does.
+//!
+//! # Configuring it
+//!
+//! Addresses and MTU are applied with `netsh`. On Linux the same work is done
+//! in process over a netlink socket; here there is no equivalent that this
+//! crate may call, because the in-process route is the IP Helper API and that
+//! is `unsafe`. `netsh` is the supported tool for the job, it is invoked by
+//! absolute path from `%SystemRoot%` so nothing on `PATH` can stand in for it,
+//! and every value handed to it is one this agent derived — an interface name
+//! from the network id, an address it allocated — never anything a peer said.
+//!
+//! # Observing it
+//!
+//! The interface table is read through `netdev`, whose enumeration is a safe
+//! wrapper over the same IP Helper API. An adapter is matched by the friendly
+//! name it was created with. A match that carries a default gateway is treated
+//! as foreign and left alone: a tsunagi overlay interface never has one, so a
+//! gateway is the mark of a real device that happens to share the name.
+//!
+//! # What Windows cannot do that Linux can
+//!
+//! Wintun exposes no way to delete an adapter this process did not create, so
+//! a leftover from a run that was killed is *reused* rather than replaced: its
+//! addresses are flushed and the planned ones applied, which reaches the same
+//! end state. A reused adapter is not removed on exit, only closed — that is a
+//! Wintun limitation, and the common case of a clean start and a clean stop is
+//! unaffected.
+
+use std::net::{IpAddr, Ipv4Addr};
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::config::Cidr;
+use super::super::tun::{TunDevice, TunRequest, open_tun};
+use super::privilege::{Privilege, probe_net_admin};
+use super::{
+    InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, Provisioned, plan_changes,
+};
+
+/// Manages the overlay interface with Wintun and `netsh`.
+#[derive(Debug)]
+pub struct WintunProvisioner {
+    /// Interfaces this process created, so they are adjusted rather than
+    /// replaced. See [`plan_changes`].
+    ours: Mutex<Vec<String>>,
+    /// Devices kept alive for as long as the interface should exist. Dropping
+    /// one closes the adapter, which removes it if this process created it.
+    held: Mutex<Vec<(String, Arc<dyn TunDevice>)>>,
+}
+
+impl WintunProvisioner {
+    /// Creates the provisioner, after checking the platform supports one.
+    ///
+    /// Whether the process is actually elevated is not checked here — that
+    /// cannot be read without an `unsafe` call — so a missing privilege
+    /// surfaces at adapter creation with a precise message, the same way the
+    /// Linux path treats the open itself as the honest answer.
+    pub fn new() -> Result<Self, OverlayError> {
+        match probe_net_admin() {
+            Privilege::Available => {}
+            Privilege::Missing(reason) => return Err(OverlayError::Unavailable(reason)),
+            Privilege::Unsupported => {
+                return Err(OverlayError::Unavailable(
+                    "interface management is not compiled in".to_string(),
+                ));
+            }
+        }
+        Ok(Self {
+            ours: Mutex::new(Vec::new()),
+            held: Mutex::new(Vec::new()),
+        })
+    }
+
+    fn is_ours(&self, name: &str) -> bool {
+        lock(&self.ours)
+            .iter()
+            .any(|owned| owned.eq_ignore_ascii_case(name))
+    }
+
+    /// Opens the Wintun adapter, which is what creates it.
+    fn create_device(&self, plan: &InterfacePlan) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        let request = TunRequest::bare(plan.name.clone(), plan.mtu);
+        open_tun(&request)
+    }
+}
+
+impl InterfaceProvisioner for WintunProvisioner {
+    fn name(&self) -> &str {
+        "wintun"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move {
+            let current = observe(&plan.name).await;
+            let changes = plan_changes(&current, plan, self.is_ours(&plan.name))?;
+
+            // Wintun cannot delete an adapter this process did not create, so a
+            // "delete then create" from `plan_changes` becomes "flush then
+            // reuse": clear the stale addresses off the interface that is
+            // there, and let the create step below reopen the same adapter.
+            // The end state — an adapter carrying exactly the plan's addresses
+            // — is identical.
+            let flush: Vec<Cidr> = if changes.delete_link {
+                tracing::info!(
+                    interface = %plan.name,
+                    "reusing an abandoned interface left by an earlier run"
+                );
+                current.addresses.clone()
+            } else {
+                Vec::new()
+            };
+
+            for cidr in flush.iter().chain(changes.remove.iter()) {
+                // Best effort: an address that is already gone is the outcome
+                // wanted, and refusing to start over one is not worth it.
+                if let Err(err) = address_del(&plan.name, *cidr).await {
+                    tracing::debug!(interface = %plan.name, %cidr, %err, "could not remove address");
+                }
+            }
+
+            let device = if changes.create_link {
+                let device = self.create_device(plan)?;
+                lock(&self.ours).push(plan.name.clone());
+                lock(&self.held).push((plan.name.clone(), Arc::clone(&device)));
+                Some(device)
+            } else {
+                None
+            };
+
+            // A failure past this point leaves an interface that exists but
+            // cannot carry traffic. Undo the bookkeeping for one we just
+            // created so it is not mistaken for a working interface, and drop
+            // the device so the adapter closes.
+            let configured = configure(plan, &changes).await;
+            if let Err(err) = configured {
+                if changes.create_link {
+                    lock(&self.held).retain(|(held, _)| held != &plan.name);
+                    lock(&self.ours).retain(|owned| owned != &plan.name);
+                }
+                return Err(err);
+            }
+
+            Ok(Provisioned { changes, device })
+        })
+    }
+
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            // Dropping the device closes the adapter, which removes it if this
+            // process created it. A reused adapter is only closed, not removed;
+            // flushing its addresses first keeps no stale configuration behind.
+            let current = observe(name).await;
+            for cidr in &current.addresses {
+                if let Err(err) = address_del(name, *cidr).await {
+                    tracing::debug!(interface = %name, %cidr, %err, "could not remove address");
+                }
+            }
+            lock(&self.held).retain(|(held, _)| !held.eq_ignore_ascii_case(name));
+            lock(&self.ours).retain(|owned| !owned.eq_ignore_ascii_case(name));
+            Ok(())
+        })
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+/// `fe80::/10` and `169.254.0.0/16`, which Windows assigns on its own.
+fn is_link_local(addr: IpAddr) -> bool {
+    match addr {
+        IpAddr::V4(addr) => addr.is_link_local(),
+        IpAddr::V6(addr) => (addr.segments()[0] & 0xffc0) == 0xfe80,
+    }
+}
+
+/// What the interface named `name` currently looks like.
+///
+/// Never fails: a name nothing is using reads as [`InterfaceState::absent`],
+/// which is what the plan then acts on.
+async fn observe(name: &str) -> InterfaceState {
+    let name = name.to_string();
+    let found = tokio::task::spawn_blocking(move || {
+        netdev::get_interfaces()
+            .into_iter()
+            .find(|iface| match &iface.friendly_name {
+                Some(friendly) => friendly.eq_ignore_ascii_case(&name),
+                None => false,
+            })
+    })
+    .await
+    .ok()
+    .flatten();
+
+    let Some(iface) = found else {
+        return InterfaceState::absent();
+    };
+
+    let mut addresses: Vec<Cidr> = Vec::new();
+    for net in &iface.ipv4 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V4(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    for net in &iface.ipv6 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V6(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    addresses.sort();
+
+    // A tsunagi overlay interface only ever carries the addresses it was given
+    // and no gateway. One that has a gateway is a real device sharing the name,
+    // so it is foreign and never touched.
+    let kind = if iface.gateway.is_some() {
+        LinkKind::Foreign("gatewayed".to_string())
+    } else {
+        LinkKind::Tun
+    };
+
+    InterfaceState {
+        kind,
+        // Wintun reports an adapter with a live session as up and an abandoned
+        // one as down, so "up" doubles as "some process is holding it".
+        attached: iface.is_up(),
+        up: iface.is_up(),
+        mtu: iface.mtu.unwrap_or(0),
+        addresses,
+    }
+}
+
+/// Applies MTU, link state and addresses.
+async fn configure(plan: &InterfacePlan, changes: &super::Changes) -> Result<(), OverlayError> {
+    if let Some(mtu) = changes.set_mtu {
+        set_mtu(&plan.name, mtu).await?;
+    }
+    for cidr in &changes.add {
+        if let Err(err) = address_add(&plan.name, *cidr).await {
+            let err_str = err.to_string();
+            if !err_str.contains("already exists") && !err_str.contains("0x80071392") {
+                return Err(err);
+            }
+        }
+    }
+    if changes.bring_up {
+        // Best effort: a Wintun adapter with a session is up already, and a
+        // failure to nudge it is not a reason to fail the whole reconcile.
+        if let Err(err) = set_up(&plan.name).await {
+            tracing::debug!(interface = %plan.name, %err, "could not enable the interface");
+        }
+    }
+    Ok(())
+}
+
+/// The dotted netmask for an IPv4 prefix length.
+fn ipv4_mask(prefix_len: u8) -> Ipv4Addr {
+    if prefix_len == 0 {
+        Ipv4Addr::UNSPECIFIED
+    } else {
+        Ipv4Addr::from(u32::MAX << (32 - u32::from(prefix_len)))
+    }
+}
+
+/// The `netsh` arguments that add an address to an interface.
+fn address_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            "interface".into(),
+            "ipv4".into(),
+            "add".into(),
+            "address".into(),
+            format!("name={name}"),
+            format!("address={addr}"),
+            format!("mask={}", ipv4_mask(cidr.prefix_len)),
+            "store=active".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            "interface".into(),
+            "ipv6".into(),
+            "add".into(),
+            "address".into(),
+            format!("interface={name}"),
+            format!("address={addr}/{}", cidr.prefix_len),
+            "store=active".into(),
+        ],
+    }
+}
+
+/// The `netsh` arguments that remove an address from an interface.
+fn address_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            "interface".into(),
+            "ipv4".into(),
+            "delete".into(),
+            "address".into(),
+            format!("name={name}"),
+            format!("address={addr}"),
+            "store=active".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            "interface".into(),
+            "ipv6".into(),
+            "delete".into(),
+            "address".into(),
+            format!("interface={name}"),
+            format!("address={addr}"),
+            "store=active".into(),
+        ],
+    }
+}
+
+/// The `netsh` arguments that set an interface's MTU, one call per family.
+fn mtu_args(name: &str, family: &str, mtu: u32) -> Vec<String> {
+    vec![
+        "interface".into(),
+        family.into(),
+        "set".into(),
+        "subinterface".into(),
+        name.into(),
+        format!("mtu={mtu}"),
+        "store=active".into(),
+    ]
+}
+
+async fn address_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    netsh(address_add_args(name, cidr)).await
+}
+
+async fn address_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    netsh(address_del_args(name, cidr)).await
+}
+
+/// Sets the MTU on both families.
+///
+/// A family that is turned off on the adapter cannot take an MTU, so this
+/// fails only when *neither* would — one family carrying the plan's MTU is
+/// enough, and refusing the whole interface because the other is disabled
+/// would be wrong.
+async fn set_mtu(name: &str, mtu: u32) -> Result<(), OverlayError> {
+    let v4 = netsh(mtu_args(name, "ipv4", mtu)).await;
+    let v6 = netsh(mtu_args(name, "ipv6", mtu)).await;
+    match (&v4, &v6) {
+        (Err(_), Err(_)) => v4,
+        _ => {
+            if let Err(err) = &v4 {
+                tracing::debug!(interface = %name, %err, "could not set the IPv4 MTU");
+            }
+            if let Err(err) = &v6 {
+                tracing::debug!(interface = %name, %err, "could not set the IPv6 MTU");
+            }
+            Ok(())
+        }
+    }
+}
+
+async fn set_up(name: &str) -> Result<(), OverlayError> {
+    netsh(vec![
+        "interface".into(),
+        "set".into(),
+        "interface".into(),
+        format!("name={name}"),
+        "admin=enabled".into(),
+    ])
+    .await
+}
+
+/// Runs `netsh` from `%SystemRoot%\System32`, so nothing on `PATH` can stand
+/// in for it, and turns a non-zero exit into an error carrying what it said.
+async fn netsh(args: Vec<String>) -> Result<(), OverlayError> {
+    let program = system32("netsh.exe");
+    let display = format!("netsh {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program).args(&args).output()
+    })
+    .await
+    .map_err(|err| OverlayError::Other(format!("could not run {display}: {err}")))?
+    .map_err(|err| OverlayError::Unavailable(format!("could not run {display}: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let message = {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        let text = if stderr.trim().is_empty() {
+            String::from_utf8_lossy(&output.stdout).trim().to_string()
+        } else {
+            stderr.trim().to_string()
+        };
+        if text.is_empty() {
+            format!("exit code {}", output.status)
+        } else {
+            text
+        }
+    };
+    Err(OverlayError::Unavailable(format!(
+        "{display} failed: {message}"
+    )))
+}
+
+/// The absolute path to a program in `System32`.
+fn system32(exe: &str) -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root).join("System32").join(exe)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::Ipv6Addr;
+
+    fn v4(last: u8, prefix: u8) -> Cidr {
+        Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)), prefix).unwrap()
+    }
+
+    fn v6(last: u16) -> Cidr {
+        Cidr::new(
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, last)),
+            64,
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn a_prefix_length_becomes_the_right_dotted_mask() {
+        assert_eq!(ipv4_mask(24), Ipv4Addr::new(255, 255, 255, 0));
+        assert_eq!(ipv4_mask(16), Ipv4Addr::new(255, 255, 0, 0));
+        assert_eq!(ipv4_mask(32), Ipv4Addr::new(255, 255, 255, 255));
+        assert_eq!(ipv4_mask(0), Ipv4Addr::UNSPECIFIED);
+    }
+
+    #[test]
+    fn a_v4_address_is_added_by_name_with_a_dotted_mask() {
+        let args = address_add_args("tsun0", v4(69, 24));
+        assert!(args.contains(&"ipv4".to_string()));
+        assert!(args.contains(&"add".to_string()));
+        assert!(args.contains(&"name=tsun0".to_string()));
+        assert!(args.contains(&"address=10.13.37.69".to_string()));
+        assert!(args.contains(&"mask=255.255.255.0".to_string()));
+    }
+
+    #[test]
+    fn a_v6_address_is_added_by_interface_with_a_prefix() {
+        let args = address_add_args("tsun0", v6(1));
+        assert!(args.contains(&"ipv6".to_string()));
+        assert!(args.contains(&"interface=tsun0".to_string()));
+        assert!(args.contains(&"address=fd00::1/64".to_string()));
+    }
+
+    #[test]
+    fn deleting_an_address_names_no_mask() {
+        let args = address_del_args("tsun0", v4(69, 24));
+        assert!(args.contains(&"delete".to_string()));
+        assert!(args.contains(&"address=10.13.37.69".to_string()));
+        assert!(!args.iter().any(|arg| arg.starts_with("mask=")));
+    }
+
+    #[test]
+    fn the_mtu_is_set_per_family_on_the_subinterface() {
+        let args = mtu_args("tsun0", "ipv4", 1280);
+        assert!(args.contains(&"subinterface".to_string()));
+        assert!(args.contains(&"tsun0".to_string()));
+        assert!(args.contains(&"mtu=1280".to_string()));
+    }
+
+    #[test]
+    fn link_local_addresses_are_not_ours_to_manage() {
+        assert!(is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfe80, 0, 0, 0, 0, 0, 0, 1
+        ))));
+        assert!(is_link_local(IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
+        assert!(!is_link_local(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 1))));
+        assert!(!is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfd00, 0, 0, 0, 0, 0, 0, 1
+        ))));
+    }
+
+    #[test]
+    fn system32_is_absolute_and_ends_with_the_program() {
+        let path = system32("netsh.exe");
+        assert!(path.is_absolute(), "{path:?}");
+        assert!(path.ends_with("netsh.exe"), "{path:?}");
+    }
+}

crates/tsunagi-core/src/overlay/router.rs

@@ -0,0 +1,976 @@
+//! Which peer a packet is for, and whether a peer may have sent it.
+//!
+//! One agent has one interface, so a packet coming off it could belong to any
+//! network the agent is in and to any protocol currently carrying traffic.
+//! Deciding which is this table's job, and it is the reason the interface
+//! belongs to the system level: the answer comes from signed state, which no
+//! protocol owns.
+//!
+//! Two questions, and they are not the same one:
+//!
+//! * **Outbound**: this destination address — whose is it? A packet for
+//!   nobody is dropped rather than broadcast.
+//! * **Inbound**: this peer decrypted a packet with this source address —
+//!   is that address actually its? A peer may not speak for anybody else,
+//!   and the check consults the signed claim rather than anything the peer
+//!   said.
+//!
+//! Nothing here is derived from a protocol's key. An address is allocated at
+//! the system level and signed by the member that holds it, so every protocol
+//! carries traffic for the same addresses and the table is the same whichever
+//! one is in use.
+
+use super::broadcast::{BroadcastPolicy, BroadcastTable};
+use arc_swap::ArcSwap;
+use std::collections::HashMap;
+use std::net::Ipv4Addr;
+use std::sync::{Arc, RwLock};
+
+use iroh::EndpointId;
+
+use crate::identity::NetworkId;
+use crate::state::Ipv4Range;
+
+/// Where an outbound packet is going.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Route {
+    /// The network the destination belongs to.
+    pub network: NetworkId,
+    /// The member that holds the destination address.
+    pub peer: EndpointId,
+}
+
+/// A network's exit-node settings, as far as packets are concerned.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct ExitPolicy {
+    /// This agent offers itself as an exit node here: packets from members
+    /// for addresses outside every overlay range are handed to the host.
+    pub offer: bool,
+    /// The member all other traffic is sent to, if this device uses one.
+    pub via: Option<EndpointId>,
+}
+
+/// What one network contributes to the table.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct NetworkRoutes {
+    /// Exit-node settings of this network.
+    pub exit: ExitPolicy,
+    /// Local broadcast policy and authenticated willing recipients.
+    pub broadcast: BroadcastPolicy,
+    /// The range this network allocates from, once it has agreed one.
+    pub range: Option<Ipv4Range>,
+    /// This agent's own address in the network.
+    pub local: Option<Ipv4Addr>,
+    /// Every other member's address, as the signed state has it.
+    pub peers: Vec<(Ipv4Addr, EndpointId)>,
+}
+
+/// Why a network cannot be added to the table.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum RouteError {
+    /// Two networks would answer for the same addresses.
+    #[error(
+        "network {network} uses {range}, which overlaps {other_range} already in use by \
+         {other}. One agent has one interface, so an address can belong to only one of \
+         them: stop the other network before starting this one."
+    )]
+    Overlap {
+        /// The network being added.
+        network: NetworkId,
+        /// The range it wants.
+        range: Ipv4Range,
+        /// The network already using an overlapping range.
+        other: NetworkId,
+        /// That network's range.
+        other_range: Ipv4Range,
+    },
+}
+
+/// Address ownership across every network this agent is in.
+#[derive(Debug, Default)]
+pub struct RoutingTable {
+    pub(crate) vips: Arc<crate::vip::VipRouter>,
+    broadcast: ArcSwap<BroadcastTable>,
+    networks: RwLock<HashMap<NetworkId, NetworkRoutes>>,
+}
+
+impl RoutingTable {
+    /// An empty table.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    fn read(&self) -> std::sync::RwLockReadGuard<'_, HashMap<NetworkId, NetworkRoutes>> {
+        match self.networks.read() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    fn write(&self) -> std::sync::RwLockWriteGuard<'_, HashMap<NetworkId, NetworkRoutes>> {
+        match self.networks.write() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    /// Replaces what one network contributes.
+    ///
+    /// Rejected when its range overlaps another network's. With one
+    /// interface an address can belong to only one network, and guessing
+    /// which would send somebody's traffic to a stranger.
+    pub fn set_network(&self, network: NetworkId, routes: NetworkRoutes) -> Result<(), RouteError> {
+        let mut networks = self.write();
+        if let Some(range) = routes.range {
+            for (other, existing) in networks.iter() {
+                if *other == network {
+                    continue;
+                }
+                let Some(other_range) = existing.range else {
+                    continue;
+                };
+                if ranges_overlap(range, other_range) {
+                    return Err(RouteError::Overlap {
+                        network,
+                        range,
+                        other: *other,
+                        other_range,
+                    });
+                }
+            }
+        }
+        networks.insert(network, routes);
+        self.broadcast
+            .store(Arc::new(BroadcastTable::build(&networks)));
+        Ok(())
+    }
+
+    /// Forgets a network.
+    pub fn remove_network(&self, network: NetworkId) {
+        self.vips.remove_network(network);
+        let mut networks = self.write();
+        networks.remove(&network);
+        self.broadcast
+            .store(Arc::new(BroadcastTable::build(&networks)));
+    }
+
+    /// Replaces only a network's broadcast participation, without changing addresses.
+    pub fn set_broadcast(&self, network: NetworkId, policy: BroadcastPolicy) {
+        let mut networks = self.write();
+        if let Some(routes) = networks.get_mut(&network) {
+            if routes.broadcast == policy {
+                return;
+            }
+            routes.broadcast = policy;
+            self.broadcast
+                .store(Arc::new(BroadcastTable::build(&networks)));
+        }
+    }
+
+    /// Replaces only a network's exit-node settings, without changing addresses.
+    ///
+    /// A network that has no entry yet (its range is not agreed) takes them
+    /// from [`NetworkRoutes::exit`] when it is added.
+    pub fn set_exit(&self, network: NetworkId, policy: ExitPolicy) {
+        if let Some(routes) = self.write().get_mut(&network) {
+            routes.exit = policy;
+        }
+    }
+
+    /// Recognizes limited and configured subnet broadcasts before unicast lookup.
+    pub fn is_broadcast(&self, destination: Ipv4Addr) -> bool {
+        self.broadcast.load().is_destination(destination)
+    }
+
+    /// A precomputed domain-scoped recipient list; no per-packet graph search.
+    pub fn broadcast_recipients(
+        &self,
+        source: Ipv4Addr,
+        destination: Ipv4Addr,
+    ) -> Option<Arc<[Route]>> {
+        self.broadcast.load().outgoing(source, destination)
+    }
+
+    /// Local admission after the plugin authenticated and decrypted the sender.
+    pub fn accepts_broadcast(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        destination: Ipv4Addr,
+    ) -> bool {
+        self.broadcast.load().accepts(network, peer, destination)
+    }
+
+    /// The peer that holds a destination address, if anybody does.
+    ///
+    /// This agent's own address returns `None`: a packet for ourselves does
+    /// not go over a tunnel, and answering with a peer would send it to one.
+    ///
+    /// An address nobody holds, outside every overlay range and an ordinary
+    /// unicast destination, goes to this device's exit node when it has one:
+    /// that is what sending *all* traffic through a member means.
+    pub fn route(&self, destination: Ipv4Addr) -> Option<Route> {
+        let networks = self.read();
+        for (network, routes) in networks.iter() {
+            if routes.local == Some(destination) {
+                return None;
+            }
+            if let Some((_, peer)) = routes
+                .peers
+                .iter()
+                .find(|(address, _)| *address == destination)
+            {
+                return Some(Route {
+                    network: *network,
+                    peer: *peer,
+                });
+            }
+        }
+        // Inside an overlay range, an unowned address is nobody's, not the
+        // internet's.
+        if !is_internet_destination(destination) || in_any_range(&networks, destination) {
+            return None;
+        }
+        exit_via(&networks).map(|(network, peer)| Route { network, peer })
+    }
+
+    /// Whether a peer may send from a source address.
+    ///
+    /// A peer speaks only for the address the network agreed it holds. The
+    /// one exception is this device's exit node: what comes back from the
+    /// internet carries the internet's addresses, so from *that* peer any
+    /// source outside the overlay ranges is accepted.
+    pub fn may_send_from(&self, network: NetworkId, peer: EndpointId, source: Ipv4Addr) -> bool {
+        let networks = self.read();
+        let Some(routes) = networks.get(&network) else {
+            return false;
+        };
+        if routes
+            .peers
+            .iter()
+            .any(|(address, holder)| *address == source && *holder == peer)
+        {
+            return true;
+        }
+        routes.exit.via == Some(peer)
+            && is_internet_destination(source)
+            && !in_any_range(&networks, source)
+    }
+
+    /// Whether this agent, as an exit node of `network`, takes a member's
+    /// packet for `destination` to the host.
+    ///
+    /// Only an ordinary unicast address outside every overlay range, so a
+    /// packet for another member, for this agent or for the local network
+    /// segment never leaves by this door. The sender's own source address is
+    /// checked separately, by [`Self::may_send_from`].
+    pub fn accepts_exit_traffic(&self, network: NetworkId, destination: Ipv4Addr) -> bool {
+        let networks = self.read();
+        networks
+            .get(&network)
+            .is_some_and(|routes| routes.exit.offer)
+            && is_internet_destination(destination)
+            && !in_any_range(&networks, destination)
+    }
+
+    /// The network and member this device sends its internet traffic through.
+    ///
+    /// One default route, so one answer; if several networks claim one (they
+    /// should not) the lowest network id decides, so two runs agree.
+    pub fn exit_via(&self) -> Option<(NetworkId, EndpointId)> {
+        exit_via(&self.read())
+    }
+
+    /// Networks that offer this agent as an exit node, with the range each
+    /// agreed. A network that has not agreed one yet has nothing to
+    /// masquerade and is left out.
+    pub fn exit_offers(&self) -> Vec<(NetworkId, Ipv4Range)> {
+        let mut offers: Vec<(NetworkId, Ipv4Range)> = self
+            .read()
+            .iter()
+            .filter(|(_, routes)| routes.exit.offer)
+            .filter_map(|(network, routes)| Some((*network, routes.range?)))
+            .collect();
+        offers.sort_by_key(|(network, _)| *network);
+        offers
+    }
+
+    /// The range of every network in the table that has agreed one, sorted.
+    pub fn overlay_ranges(&self) -> Vec<Ipv4Range> {
+        let mut ranges: Vec<Ipv4Range> = self
+            .read()
+            .values()
+            .filter_map(|routes| routes.range)
+            .collect();
+        ranges.sort_by_key(|range| (range.base, range.prefix_len));
+        ranges.dedup();
+        ranges
+    }
+
+    /// One network's exit-node settings and its range, if it is in the table.
+    pub fn exit_settings(&self, network: NetworkId) -> Option<(ExitPolicy, Option<Ipv4Range>)> {
+        self.read()
+            .get(&network)
+            .map(|routes| (routes.exit, routes.range))
+    }
+
+    /// Whether an ordinary packet terminates on this host in this network.
+    /// Exported LAN subnets must extend this admission policy explicitly later.
+    pub fn is_local_destination(&self, network: NetworkId, destination: Ipv4Addr) -> bool {
+        self.read()
+            .get(&network)
+            .is_some_and(|routes| routes.local == Some(destination))
+    }
+
+    /// Every address this agent should answer to, with its prefix length.
+    pub fn local_addresses(&self) -> Vec<(Ipv4Addr, u8)> {
+        let mut addresses: Vec<(Ipv4Addr, u8)> = self
+            .read()
+            .values()
+            .filter_map(|routes| Some((routes.local?, routes.range?.prefix_len)))
+            .collect();
+        addresses.sort();
+        addresses
+    }
+
+    /// The one network whose address the limited broadcast route is bound to.
+    ///
+    /// The host has a single `255.255.255.255` route per interface, and one
+    /// interface carries every network, so only one overlay address can be its
+    /// source. The choice is the lowest [`NetworkId`] among networks that have
+    /// broadcast enabled, a range and a local address: deterministic, so two
+    /// runs with the same networks pick the same one.
+    pub fn broadcast_host_target(&self) -> Option<(Ipv4Addr, Ipv4Range)> {
+        let networks = self.read();
+        networks
+            .iter()
+            .filter(|(_, routes)| routes.broadcast.enabled)
+            .filter_map(|(network, routes)| Some((*network, routes.local?, routes.range?)))
+            .min_by_key(|(network, _, _)| *network)
+            .map(|(_, local, range)| (local, range))
+    }
+
+    /// Whether a range would collide with one another network already uses.
+    ///
+    /// Asked before proposing a range rather than after: with one interface
+    /// an agent that claimed an address it could not route would also be
+    /// telling everybody else to use that range, and "the range of the
+    /// lowest author wins" would spread the collision instead of containing
+    /// it.
+    pub fn would_overlap(&self, network: NetworkId, range: Ipv4Range) -> Option<Ipv4Range> {
+        self.read().iter().find_map(|(other, existing)| {
+            if *other == network {
+                return None;
+            }
+            let other_range = existing.range?;
+            ranges_overlap(range, other_range).then_some(other_range)
+        })
+    }
+
+    /// All active networks whose reserved ranges intersect this range.
+    pub fn overlapping_networks(
+        &self,
+        network: NetworkId,
+        range: Ipv4Range,
+    ) -> Vec<(NetworkId, Ipv4Range)> {
+        let mut conflicts: Vec<_> = self
+            .read()
+            .iter()
+            .filter_map(|(other, routes)| {
+                let other_range = routes.range?;
+                (*other != network && ranges_overlap(range, other_range))
+                    .then_some((*other, other_range))
+            })
+            .collect();
+        conflicts.sort_by_key(|(id, _)| *id);
+        conflicts
+    }
+
+    /// Current signed allocations used for encrypted data delivery checks.
+    pub(crate) fn probe_addresses(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+    ) -> Option<(Ipv4Addr, Ipv4Addr)> {
+        let table = self.read();
+        let routes = table.get(&network)?;
+        Some((
+            routes.local?,
+            routes
+                .peers
+                .iter()
+                .find(|(_, holder)| *holder == peer)
+                .map(|(address, _)| *address)?,
+        ))
+    }
+
+    /// How many networks the table covers.
+    pub fn len(&self) -> usize {
+        self.read().len()
+    }
+
+    /// Whether it covers none.
+    pub fn is_empty(&self) -> bool {
+        self.read().is_empty()
+    }
+}
+
+/// Whether an address is an ordinary unicast one: not this host, not a
+/// multicast or broadcast group, not link-local, not unspecified.
+fn is_internet_destination(address: Ipv4Addr) -> bool {
+    !(address.is_unspecified()
+        || address.is_loopback()
+        || address.is_multicast()
+        || address.is_broadcast()
+        || address.is_link_local())
+}
+
+/// Whether an address lies in any network's range.
+fn in_any_range(networks: &HashMap<NetworkId, NetworkRoutes>, address: Ipv4Addr) -> bool {
+    networks
+        .values()
+        .any(|routes| routes.range.is_some_and(|range| range.contains(address)))
+}
+
+/// The exit node in use, if any, from the lowest network that names one.
+fn exit_via(networks: &HashMap<NetworkId, NetworkRoutes>) -> Option<(NetworkId, EndpointId)> {
+    networks
+        .iter()
+        .filter_map(|(network, routes)| Some((*network, routes.exit.via?)))
+        .min_by_key(|(network, _)| *network)
+}
+
+/// Whether two ranges share any address.
+fn ranges_overlap(one: Ipv4Range, other: Ipv4Range) -> bool {
+    // A range contains the other's base, or the other way round. With
+    // prefix-aligned ranges that is the whole of it: two blocks either nest
+    // or are disjoint.
+    one.contains(other.base) || other.contains(one.base)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+
+    fn network(name: &str) -> NetworkId {
+        NetworkKeys::derive(
+            &NetworkName::new(name).unwrap(),
+            &NetworkSecret::from_bytes(vec![4u8; 32]).unwrap(),
+        )
+        .network_id()
+    }
+
+    fn peer(seed: u8) -> EndpointId {
+        iroh::SecretKey::from_bytes(&[seed; 32]).public()
+    }
+
+    fn addr(last: u8) -> Ipv4Addr {
+        Ipv4Addr::new(10, 13, 37, last)
+    }
+
+    fn routes() -> NetworkRoutes {
+        NetworkRoutes {
+            exit: Default::default(),
+            broadcast: Default::default(),
+            range: Some("10.13.37.0/24".parse().unwrap()),
+            local: Some(addr(1)),
+            peers: vec![(addr(2), peer(2)), (addr(3), peer(3))],
+        }
+    }
+
+    #[test]
+    fn a_packet_goes_to_the_member_that_holds_its_destination() {
+        let table = RoutingTable::new();
+        let id = network("routing");
+        table.set_network(id, routes()).unwrap();
+
+        assert_eq!(
+            table.route(addr(2)),
+            Some(Route {
+                network: id,
+                peer: peer(2)
+            })
+        );
+        assert_eq!(
+            table.route(addr(3)),
+            Some(Route {
+                network: id,
+                peer: peer(3)
+            })
+        );
+    }
+
+    #[test]
+    fn a_packet_for_nobody_has_nowhere_to_go() {
+        // Dropped rather than sent to everybody: a tunnel is not a broadcast
+        // domain, and flooding would hand one member another's traffic.
+        let table = RoutingTable::new();
+        table.set_network(network("routing"), routes()).unwrap();
+        assert_eq!(table.route(addr(200)), None);
+        assert_eq!(table.route(Ipv4Addr::new(192, 0, 2, 1)), None);
+    }
+
+    #[test]
+    fn a_packet_for_ourselves_is_not_sent_to_a_peer() {
+        let table = RoutingTable::new();
+        table.set_network(network("routing"), routes()).unwrap();
+        assert_eq!(table.route(addr(1)), None, "that address is ours");
+    }
+
+    #[test]
+    fn a_peer_may_only_send_from_the_address_it_holds() {
+        let table = RoutingTable::new();
+        let id = network("ownership");
+        table.set_network(id, routes()).unwrap();
+
+        assert!(table.may_send_from(id, peer(2), addr(2)));
+        // Somebody else's address, and its own network's.
+        assert!(!table.may_send_from(id, peer(2), addr(3)));
+        assert!(!table.may_send_from(id, peer(2), addr(1)));
+        assert!(!table.may_send_from(id, peer(2), addr(99)));
+        // A peer that holds nothing here.
+        assert!(!table.may_send_from(id, peer(9), addr(2)));
+        // And the question is per network, not global.
+        assert!(!table.may_send_from(network("elsewhere"), peer(2), addr(2)));
+    }
+
+    #[test]
+    fn two_networks_with_overlapping_ranges_are_refused() {
+        // One interface means an address belongs to one network. Accepting
+        // both and guessing would send somebody's traffic to a stranger.
+        let table = RoutingTable::new();
+        let first = network("first");
+        let second = network("second");
+        table.set_network(first, routes()).unwrap();
+
+        let err = table.set_network(second, routes()).unwrap_err();
+        let RouteError::Overlap { other, .. } = err;
+        assert_eq!(other, first);
+        assert!(err.to_string().contains("stop the other network"), "{err}");
+        assert_eq!(table.len(), 1, "the overlapping network is not added");
+
+        // A range of its own is fine.
+        table
+            .set_network(
+                second,
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    peers: vec![(Ipv4Addr::new(10, 99, 0, 2), peer(4))],
+                },
+            )
+            .unwrap();
+        assert_eq!(table.len(), 2);
+        assert_eq!(
+            table.route(Ipv4Addr::new(10, 99, 0, 2)),
+            Some(Route {
+                network: second,
+                peer: peer(4)
+            })
+        );
+        // And the first network still answers for its own.
+        assert_eq!(table.route(addr(2)).map(|route| route.network), Some(first));
+    }
+
+    #[test]
+    fn the_host_broadcast_target_is_the_lowest_enabled_network_with_an_address() {
+        let table = RoutingTable::new();
+        assert_eq!(table.broadcast_host_target(), None);
+
+        let (low, high) = {
+            let (a, b) = (network("first"), network("second"));
+            if a < b { (a, b) } else { (b, a) }
+        };
+        let mk = |enabled: bool, base: [u8; 4], local: Option<Ipv4Addr>| NetworkRoutes {
+            exit: Default::default(),
+            broadcast: BroadcastPolicy {
+                enabled,
+                ..Default::default()
+            },
+            range: Some(
+                format!("{}.{}.{}.0/24", base[0], base[1], base[2])
+                    .parse()
+                    .unwrap(),
+            ),
+            local,
+            peers: Vec::new(),
+        };
+        let a = Ipv4Addr::new(10, 1, 0, 1);
+        let b = Ipv4Addr::new(10, 2, 0, 1);
+
+        // The lower id has broadcast off, so the higher one is chosen.
+        table
+            .set_network(low, mk(false, [10, 1, 0, 0], Some(a)))
+            .unwrap();
+        table
+            .set_network(high, mk(true, [10, 2, 0, 0], Some(b)))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((b, "10.2.0.0/24".parse().unwrap()))
+        );
+
+        // Both enabled: the lowest id wins, whatever the insertion order.
+        table
+            .set_network(low, mk(true, [10, 1, 0, 0], Some(a)))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((a, "10.1.0.0/24".parse().unwrap()))
+        );
+
+        // No local address yet: it cannot be the route's source.
+        table
+            .set_network(low, mk(true, [10, 1, 0, 0], None))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((b, "10.2.0.0/24".parse().unwrap()))
+        );
+
+        table.remove_network(high);
+        assert_eq!(table.broadcast_host_target(), None);
+    }
+
+    #[test]
+    fn a_range_can_be_asked_about_before_it_is_proposed() {
+        // The point of asking first: an agent that claims an address it
+        // cannot route also tells everybody else to use that range.
+        let table = RoutingTable::new();
+        let first = network("first");
+        let second = network("second");
+        table.set_network(first, routes()).unwrap();
+
+        let mine: Ipv4Range = "10.13.37.0/24".parse().unwrap();
+        assert_eq!(table.would_overlap(second, mine), Some(mine));
+        // Its own range is not a collision with itself.
+        assert_eq!(table.would_overlap(first, mine), None);
+        assert_eq!(
+            table.would_overlap(second, "10.99.0.0/16".parse().unwrap()),
+            None
+        );
+    }
+
+    #[test]
+    fn a_nested_range_counts_as_overlapping() {
+        let table = RoutingTable::new();
+        table
+            .set_network(
+                network("wide"),
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.0.0.0/8".parse().unwrap()),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        // Inside the /8, in either direction.
+        assert!(
+            table
+                .set_network(
+                    network("narrow"),
+                    NetworkRoutes {
+                        exit: Default::default(),
+                        broadcast: Default::default(),
+                        range: Some("10.13.37.0/24".parse().unwrap()),
+                        ..Default::default()
+                    }
+                )
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn replacing_a_network_does_not_count_as_overlapping_itself() {
+        let table = RoutingTable::new();
+        let id = network("refresh");
+        table.set_network(id, routes()).unwrap();
+
+        let mut updated = routes();
+        updated.peers.push((addr(4), peer(4)));
+        table.set_network(id, updated).unwrap();
+        assert_eq!(table.len(), 1);
+        assert!(table.route(addr(4)).is_some());
+    }
+
+    #[test]
+    fn forgetting_a_network_takes_its_addresses_with_it() {
+        let table = RoutingTable::new();
+        let id = network("leaving");
+        table.set_network(id, routes()).unwrap();
+        table.remove_network(id);
+
+        assert!(table.is_empty());
+        assert_eq!(table.route(addr(2)), None);
+        assert!(!table.may_send_from(id, peer(2), addr(2)));
+    }
+
+    #[test]
+    fn the_local_addresses_are_what_the_interface_should_carry() {
+        let table = RoutingTable::new();
+        table.set_network(network("one"), routes()).unwrap();
+        table
+            .set_network(
+                network("two"),
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    peers: Vec::new(),
+                },
+            )
+            .unwrap();
+
+        assert_eq!(
+            table.local_addresses(),
+            vec![(addr(1), 24), (Ipv4Addr::new(10, 99, 0, 1), 16)]
+        );
+    }
+
+    #[test]
+    fn a_network_with_no_address_yet_contributes_nothing() {
+        let table = RoutingTable::new();
+        let id = network("waiting");
+        table.set_network(id, NetworkRoutes::default()).unwrap();
+
+        assert!(table.local_addresses().is_empty());
+        assert_eq!(table.route(addr(2)), None);
+        assert_eq!(table.len(), 1, "the network is known, it just has nothing");
+    }
+
+    fn with_exit(offer: bool, via: Option<EndpointId>) -> NetworkRoutes {
+        NetworkRoutes {
+            exit: ExitPolicy { offer, via },
+            ..routes()
+        }
+    }
+
+    const INTERNET: Ipv4Addr = Ipv4Addr::new(93, 184, 216, 34);
+
+    #[test]
+    fn without_an_exit_node_the_internet_has_nowhere_to_go() {
+        let table = RoutingTable::new();
+        table.set_network(network("plain"), routes()).unwrap();
+        assert_eq!(table.route(INTERNET), None);
+        assert_eq!(table.exit_via(), None);
+    }
+
+    #[test]
+    fn with_an_exit_node_every_ordinary_destination_goes_to_it() {
+        let table = RoutingTable::new();
+        let id = network("client");
+        table
+            .set_network(id, with_exit(false, Some(peer(2))))
+            .unwrap();
+        let exit = Route {
+            network: id,
+            peer: peer(2),
+        };
+        assert_eq!(table.route(INTERNET), Some(exit));
+        assert_eq!(table.route(Ipv4Addr::new(192, 168, 1, 10)), Some(exit));
+        assert_eq!(table.exit_via(), Some((id, peer(2))));
+        // A member's own address still goes to that member.
+        assert_eq!(
+            table.route(addr(3)),
+            Some(Route {
+                network: id,
+                peer: peer(3)
+            })
+        );
+    }
+
+    #[test]
+    fn the_exit_node_is_never_the_answer_for_what_is_not_the_internet() {
+        let table = RoutingTable::new();
+        table
+            .set_network(network("client"), with_exit(false, Some(peer(2))))
+            .unwrap();
+        // Ourselves, a free address inside the overlay range, and everything
+        // that is not an ordinary unicast destination.
+        for destination in [
+            addr(1),
+            addr(200),
+            Ipv4Addr::new(127, 0, 0, 1),
+            Ipv4Addr::new(224, 0, 0, 251),
+            Ipv4Addr::new(239, 255, 255, 250),
+            Ipv4Addr::BROADCAST,
+            Ipv4Addr::UNSPECIFIED,
+            Ipv4Addr::new(169, 254, 1, 1),
+        ] {
+            assert_eq!(table.route(destination), None, "{destination}");
+        }
+    }
+
+    #[test]
+    fn another_networks_range_is_not_the_internet_either() {
+        let table = RoutingTable::new();
+        table
+            .set_network(network("client"), with_exit(false, Some(peer(2))))
+            .unwrap();
+        table
+            .set_network(
+                network("other"),
+                NetworkRoutes {
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        assert_eq!(table.route(Ipv4Addr::new(10, 99, 7, 7)), None);
+    }
+
+    #[test]
+    fn two_exit_choices_resolve_to_the_lowest_network_id() {
+        let table = RoutingTable::new();
+        let (a, b) = (network("one"), network("two"));
+        table
+            .set_network(a, with_exit(false, Some(peer(2))))
+            .unwrap();
+        table
+            .set_network(
+                b,
+                NetworkRoutes {
+                    exit: ExitPolicy {
+                        offer: false,
+                        via: Some(peer(9)),
+                    },
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        let (chosen, _) = table.exit_via().unwrap();
+        assert_eq!(chosen, a.min(b));
+        assert_eq!(table.route(INTERNET).unwrap().network, a.min(b));
+    }
+
+    #[test]
+    fn the_exit_peer_may_send_from_any_internet_address_but_nobody_else_may() {
+        let table = RoutingTable::new();
+        let id = network("client");
+        table
+            .set_network(id, with_exit(false, Some(peer(2))))
+            .unwrap();
+
+        // What comes back from the internet carries the internet's addresses.
+        assert!(table.may_send_from(id, peer(2), INTERNET));
+        // Another member still speaks only for its own address.
+        assert!(!table.may_send_from(id, peer(3), INTERNET));
+        assert!(table.may_send_from(id, peer(3), addr(3)));
+        // The exit peer cannot pass itself off as somebody else in the
+        // overlay, nor as this agent, nor as something that is not a source.
+        assert!(!table.may_send_from(id, peer(2), addr(3)));
+        assert!(!table.may_send_from(id, peer(2), addr(1)));
+        assert!(!table.may_send_from(id, peer(2), addr(200)));
+        for source in [
+            Ipv4Addr::LOCALHOST,
+            Ipv4Addr::UNSPECIFIED,
+            Ipv4Addr::new(224, 0, 0, 1),
+            Ipv4Addr::BROADCAST,
+        ] {
+            assert!(!table.may_send_from(id, peer(2), source), "{source}");
+        }
+        // Without choosing an exit node nobody gets the exception.
+        let plain = RoutingTable::new();
+        plain.set_network(id, routes()).unwrap();
+        assert!(!plain.may_send_from(id, peer(2), INTERNET));
+    }
+
+    #[test]
+    fn an_exit_node_takes_a_members_internet_packets_only_while_it_offers() {
+        let table = RoutingTable::new();
+        let id = network("server");
+        table.set_network(id, with_exit(false, None)).unwrap();
+        assert!(!table.accepts_exit_traffic(id, INTERNET));
+
+        table.set_exit(
+            id,
+            ExitPolicy {
+                offer: true,
+                via: None,
+            },
+        );
+        assert!(table.accepts_exit_traffic(id, INTERNET));
+        assert!(table.accepts_exit_traffic(id, Ipv4Addr::new(192, 168, 1, 10)));
+        // Never for another member, for this agent, a free overlay address,
+        // or anything that is not an ordinary unicast destination.
+        for destination in [
+            addr(1),
+            addr(2),
+            addr(200),
+            Ipv4Addr::LOCALHOST,
+            Ipv4Addr::new(224, 0, 0, 251),
+            Ipv4Addr::BROADCAST,
+            Ipv4Addr::UNSPECIFIED,
+        ] {
+            assert!(
+                !table.accepts_exit_traffic(id, destination),
+                "{destination}"
+            );
+        }
+        // An unknown network offers nothing.
+        assert!(!table.accepts_exit_traffic(network("elsewhere"), INTERNET));
+
+        table.set_exit(id, ExitPolicy::default());
+        assert!(!table.accepts_exit_traffic(id, INTERNET));
+    }
+
+    #[test]
+    fn offers_are_listed_by_network_with_the_range_each_agreed() {
+        let table = RoutingTable::new();
+        let (a, b) = (network("a"), network("b"));
+        table.set_network(a, with_exit(true, None)).unwrap();
+        table
+            .set_network(
+                b,
+                NetworkRoutes {
+                    exit: ExitPolicy {
+                        offer: true,
+                        via: None,
+                    },
+                    // No range agreed yet: nothing to masquerade.
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        assert_eq!(
+            table.exit_offers(),
+            vec![(a, "10.13.37.0/24".parse().unwrap())]
+        );
+        assert_eq!(
+            table.exit_settings(a),
+            Some((
+                ExitPolicy {
+                    offer: true,
+                    via: None
+                },
+                Some("10.13.37.0/24".parse().unwrap())
+            ))
+        );
+        assert_eq!(table.exit_settings(network("none")), None);
+    }
+
+    #[test]
+    fn setting_exit_for_a_network_with_no_entry_is_remembered_by_the_caller() {
+        // The runtime keeps the policy and passes it with the network when
+        // its range is agreed; a setter for a network the table does not know
+        // yet changes nothing.
+        let table = RoutingTable::new();
+        table.set_exit(
+            network("later"),
+            ExitPolicy {
+                offer: true,
+                via: None,
+            },
+        );
+        assert!(table.is_empty());
+    }
+}

crates/tsunagi-core/src/overlay/tun.rs

@@ -0,0 +1,569 @@
+//! The boundary to the operating system's packet interface.
+//!
+//! One agent has one of these, and it belongs to the agent rather than to
+//! any protocol: a packet arriving here is routed to whichever peer owns its
+//! destination, over whichever protocol has a link to that peer.
+//!
+//! Two implementations:
+//!
+//! * [`MemoryTun`] keeps packets in memory. It needs no privileges at all and
+//!   is what the test suite uses, so the entire data plane — handshake,
+//!   encryption, routing — is exercised without touching the host.
+//! * `SystemTun`, behind the `tun-device` feature, is a real TUN interface.
+//!   Creating one needs `CAP_NET_ADMIN`, and it is
+//!   [`provision`](super::provision) that holds that and creates it.
+
+use std::sync::Arc;
+
+use bytes::Bytes;
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+/// What a device should look like once created.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct TunRequest {
+    /// Interface name to ask for.
+    pub name: String,
+    /// Every overlay address this host answers to, and no others.
+    ///
+    /// One per network the agent is in — one agent has one interface, and
+    /// it carries them all. A list rather than one address because
+    /// carrying only the first is carrying only the first network: the
+    /// others have addresses the operating system has never heard of, and
+    /// their traffic goes nowhere.
+    ///
+    /// Allocated and signed at the system level, never derived from a
+    /// protocol's key: every protocol carries traffic for the same
+    /// addresses.
+    pub addresses: Vec<super::Cidr>,
+    /// Interface MTU.
+    pub mtu: u32,
+}
+
+impl TunRequest {
+    /// A request carrying nothing but a name and an MTU.
+    ///
+    /// Used where the addresses have already been applied to the host, so the
+    /// device itself only needs opening.
+    pub fn bare(name: impl Into<String>, mtu: u32) -> Self {
+        Self {
+            name: name.into(),
+            addresses: Vec::new(),
+            mtu,
+        }
+    }
+}
+
+/// A packet interface.
+///
+/// `recv` yields packets the operating system wants sent; `send` delivers
+/// packets that arrived from a peer.
+pub trait TunDevice: Send + Sync + std::fmt::Debug + 'static {
+    /// The interface name the operating system actually gave us.
+    fn name(&self) -> &str;
+
+    /// The interface MTU.
+    fn mtu(&self) -> u32;
+
+    /// The next packet the operating system wants to send, or `None` once the
+    /// device is gone.
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>>;
+
+    /// Delivers a packet to the operating system.
+    fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>>;
+}
+
+/// Creates packet interfaces.
+pub trait TunFactory: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Whether what this factory creates exists on the host.
+    ///
+    /// `false` for an in-memory device: it has a name and an MTU and nothing
+    /// else — no link, no addresses, nothing for the operating system to
+    /// route through or to attach a resolver setting to. Stated rather than
+    /// guessed from the name, because a name is not evidence: an in-memory
+    /// `tsun0` and a real `tsun0` belonging to another agent look identical
+    /// from here, and configuring the operating system for the first would
+    /// land on the second.
+    fn on_host(&self) -> bool;
+
+    /// Creates a device.
+    fn create<'a>(
+        &'a self,
+        request: TunRequest,
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>>;
+
+    /// Applies a changed request to an interface that already exists.
+    ///
+    /// The overlay IPv4 address is allocated at run time, so it can change
+    /// while the agent runs. A factory that manages the host applies that to
+    /// the live interface, without recreating it: recreating would drop every
+    /// tunnel riding on it.
+    ///
+    /// The default does nothing, which is right for a factory that only
+    /// attaches to an interface somebody else prepared.
+    fn reconfigure<'a>(&'a self, _request: TunRequest) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move { Ok(()) })
+    }
+
+    /// The host rules for this factory's interface, when it manages the host.
+    ///
+    /// `None` for an in-memory device or an interface somebody else prepared:
+    /// there is nothing of ours to route or to open in a firewall.
+    fn host_rules(&self) -> Option<Arc<dyn super::hostrules::BroadcastHostRules>> {
+        None
+    }
+
+    /// Removes an interface this factory created.
+    ///
+    /// Runs on the teardown path, so it reports rather than fails: there is
+    /// nothing useful to do about a failure at that point, and an interface
+    /// that is already gone is the desired outcome anyway.
+    fn destroy<'a>(&'a self, _name: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// An in-memory packet interface.
+///
+/// Nothing reaches the operating system. Packets the device "sends" can be
+/// read back with [`MemoryTun::pop_to_os`], and packets can be injected as if
+/// the operating system produced them with [`MemoryTun::push_from_os`].
+#[derive(Debug)]
+pub struct MemoryTun {
+    name: String,
+    mtu: u32,
+    from_os_tx: tokio::sync::mpsc::UnboundedSender<Bytes>,
+    from_os_rx: tokio::sync::Mutex<tokio::sync::mpsc::UnboundedReceiver<Bytes>>,
+    to_os_tx: tokio::sync::mpsc::UnboundedSender<Bytes>,
+    to_os_rx: tokio::sync::Mutex<tokio::sync::mpsc::UnboundedReceiver<Bytes>>,
+    /// Set once the device is gone, so a reader stops rather than waiting
+    /// for a packet that will never come.
+    ///
+    /// A `watch` rather than a `Notify`: closing must be seen by a reader
+    /// that has not started waiting yet as well as by one already waiting —
+    /// a device can be removed before the loop reading it has been polled
+    /// even once. The receiver is kept for the same reason: a `watch` send
+    /// with nobody subscribed does nothing at all.
+    closed: tokio::sync::watch::Sender<bool>,
+    closed_rx: tokio::sync::watch::Receiver<bool>,
+}
+
+impl MemoryTun {
+    /// Creates a device with the given name and MTU.
+    pub fn new(name: impl Into<String>, mtu: u32) -> Arc<Self> {
+        let (from_os_tx, from_os_rx) = tokio::sync::mpsc::unbounded_channel();
+        let (to_os_tx, to_os_rx) = tokio::sync::mpsc::unbounded_channel();
+        let (closed, closed_rx) = tokio::sync::watch::channel(false);
+        Arc::new(Self {
+            name: name.into(),
+            mtu,
+            from_os_tx,
+            from_os_rx: tokio::sync::Mutex::new(from_os_rx),
+            to_os_tx,
+            to_os_rx: tokio::sync::Mutex::new(to_os_rx),
+            closed,
+            closed_rx,
+        })
+    }
+
+    /// Marks the device gone, so the packet loop reading it ends.
+    ///
+    /// A real interface reports end of stream when it is removed; this is
+    /// how that happens here. Without it teardown waits out the whole grace
+    /// period for a reader that had no way of knowing.
+    pub fn close(&self) {
+        let _ = self.closed.send(true);
+    }
+
+    /// Injects a packet as if the operating system had produced it.
+    pub fn push_from_os(&self, packet: Bytes) {
+        let _ = self.from_os_tx.send(packet);
+    }
+
+    /// Takes the next packet the device delivered to the operating system.
+    pub async fn pop_to_os(&self) -> Option<Bytes> {
+        self.to_os_rx.lock().await.recv().await
+    }
+}
+
+impl TunDevice for MemoryTun {
+    fn name(&self) -> &str {
+        &self.name
+    }
+
+    fn mtu(&self) -> u32 {
+        self.mtu
+    }
+
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut closed = self.closed_rx.clone();
+            if *closed.borrow_and_update() {
+                return None;
+            }
+            let mut queue = self.from_os_rx.lock().await;
+            tokio::select! {
+                packet = queue.recv() => packet,
+                _ = closed.changed() => None,
+            }
+        })
+    }
+
+    fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            let _ = self.to_os_tx.send(packet);
+            Ok(())
+        })
+    }
+}
+
+/// Whether an address is assigned to some interface on this host.
+///
+/// Binding a UDP socket to a specific address only succeeds when the address
+/// is local, which makes this a cheap check that needs no privileges and no
+/// platform-specific code. It does not say *which* interface has it, which is
+/// enough here: the agent chose the address, so anything else holding it is a
+/// problem in its own right.
+pub fn address_is_local(address: std::net::IpAddr) -> bool {
+    if std::net::UdpSocket::bind((address, 0)).is_ok() {
+        return true;
+    }
+    #[cfg(all(feature = "tun-device", target_os = "windows"))]
+    {
+        if let Ok(interfaces) = std::panic::catch_unwind(netdev::get_interfaces) {
+            for iface in interfaces {
+                if iface.is_up() {
+                    match address {
+                        std::net::IpAddr::V4(v4) => {
+                            if iface.ipv4.iter().any(|ip| ip.addr() == v4) {
+                                return true;
+                            }
+                        }
+                        std::net::IpAddr::V6(v6) => {
+                            if iface.ipv6.iter().any(|ip| ip.addr() == v6) {
+                                return true;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    false
+}
+
+/// Creates [`MemoryTun`] devices.
+#[derive(Debug, Clone, Default)]
+pub struct MemoryTunFactory {
+    created: Arc<std::sync::Mutex<Vec<Arc<MemoryTun>>>>,
+}
+
+impl MemoryTunFactory {
+    /// Creates a factory.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// The device created for an interface name, if any.
+    pub fn device(&self, name: &str) -> Option<Arc<MemoryTun>> {
+        let guard = match self.created.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        guard
+            .iter()
+            .find(|device| device.name() == name)
+            .map(Arc::clone)
+    }
+
+    /// Every device created so far.
+    pub fn devices(&self) -> Vec<Arc<MemoryTun>> {
+        let guard = match self.created.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        guard.clone()
+    }
+}
+
+impl TunFactory for MemoryTunFactory {
+    fn name(&self) -> &str {
+        "memory"
+    }
+
+    /// Nothing here is on the host, which is the whole point of it.
+    fn on_host(&self) -> bool {
+        false
+    }
+
+    /// Closes the device so whatever is reading it stops.
+    ///
+    /// The device itself is kept, because a test asks what went through it
+    /// after the agent that owned it has gone.
+    fn destroy<'a>(&'a self, name: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let Some(device) = self.device(name) {
+                device.close();
+            }
+        })
+    }
+
+    fn create<'a>(
+        &'a self,
+        request: TunRequest,
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>> {
+        Box::pin(async move {
+            let device = MemoryTun::new(request.name, request.mtu);
+            let mut guard = match self.created.lock() {
+                Ok(guard) => guard,
+                Err(poisoned) => poisoned.into_inner(),
+            };
+            guard.push(Arc::clone(&device));
+            Ok(device as Arc<dyn TunDevice>)
+        })
+    }
+}
+
+#[cfg(feature = "tun-device")]
+pub(crate) use system::open_tun;
+
+#[cfg(feature = "tun-device")]
+mod system {
+    use std::sync::Arc;
+
+    use bytes::Bytes;
+    use tokio::sync::Mutex;
+
+    use super::{TunDevice, TunRequest};
+    use crate::BoxFuture;
+    use crate::overlay::OverlayError;
+
+    /// A real TUN interface.
+    ///
+    /// On Linux it is created by opening `/dev/net/tun`, which needs
+    /// `CAP_NET_ADMIN`; on Windows it is a Wintun adapter, which needs the
+    /// process to be elevated and `wintun.dll` to be reachable. Either way
+    /// [`open_tun`] is only ever called from
+    /// [`provision`](super::super::provision), which holds whatever privilege
+    /// the platform requires for the length of the call and no longer.
+    ///
+    /// It is deliberately **not** made persistent, so the operating system
+    /// removes the interface when this value is dropped — however the process
+    /// ends. (A Wintun adapter created this way is likewise torn down when the
+    /// last handle to it closes, which the reader and writer below hold.)
+    pub struct SystemTun {
+        name: String,
+        mtu: u32,
+        reader: Mutex<tokio::io::ReadHalf<tun::AsyncDevice>>,
+        writer: Mutex<tokio::io::WriteHalf<tun::AsyncDevice>>,
+    }
+
+    impl std::fmt::Debug for SystemTun {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            f.debug_struct("SystemTun")
+                .field("name", &self.name)
+                .field("mtu", &self.mtu)
+                .finish()
+        }
+    }
+
+    impl TunDevice for SystemTun {
+        fn name(&self) -> &str {
+            &self.name
+        }
+
+        fn mtu(&self) -> u32 {
+            self.mtu
+        }
+
+        fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+            Box::pin(async move {
+                use tokio::io::AsyncReadExt;
+                let mut buffer = vec![0u8; self.mtu as usize + 64];
+                let mut reader = self.reader.lock().await;
+                match reader.read(&mut buffer).await {
+                    Ok(0) => None,
+                    Ok(read) => {
+                        buffer.truncate(read);
+                        Some(Bytes::from(buffer))
+                    }
+                    Err(err) => {
+                        tracing::debug!(%err, "tun read failed");
+                        None
+                    }
+                }
+            })
+        }
+
+        fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>> {
+            Box::pin(async move {
+                use tokio::io::AsyncWriteExt;
+                let mut writer = self.writer.lock().await;
+                writer
+                    .write_all(&packet)
+                    .await
+                    .map_err(|err| OverlayError::Other(format!("tun write failed: {err}")))
+            })
+        }
+    }
+
+    /// Creates the TUN interface by opening it.
+    ///
+    /// Synchronous, and deliberately so: the caller holds a capability guard
+    /// across this call, and such a guard must not span an `await` because
+    /// Linux capabilities are per thread.
+    pub(crate) fn open_tun(request: &TunRequest) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        // macOS mutates neither field below (it neither names the interface nor
+        // sets a platform config), so `mut` is unused only there.
+        #[cfg_attr(target_os = "macos", allow(unused_mut))]
+        let mut config = tun::Configuration::default();
+        // macOS names its tunnels `utunN` and assigns the unit number itself.
+        // A name derived from the network id (`tsun…`) is not a legal utun
+        // name, so there the kernel is left to choose and the real name is read
+        // back below. Linux and Windows take the requested name verbatim.
+        #[cfg(not(target_os = "macos"))]
+        config.tun_name(&request.name);
+        #[cfg(target_os = "linux")]
+        config.platform_config(|platform| {
+            // The crate's own root check is not the check we want: this holds
+            // CAP_NET_ADMIN without being root. Whether the open succeeds is
+            // the honest answer.
+            platform.ensure_root_privileges(false);
+        });
+        // Packet information stays off, so reads and writes are raw IP
+        // packets. `ip tuntap add ... mode tun` also defaults to no packet
+        // information, so the flags match when attaching to one. The address
+        // and MTU are left off the configuration on purpose: the provisioner
+        // applies those to the interface separately, the same way on every
+        // platform, so the plan stays the single source of what it carries.
+
+        let device = tun::create_as_async(&config).map_err(|err| open_error(&request.name, err))?;
+
+        // The interface name the operating system actually gave us: the
+        // requested one everywhere except macOS, where the kernel chose the
+        // utun unit and so the real name has to be read back off the device.
+        #[cfg(target_os = "macos")]
+        let name = {
+            use tun::AbstractDevice;
+            device.tun_name().unwrap_or_else(|_| request.name.clone())
+        };
+        #[cfg(not(target_os = "macos"))]
+        let name = request.name.clone();
+
+        let (reader, writer) = tokio::io::split(device);
+        Ok(Arc::new(SystemTun {
+            name,
+            mtu: request.mtu,
+            reader: Mutex::new(reader),
+            writer: Mutex::new(writer),
+        }) as Arc<dyn TunDevice>)
+    }
+
+    fn open_error(name: &str, error: tun::Error) -> OverlayError {
+        // Wintun preserves the Windows code in its I/O variant. Normalize it
+        // before testing permissions; matching localized error text would
+        // miss the same failure on another Windows language.
+        #[cfg(target_os = "windows")]
+        let error = match error {
+            tun::Error::WintunError(error) => tun::Error::Io(error.into()),
+            error => error,
+        };
+        let denied = match &error {
+            tun::Error::Io(error) => {
+                error.kind() == std::io::ErrorKind::PermissionDenied
+                    || (cfg!(target_os = "windows")
+                        && matches!(error.raw_os_error(), Some(5 | 740 | 1314)))
+            }
+            _ => false,
+        };
+        if denied {
+            let hint = if cfg!(target_os = "windows") {
+                "Open PowerShell or Command Prompt with 'Run as administrator' and \
+                 start `tsng up` there."
+            } else if cfg!(target_os = "linux") {
+                "Start `tsng up` as root or grant CAP_NET_ADMIN with \
+                 `sudo setcap cap_net_admin,cap_net_bind_service+p /path/to/tsunagi`."
+            } else if cfg!(target_os = "macos") {
+                "Creating a utun interface needs root on macOS; start `tsng up` \
+                 with `sudo`, or install it as a root LaunchDaemon."
+            } else {
+                "Start `tsng up` with the privileges required to create a TUN interface."
+            };
+            return OverlayError::Unavailable(format!(
+                "permission denied while creating the TUN interface `{name}`. \
+                 {hint} OS error: {error}"
+            ));
+        }
+        OverlayError::Unavailable(format!(
+            "cannot create the TUN interface `{name}`: {error}. {}",
+            open_hint()
+        ))
+    }
+
+    /// The platform-appropriate tail of a "cannot create the interface" error:
+    /// what privilege it needs and the fallback that always works.
+    fn open_hint() -> &'static str {
+        #[cfg(target_os = "linux")]
+        {
+            "Creating one needs CAP_NET_ADMIN; grant it with `setcap cap_net_admin,cap_net_bind_service+p`, \
+             or run with `--no-tun` to keep the tunnels off the operating system."
+        }
+        #[cfg(target_os = "windows")]
+        {
+            "Creating one needs an elevated process and `wintun.dll` on the search path; \
+             run as Administrator with the DLL beside the executable, or run with `--no-tun` \
+             to keep the tunnels off the operating system."
+        }
+        #[cfg(target_os = "macos")]
+        {
+            "Creating a utun interface needs root; start the agent with `sudo` or as a root \
+             LaunchDaemon, or run with `--no-tun` to keep the tunnels off the operating system."
+        }
+        #[cfg(not(any(target_os = "linux", target_os = "windows", target_os = "macos")))]
+        {
+            "Run with `--no-tun` to keep the tunnels off the operating system."
+        }
+    }
+
+    #[cfg(test)]
+    mod tests {
+        use super::*;
+
+        #[test]
+        fn a_tun_permission_error_explains_how_to_start_the_agent() {
+            let error = tun::Error::Io(std::io::Error::from(std::io::ErrorKind::PermissionDenied));
+            let message = open_error("tsun0", error).to_string();
+            assert!(message.contains("permission denied"), "{message}");
+            assert!(message.contains("tsun0"), "{message}");
+            assert!(message.contains("tsng up"), "{message}");
+            #[cfg(target_os = "windows")]
+            assert!(message.contains("Run as administrator"), "{message}");
+            #[cfg(target_os = "linux")]
+            assert!(message.contains("CAP_NET_ADMIN"), "{message}");
+        }
+
+        #[cfg(target_os = "windows")]
+        #[test]
+        fn wintun_access_and_elevation_errors_get_the_permission_hint() {
+            for code in [5, 740, 1314] {
+                let error = tun::Error::WintunError(std::io::Error::from_raw_os_error(code).into());
+                let message = open_error("tsun0", error).to_string();
+                assert!(message.contains("permission denied"), "{message}");
+                assert!(message.contains("Run as administrator"), "{message}");
+                assert!(message.contains(&format!("os error {code}")), "{message}");
+                assert!(!message.contains("wintun.dll"), "{message}");
+            }
+        }
+
+        #[test]
+        fn other_tun_errors_keep_their_cause_without_claiming_access_was_denied() {
+            let message = open_error("tsun0", tun::Error::InvalidName).to_string();
+            assert!(message.contains("invalid device tun name"), "{message}");
+            assert!(!message.contains("permission denied"), "{message}");
+        }
+    }
+}

crates/tsunagi-core/src/proto/message.rs

@@ -0,0 +1,395 @@
+//! Control message formats.
+//!
+//! This is deliberately a small, closed set of messages, not a general RPC
+//! framework. Bodies are encoded with [postcard], a compact, deterministic,
+//! non-self-describing serde format.
+//!
+//! Every message that arrives from the network passes [`validate`] against the
+//! configured [`Limits`] before it reaches anything else.
+//!
+//! [postcard]: https://docs.rs/postcard
+
+use serde::{Deserialize, Serialize};
+
+use crate::config::Limits;
+use crate::dataplane::{MAX_PROTOCOL_ID_LEN, PluginCapability};
+use crate::error::ProtocolError;
+
+/// ALPN of the tsunagi control plane.
+///
+/// The version in the ALPN is the wire-compatibility version of the control
+/// protocol. It is independent of the network identity scheme version, so
+/// bumping it must not change any existing [`crate::NetworkId`].
+pub const ALPN: &[u8] = b"tsunagi/ctrl/8";
+
+/// ALPN of the tsunagi data plane.
+///
+/// Data plane connections are deliberately separate from control plane ones.
+/// They carry one IP plugin's packets for one network and nothing else, so a
+/// saturated or broken data plane cannot disturb control traffic, and the
+/// transport underneath can be replaced without touching the control protocol.
+///
+/// Version 4 adds source, destination, hop limit and stable flow id above
+/// transport fragmentation. Version 5 adds the end-to-end protocol to that
+/// envelope, so a tunnel no longer depends on the kind of link under it, and
+/// a first-claimed time to a signed hostname claim. All members must upgrade
+/// together. Persistent identities and network configuration do not change.
+pub const DATA_ALPN: &[u8] = b"tsunagi/data/5";
+
+/// Largest plugin protocol identifier accepted when opening a data channel.
+pub const MAX_DATA_PROTOCOL_LEN: usize = 32;
+
+/// Largest accepted signature on a signed record, in bytes.
+pub const MAX_SIGNATURE_LEN: usize = 64;
+
+/// Control protocol version carried inside the handshake.
+pub const PROTOCOL_VERSION: u16 = 8;
+
+pub use crate::dataplane::transport::{MAX_TRANSPORT_PAYLOAD, TransportPayload};
+
+/// First message of the handshake, sent by the initiator.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct Hello {
+    /// Control protocol version the initiator speaks.
+    pub version: u16,
+    /// Public network identifier the initiator wants to join.
+    pub network_id: [u8; 32],
+    /// Initiator's fresh handshake nonce.
+    pub nonce: [u8; 16],
+}
+
+/// Responder's reply to [`Hello`]. Carries no proof yet.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct HelloAck {
+    /// Control protocol version the responder speaks.
+    pub version: u16,
+    /// Responder's fresh handshake nonce.
+    pub nonce: [u8; 16],
+}
+
+/// A handshake proof, i.e. one HMAC tag over a role-specific transcript.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct AuthProof {
+    /// HMAC-SHA256 tag.
+    pub proof: [u8; 32],
+}
+
+/// What this agent tells a peer about itself.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct Announcement {
+    /// Local opt-in to receive LAN broadcasts in this network.
+    pub broadcast: bool,
+    /// This agent offers itself as an exit node in this network: members may
+    /// send all their internet traffic through it. Off unless its owner
+    /// turned it on, and only said once the rules that make it work are in
+    /// place.
+    pub exit_node: bool,
+    /// Human-readable hostname. A mutable binding, not an identity.
+    pub hostname: String,
+    /// Announced IP plugin capabilities. Opaque to the core.
+    pub capabilities: Vec<PluginCapability>,
+}
+
+/// Opens a data channel, sent by the initiator right after the membership
+/// handshake on a [`DATA_ALPN`] connection.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DataOpen {
+    /// Which IP plugin's packets this channel will carry.
+    pub protocol: String,
+    /// Maximum reassembled payload the initiator is willing to receive.
+    pub max_datagram: u32,
+}
+
+/// The responder's answer to [`DataOpen`].
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DataOpenAck {
+    /// Whether the channel was accepted.
+    ///
+    /// A channel is refused when the responder has no plugin for that
+    /// protocol in that network. That is an ordinary outcome, not an error.
+    pub accepted: bool,
+    /// Largest datagram the responder is willing to receive, in bytes.
+    pub max_datagram: u32,
+}
+
+/// A control message exchanged after a successful handshake.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[non_exhaustive]
+pub enum ControlMessage {
+    /// Private transport negotiation; never persisted or rebroadcast.
+    Transport {
+        /// Transport protocol id; the payload stays opaque to the core.
+        protocol: String,
+        /// Secret negotiation data, redacted in diagnostics.
+        payload: TransportPayload,
+    },
+    /// Hostname and capability announcement.
+    Announce(Announcement),
+    /// A small request used to verify that the exchange works.
+    Ping {
+        /// Caller-chosen sequence number, echoed back.
+        seq: u64,
+        /// Opaque bounded payload, echoed back.
+        payload: Vec<u8>,
+    },
+    /// The reply to a [`ControlMessage::Ping`].
+    Pong {
+        /// Sequence number of the request being answered.
+        seq: u64,
+        /// Echoed payload.
+        payload: Vec<u8>,
+    },
+    /// A snapshot of signed records this agent holds for the network.
+    ///
+    /// A snapshot is merged into what the receiver already has, never
+    /// substituted for it: an author missing from the batch is left alone,
+    /// because absence is not deletion.
+    State {
+        /// The records. Bounded by [`crate::config::Limits::max_state_records`].
+        records: Vec<crate::state::SignedRecord>,
+    },
+    /// Members this sender knows of, and where it has seen them.
+    ///
+    /// An introduction, not a vouching: everything here is an unverified
+    /// candidate, and membership is still decided by the handshake. It is
+    /// what turns a star around whoever was named on the command line into
+    /// a mesh — a device that was told about one member ends up talking to
+    /// all of them — and it is the same shape a lookup in a distributed
+    /// hash table would return.
+    Peers {
+        /// The members, with whatever addresses the sender has for them.
+        peers: Vec<PeerHint>,
+    },
+    /// Which peers this sender has a live data link with, right now.
+    ///
+    /// First-hand and nothing else: a sender speaks only for itself, never
+    /// about what somebody else can reach. That is what makes it usable
+    /// without weighing hearsay — the claim is proved or disproved by
+    /// sending through it.
+    ///
+    /// A snapshot rather than a change, because a snapshot is idempotent
+    /// and repairs itself; and volatile, so it lives here and not in
+    /// signed state, which is for what has to survive a member being away.
+    Reachable {
+        /// Direct links, scoped to the plugin protocol they actually carry.
+        links: Vec<ReachableLink>,
+    },
+    /// Graceful goodbye.
+    ///
+    /// A peer going away is not a revocation of anything.
+    Bye {
+        /// Short free-text reason.
+        reason: String,
+    },
+    /// One complete signed VIP field snapshot, merged atomically.
+    Vips {
+        /// Signed registers bounded by the per-VIP field limit.
+        entries: Vec<crate::vip::VipEntry>,
+    },
+    /// Signed cleanup markers, each proving the retired claim version.
+    Retirements {
+        /// Bounded by the per-message signed-state limit.
+        markers: Vec<crate::state::SignedRetirement>,
+    },
+}
+
+/// Somewhere a member has been seen.
+///
+/// Addresses are in iroh's own text form, which is what the endpoint takes
+/// back: this is a hint to try, never a fact about where anybody is.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PeerHint {
+    /// The member's endpoint key.
+    pub endpoint: [u8; 32],
+    /// Where the sender has seen it, as `ip:<socket>` or `relay:<url>`.
+    ///
+    /// The same spelling the disposable cache uses, so one decoder serves
+    /// both and neither can drift from the other. An empty list is still
+    /// worth sending: an id alone is enough for a receiver whose endpoint
+    /// can resolve it.
+    pub addrs: Vec<String>,
+}
+
+/// One directly observed edge, advertised only by its authenticated source.
+#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub struct ReachableLink {
+    /// The other end of this transport link.
+    pub peer: [u8; 32],
+    /// Protocol carried by the link.
+    pub protocol: String,
+}
+
+/// Longest address text accepted in a hint.
+pub const MAX_HINT_ADDR_LEN: usize = 128;
+
+/// Most addresses accepted for one member in a hint.
+pub const MAX_HINT_ADDRS: usize = 8;
+
+/// A control message together with the network it belongs to.
+///
+/// Every session is bound to exactly one network at handshake time. The
+/// `network_id` here is re-checked on every message, so an authenticated
+/// session for network A can never be used to speak to network B, even over a
+/// shared physical connection.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct Envelope {
+    /// Network this message belongs to.
+    pub network_id: [u8; 32],
+    /// The message itself.
+    pub message: ControlMessage,
+}
+
+/// Encodes a value into a postcard byte vector.
+pub fn encode<T: Serialize>(value: &T) -> Result<Vec<u8>, ProtocolError> {
+    postcard::to_stdvec(value).map_err(|_| ProtocolError::Malformed("cannot encode message"))
+}
+
+/// Decodes a value from postcard bytes.
+pub fn decode<T: for<'de> Deserialize<'de>>(bytes: &[u8]) -> Result<T, ProtocolError> {
+    postcard::from_bytes(bytes).map_err(|_| ProtocolError::Malformed("cannot decode message"))
+}
+
+fn check_len(field: &'static str, len: usize, limit: usize) -> Result<(), ProtocolError> {
+    if len > limit {
+        return Err(ProtocolError::FieldTooLarge { field, len, limit });
+    }
+    Ok(())
+}
+
+/// Validates a decoded capability against the configured limits.
+pub fn validate_capability(
+    capability: &PluginCapability,
+    limits: &Limits,
+) -> Result<(), ProtocolError> {
+    if capability.protocol.is_empty() {
+        return Err(ProtocolError::Malformed("empty plugin protocol id"));
+    }
+    check_len(
+        "capability.protocol",
+        capability.protocol.len(),
+        MAX_PROTOCOL_ID_LEN,
+    )?;
+    check_len(
+        "capability.data",
+        capability.data.len(),
+        limits.max_capability_data_len,
+    )?;
+    Ok(())
+}
+
+/// Validates a decoded control message against the configured limits.
+///
+/// Returning an error rejects that single message. It never stops the session's
+/// network, the other networks or the agent.
+pub fn validate(message: &ControlMessage, limits: &Limits) -> Result<(), ProtocolError> {
+    match message {
+        ControlMessage::Vips { entries } => {
+            check_len("vip.fields", entries.len(), limits.max_vip_fields)?;
+            for e in entries {
+                check_len("vip.id", e.vip.len(), 32)?;
+                check_len("vip.key", e.key.len(), 160)?;
+                check_len("vip.value", e.value.as_ref().map_or(0, Vec::len), 256)?;
+                check_len("vip.signature", e.signature.len(), 64)?;
+            }
+        }
+        ControlMessage::Retirements { markers } => {
+            check_len("retirements", markers.len(), limits.max_state_records)?;
+            for marker in markers {
+                check_len(
+                    "retirement.signature",
+                    marker.signature.len(),
+                    MAX_SIGNATURE_LEN,
+                )?;
+                check_len(
+                    "retirement.target.signature",
+                    marker.target.signature.len(),
+                    MAX_SIGNATURE_LEN,
+                )?;
+                if let Some(name) = marker.target.body.hostname() {
+                    check_len(
+                        "retirement.target.hostname",
+                        name.len(),
+                        limits.max_hostname_len,
+                    )?;
+                }
+            }
+        }
+        ControlMessage::Transport { protocol, payload } => {
+            if protocol.is_empty() {
+                return Err(ProtocolError::Malformed("empty transport protocol"));
+            }
+            check_len("transport.protocol", protocol.len(), MAX_PROTOCOL_ID_LEN)?;
+            check_len(
+                "transport.payload",
+                payload.as_bytes().len(),
+                limits.max_transport_payload_len.min(MAX_TRANSPORT_PAYLOAD),
+            )?;
+        }
+        ControlMessage::Peers { peers } => {
+            check_len("peers", peers.len(), limits.max_state_records)?;
+            for hint in peers {
+                check_len("peers.addrs", hint.addrs.len(), MAX_HINT_ADDRS)?;
+                for addr in &hint.addrs {
+                    check_len("peers.addr", addr.len(), MAX_HINT_ADDR_LEN)?;
+                }
+            }
+        }
+        ControlMessage::Reachable { links } => {
+            check_len("reachable.links", links.len(), limits.max_state_records)?;
+            for link in links {
+                if link.protocol.is_empty() {
+                    return Err(ProtocolError::Malformed("empty routing protocol"));
+                }
+                check_len(
+                    "reachable.protocol",
+                    link.protocol.len(),
+                    MAX_PROTOCOL_ID_LEN,
+                )?;
+            }
+        }
+        ControlMessage::Announce(announcement) => {
+            check_len(
+                "announce.hostname",
+                announcement.hostname.len(),
+                limits.max_hostname_len,
+            )?;
+            check_len(
+                "announce.capabilities",
+                announcement.capabilities.len(),
+                limits.max_capabilities,
+            )?;
+            for capability in &announcement.capabilities {
+                validate_capability(capability, limits)?;
+            }
+        }
+        ControlMessage::Ping { payload, .. } | ControlMessage::Pong { payload, .. } => {
+            check_len("echo.payload", payload.len(), limits.max_echo_payload_len)?;
+        }
+        ControlMessage::State { records } => {
+            check_len("state.records", records.len(), limits.max_state_records)?;
+            for record in records {
+                check_len("state.signature", record.signature.len(), MAX_SIGNATURE_LEN)?;
+            }
+        }
+        ControlMessage::Bye { reason } => {
+            check_len("bye.reason", reason.len(), limits.max_reason_len)?;
+        }
+    }
+    Ok(())
+}
+
+/// A short, stable label for a message kind, for metrics and diagnostics.
+pub fn kind(message: &ControlMessage) -> &'static str {
+    match message {
+        ControlMessage::Transport { .. } => "transport",
+        ControlMessage::Announce(_) => "announce",
+        ControlMessage::Ping { .. } => "ping",
+        ControlMessage::Pong { .. } => "pong",
+        ControlMessage::State { .. } => "state",
+        ControlMessage::Retirements { .. } => "retirements",
+        ControlMessage::Vips { .. } => "vips",
+        ControlMessage::Reachable { .. } => "reachable",
+        ControlMessage::Peers { .. } => "peers",
+        ControlMessage::Bye { .. } => "bye",
+    }
+}

crates/tsunagi-core/src/state/allocator.rs

@@ -0,0 +1,284 @@
+//! Choosing a free overlay address.
+//!
+//! Allocation, not derivation. Derivation needs no coordination but cannot
+//! avoid collisions in a space as small as IPv4; allocation avoids them but
+//! has to look at what everybody else already holds. The signed records in
+//! [`super`] are what makes that possible without a coordinator.
+//!
+//! The rules:
+//!
+//! * an address a participant already holds is kept, because stability across
+//!   an absence is the whole point;
+//! * otherwise the search starts at a position derived from the participant's
+//!   own identity, so two participants joining at once rarely start in the
+//!   same place;
+//! * the search then walks the range, so a free address is found whenever one
+//!   exists.
+
+use std::collections::HashSet;
+use std::net::Ipv4Addr;
+
+use iroh::EndpointId;
+use sha2::{Digest, Sha256};
+
+use super::Ipv4Range;
+use crate::identity::NetworkId;
+
+/// Why no address could be allocated.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+#[non_exhaustive]
+pub enum AllocationError {
+    /// Every address in the range is taken.
+    #[error("the overlay range {range} is full: {holders} of {usable} addresses are taken")]
+    RangeFull {
+        /// The range that is full.
+        range: Ipv4Range,
+        /// How many are held.
+        holders: usize,
+        /// How many the range has.
+        usable: u64,
+    },
+    /// The range has no usable host addresses.
+    #[error("the overlay range {range} has no room for hosts")]
+    NoRoom {
+        /// The offending range.
+        range: Ipv4Range,
+    },
+}
+
+/// How many host addresses a range holds, excluding network and broadcast.
+pub fn usable_addresses(range: Ipv4Range) -> u64 {
+    let host_bits = 32u32.saturating_sub(u32::from(range.prefix_len));
+    if host_bits < 2 {
+        return 0;
+    }
+    (1u64 << host_bits) - 2
+}
+
+/// The nth host address of a range.
+fn address_at(range: Ipv4Range, offset: u64) -> Ipv4Addr {
+    let host_bits = 32u32.saturating_sub(u32::from(range.prefix_len));
+    let mask = if host_bits >= 32 {
+        0
+    } else {
+        u32::MAX << host_bits
+    };
+    let network_part = u32::from(range.base) & mask;
+    // Offsets run 1..=usable, so the network address is never handed out.
+    Ipv4Addr::from(network_part | ((offset % (1u64 << host_bits)) as u32))
+}
+
+/// Picks an address for `author`, keeping `current` if it is still usable.
+///
+/// `taken` is what every other participant is known to hold.
+pub fn allocate(
+    network: NetworkId,
+    author: EndpointId,
+    range: Ipv4Range,
+    taken: &HashSet<Ipv4Addr>,
+    current: Option<Ipv4Addr>,
+) -> Result<Ipv4Addr, AllocationError> {
+    let usable = usable_addresses(range);
+    if usable == 0 {
+        return Err(AllocationError::NoRoom { range });
+    }
+
+    // Keeping what we already hold is what lets a participant come back to
+    // the same address after any length of absence.
+    if let Some(current) = current
+        && range.contains(current)
+        && !taken.contains(&current)
+    {
+        return Ok(current);
+    }
+
+    // Start somewhere derived from who we are, so two newcomers do not both
+    // begin at the first address and collide every time.
+    let mut hash = Sha256::new();
+    hash.update(b"tsunagi-ipv4-allocation-v1");
+    hash.update(network.as_bytes());
+    hash.update(author.as_bytes());
+    let digest = hash.finalize();
+    let seed = u64::from_be_bytes([
+        digest[0], digest[1], digest[2], digest[3], digest[4], digest[5], digest[6], digest[7],
+    ]);
+
+    for step in 0..usable {
+        let offset = ((seed.wrapping_add(step)) % usable) + 1;
+        let candidate = address_at(range, offset);
+        if !taken.contains(&candidate) {
+            return Ok(candidate);
+        }
+    }
+
+    Err(AllocationError::RangeFull {
+        range,
+        holders: taken.len(),
+        usable,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+    use iroh::SecretKey;
+
+    fn network(name: &str) -> NetworkId {
+        NetworkKeys::derive(
+            &NetworkName::new(name).unwrap(),
+            &NetworkSecret::from_bytes(vec![7u8; 32]).unwrap(),
+        )
+        .network_id()
+    }
+
+    fn slash24() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    #[test]
+    fn a_range_reports_its_usable_size() {
+        assert_eq!(usable_addresses(slash24()), 254);
+        assert_eq!(usable_addresses("10.0.0.0/16".parse().unwrap()), 65534);
+        assert_eq!(usable_addresses("10.0.0.0/30".parse().unwrap()), 2);
+        // The type refuses a /31, but the function is defensive anyway.
+        assert!("10.0.0.0/31".parse::<Ipv4Range>().is_err());
+        assert_eq!(
+            usable_addresses(Ipv4Range {
+                base: "10.0.0.0".parse().unwrap(),
+                prefix_len: 31
+            }),
+            0
+        );
+    }
+
+    #[test]
+    fn an_allocated_address_is_inside_the_range_and_not_its_edges() {
+        let id = network("inside");
+        let taken = HashSet::new();
+        for _ in 0..64 {
+            let author = SecretKey::generate().public();
+            let address = allocate(id, author, slash24(), &taken, None).unwrap();
+            assert!(slash24().contains(address));
+            assert_ne!(address.octets()[3], 0, "never the network address");
+            assert_ne!(address.octets()[3], 255, "never the broadcast address");
+        }
+    }
+
+    #[test]
+    fn an_address_already_held_is_kept() {
+        let id = network("sticky");
+        let author = SecretKey::generate().public();
+        let mine: Ipv4Addr = "10.13.37.42".parse().unwrap();
+
+        // This is what lets a participant return to the same address.
+        let taken = HashSet::new();
+        assert_eq!(
+            allocate(id, author, slash24(), &taken, Some(mine)).unwrap(),
+            mine
+        );
+
+        // Unless somebody else took it while we were away.
+        let taken = HashSet::from([mine]);
+        assert_ne!(
+            allocate(id, author, slash24(), &taken, Some(mine)).unwrap(),
+            mine
+        );
+
+        // Or unless the range changed under us.
+        let elsewhere: Ipv4Range = "10.99.0.0/16".parse().unwrap();
+        let moved = allocate(id, author, elsewhere, &HashSet::new(), Some(mine)).unwrap();
+        assert!(elsewhere.contains(moved));
+    }
+
+    #[test]
+    fn allocation_is_deterministic_and_spread_out() {
+        let id = network("spread");
+        // Fixed keys, not random ones. With 40 random authors in a /24 the
+        // birthday problem alone makes a handful of collisions likely, so a
+        // threshold on the count was a coin flip rather than a property.
+        let authors: Vec<_> = (0..40u8)
+            .map(|seed| SecretKey::from_bytes(&[seed; 32]).public())
+            .collect();
+
+        let first: Vec<_> = authors
+            .iter()
+            .map(|author| allocate(id, *author, slash24(), &HashSet::new(), None).unwrap())
+            .collect();
+        let again: Vec<_> = authors
+            .iter()
+            .map(|author| allocate(id, *author, slash24(), &HashSet::new(), None).unwrap())
+            .collect();
+        assert_eq!(first, again, "the same inputs give the same answer");
+
+        // Starting points are spread, so concurrent newcomers rarely clash.
+        let distinct: HashSet<_> = first.iter().collect();
+        assert!(
+            distinct.len() >= 35,
+            "only {} distinct starting points out of 40",
+            distinct.len()
+        );
+    }
+
+    #[test]
+    fn the_search_walks_past_everything_taken() {
+        let id = network("crowded");
+        let author = SecretKey::generate().public();
+
+        // Everything taken except one address.
+        let free: Ipv4Addr = "10.13.37.200".parse().unwrap();
+        let taken: HashSet<Ipv4Addr> = (1..=254u8)
+            .map(|host| Ipv4Addr::new(10, 13, 37, host))
+            .filter(|addr| *addr != free)
+            .collect();
+        assert_eq!(allocate(id, author, slash24(), &taken, None).unwrap(), free);
+    }
+
+    #[test]
+    fn a_full_range_is_an_error_rather_than_a_duplicate() {
+        let id = network("full");
+        let author = SecretKey::generate().public();
+        let taken: HashSet<Ipv4Addr> = (1..=254u8)
+            .map(|host| Ipv4Addr::new(10, 13, 37, host))
+            .collect();
+
+        assert!(matches!(
+            allocate(id, author, slash24(), &taken, None),
+            Err(AllocationError::RangeFull { .. })
+        ));
+        assert!(matches!(
+            allocate(
+                id,
+                author,
+                Ipv4Range {
+                    base: "10.0.0.0".parse().unwrap(),
+                    prefix_len: 31
+                },
+                &HashSet::new(),
+                None
+            ),
+            Err(AllocationError::NoRoom { .. })
+        ));
+    }
+
+    #[test]
+    fn every_address_in_a_small_range_can_be_handed_out() {
+        let id = network("exhaustive");
+        let small: Ipv4Range = "10.13.37.0/29".parse().unwrap();
+        let mut taken = HashSet::new();
+        let mut handed = Vec::new();
+
+        for _ in 0..usable_addresses(small) {
+            let author = SecretKey::generate().public();
+            let address = allocate(id, author, small, &taken, None).unwrap();
+            assert!(taken.insert(address), "handed out {address} twice");
+            handed.push(address);
+        }
+        assert_eq!(handed.len(), 6);
+        // And then it is genuinely full.
+        let author = SecretKey::generate().public();
+        assert!(allocate(id, author, small, &taken, None).is_err());
+    }
+}

crates/tsunagi-core/src/state/retirement.rs

@@ -0,0 +1,219 @@
+//! A member may retire an observed claim, never another identity's future.
+use iroh::{EndpointId, SecretKey, Signature};
+use serde::{Deserialize, Serialize};
+
+use super::{RecordBody, SignedRecord, StateError, push_lp};
+use crate::NetworkId;
+
+/// Maximum encoded retirement size, including its bounded claim proof.
+pub const MAX_RETIREMENT_LEN: usize = 1024;
+
+/// A network member's signed retirement of a proved claim version.
+///
+/// The original author's signed claim proves the cutoff really existed.
+/// Newer claims remain valid. This is neither a forged Release nor a ban.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct SignedRetirement {
+    /// Device identity of the member requesting cleanup.
+    pub author: [u8; 32],
+    /// Network in which cleanup applies.
+    pub network: [u8; 32],
+    /// Requester's durable author counter, committed with this marker.
+    pub version: u64,
+    /// The target's original signed claim; bounds the retired version.
+    pub target: SignedRecord,
+    /// Requester's Ed25519 signature over this marker and its proof.
+    pub signature: Vec<u8>,
+}
+
+impl SignedRetirement {
+    fn canonical_bytes(&self) -> Vec<u8> {
+        let mut out = Vec::with_capacity(768);
+        push_lp(&mut out, b"tsunagi-node-retirement/v1");
+        push_lp(&mut out, &self.network);
+        push_lp(&mut out, &self.author);
+        push_lp(&mut out, &self.version.to_be_bytes());
+        // Bind every field of the original signed claim, including its signature.
+        push_lp(&mut out, &self.target.network);
+        push_lp(&mut out, &self.target.author);
+        push_lp(&mut out, &self.target.version.to_be_bytes());
+        push_lp(&mut out, &self.target.body.canonical());
+        push_lp(&mut out, &self.target.signature);
+        out
+    }
+
+    /// Signs a marker with the requesting member's persistent device key.
+    pub fn sign(
+        secret: &SecretKey,
+        network: NetworkId,
+        version: u64,
+        target: SignedRecord,
+    ) -> Self {
+        let mut marker = Self {
+            author: *secret.public().as_bytes(),
+            network: *network.as_bytes(),
+            version,
+            target,
+            signature: Vec::new(),
+        };
+        marker.signature = secret.sign(&marker.canonical_bytes()).to_bytes().to_vec();
+        marker
+    }
+
+    /// Verifies both authors, scope and a finite, genuinely signed cutoff.
+    pub fn verify(&self, network: NetworkId) -> Result<EndpointId, StateError> {
+        if self.network != *network.as_bytes() {
+            return Err(StateError::WrongNetwork);
+        }
+        if self.version == 0 || self.version > i64::MAX as u64 {
+            return Err(StateError::Malformed("retirement version is out of range"));
+        }
+        if !matches!(self.target.body, RecordBody::Claim { .. }) {
+            return Err(StateError::Malformed("only a claim can be retired"));
+        }
+        let target = self.target.verify(network)?;
+        let author = EndpointId::from_bytes(&self.author).map_err(|_| StateError::BadAuthor)?;
+        if author == target {
+            return Err(StateError::Malformed(
+                "use Release to retire your own claim",
+            ));
+        }
+        let raw: [u8; Signature::LENGTH] =
+            self.signature
+                .as_slice()
+                .try_into()
+                .map_err(|_| StateError::BadSignatureLength {
+                    expected: Signature::LENGTH,
+                })?;
+        author
+            .verify(&self.canonical_bytes(), &Signature::from_bytes(&raw))
+            .map_err(|_| StateError::BadSignature)?;
+        Ok(target)
+    }
+
+    /// A deterministic strongest marker per target, independent of arrival order.
+    pub(crate) fn supersedes(&self, other: &Self) -> bool {
+        self.target.version > other.target.version
+            || (self.target.version == other.target.version
+                && (self.canonical_bytes(), &self.signature)
+                    < (other.canonical_bytes(), &other.signature))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::state::{Ipv4Range, StateSet};
+
+    fn claim(key: &SecretKey, id: NetworkId, version: u64) -> SignedRecord {
+        SignedRecord::sign(
+            key,
+            id,
+            version,
+            RecordBody::Claim {
+                address: Some("10.23.0.2".parse().unwrap()),
+                range: Some("10.23.0.0/24".parse::<Ipv4Range>().unwrap()),
+                hostname: Some("retired".into()),
+                hostname_since: Some(123),
+            },
+        )
+    }
+
+    #[test]
+    fn markers_release_claims_in_any_merge_order_and_older_snapshots_cannot_restore_them() {
+        let id = NetworkId::from_bytes([7; 32]);
+        let target = SecretKey::generate();
+        let cleaner = SecretKey::generate();
+        let record = claim(&target, id, 5);
+        let marker = SignedRetirement::sign(&cleaner, id, 8, record.clone());
+        for first in [true, false] {
+            let mut state = StateSet::new();
+            if first {
+                state.merge_retirement(id, marker.clone(), 16).unwrap();
+            }
+            state.merge(id, record.clone()).unwrap();
+            if !first {
+                state.merge_retirement(id, marker.clone(), 16).unwrap();
+            }
+            assert!(state.address_holders().is_empty());
+            assert!(state.hostname_holders().is_empty());
+            state.merge(id, claim(&target, id, 4)).unwrap();
+            assert!(state.is_retired(&target.public()));
+            assert!(
+                state.agreed_range().is_some(),
+                "cleanup does not renumber the network"
+            );
+            state.merge(id, claim(&target, id, 6)).unwrap();
+            assert!(!state.is_retired(&target.public()));
+            assert_eq!(state.hostname_of(&target.public()), Some("retired"));
+        }
+    }
+
+    #[test]
+    fn proof_signature_scope_tampering_future_cutoffs_and_capacity_are_checked() {
+        let id = NetworkId::from_bytes([7; 32]);
+        let target = SecretKey::generate();
+        let cleaner = SecretKey::generate();
+        let marker = SignedRetirement::sign(&cleaner, id, 8, claim(&target, id, 5));
+        assert!(marker.verify(NetworkId::from_bytes([8; 32])).is_err());
+        let mut forged = marker.clone();
+        forged.target.version = 99;
+        assert!(forged.verify(id).is_err());
+        let mut forged = marker.clone();
+        forged.signature[0] ^= 1;
+        assert!(forged.verify(id).is_err());
+        let mut proof = marker.target.clone();
+        proof.version = 99;
+        assert!(
+            SignedRetirement::sign(&cleaner, id, 8, proof)
+                .verify(id)
+                .is_err()
+        );
+        let mut state = StateSet::new();
+        assert!(state.merge_retirement(id, marker.clone(), 0).is_err());
+        assert!(state.merge_retirement(id, marker.clone(), 1).unwrap());
+        assert!(!state.merge_retirement(id, marker.clone(), 1).unwrap());
+        assert!(postcard::to_stdvec(&marker).unwrap().len() <= MAX_RETIREMENT_LEN);
+        let another = SecretKey::generate();
+        assert!(
+            state
+                .merge_retirement(
+                    id,
+                    SignedRetirement::sign(&cleaner, id, 9, claim(&another, id, 1)),
+                    1
+                )
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn competing_cleanup_markers_converge_and_keep_the_highest_observed_cutoff() {
+        let id = NetworkId::from_bytes([7; 32]);
+        let target = SecretKey::generate();
+        let a = SecretKey::generate();
+        let b = SecretKey::generate();
+        let markers = [
+            SignedRetirement::sign(&a, id, 1, claim(&target, id, 4)),
+            SignedRetirement::sign(&a, id, 2, claim(&target, id, 5)),
+            SignedRetirement::sign(&b, id, 3, claim(&target, id, 5)),
+        ];
+        let mut forward = StateSet::new();
+        let mut reverse = StateSet::new();
+        for marker in &markers {
+            forward.merge_retirement(id, marker.clone(), 16).unwrap();
+        }
+        for marker in markers.iter().rev() {
+            reverse.merge_retirement(id, marker.clone(), 16).unwrap();
+        }
+        assert_eq!(forward.retirements(), reverse.retirements());
+        assert_eq!(
+            forward
+                .retirement_of(&target.public())
+                .unwrap()
+                .target
+                .version,
+            5
+        );
+    }
+}

crates/tsunagi-core/src/state/vip.rs

@@ -0,0 +1,439 @@
+//! Signed multi-writer virtual service registers and deterministic merge rules.
+use crate::error::Result;
+use crate::vip::{Backend, Listener, Protocol, VipConfig, VipStatus, invalid};
+use crate::{Limits, NetworkId};
+use iroh::{EndpointId, SecretKey, Signature};
+use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
+use std::collections::{BTreeMap, BTreeSet};
+use std::net::Ipv4Addr;
+
+/// A signed, independently mergeable field. A removed field keeps its tombstone.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct VipEntry {
+    /// Network scope.
+    pub network: [u8; 32],
+    /// Stable service id.
+    pub vip: String,
+    /// Canonical field key.
+    pub key: String,
+    /// Encoded field value; absent means deletion.
+    pub value: Option<Vec<u8>>,
+    /// Lamport clock, independent of wall time.
+    pub clock: u64,
+    /// Persistent signing identity.
+    pub author: [u8; 32],
+    /// Author's durable network counter.
+    pub version: u64,
+    /// Ed25519 signature.
+    pub signature: Vec<u8>,
+}
+impl VipEntry {
+    fn bytes(&self) -> Vec<u8> {
+        let mut e = self.clone();
+        e.signature.clear();
+        let mut out = b"tsunagi-vip-field/v1\0".to_vec();
+        if let Ok(bytes) = postcard::to_stdvec(&e) {
+            out.extend(bytes);
+        }
+        out
+    }
+    /// Verifies bounds, network and authorship before the entry is merged.
+    pub fn verify(&self, network: NetworkId) -> Result<()> {
+        if self.network != *network.as_bytes()
+            || self.vip.len() != 32
+            || hex::decode(&self.vip).is_err()
+            || self.key.len() > 160
+            || self.value.as_ref().is_some_and(|v| v.len() > 256)
+            || self.version == 0
+            || self.version > i64::MAX as u64
+            || self.clock == 0
+            || self.clock > i64::MAX as u64
+        {
+            return Err(invalid("invalid VIP field bounds or network"));
+        }
+        let author =
+            EndpointId::from_bytes(&self.author).map_err(|_| invalid("invalid VIP author"))?;
+        let raw: [u8; 64] = self
+            .signature
+            .as_slice()
+            .try_into()
+            .map_err(|_| invalid("invalid VIP signature length"))?;
+        author
+            .verify(&self.bytes(), &Signature::from_bytes(&raw))
+            .map_err(|_| invalid("invalid VIP signature"))?;
+        if self.key != "deleted"
+            && self.key != "name"
+            && !self.key.starts_with("l/")
+            && !self.key.starts_with("b/")
+        {
+            return Err(invalid("unknown VIP field"));
+        }
+        match self.key.as_str() {
+            "deleted" => {
+                if self.value.as_deref() != Some(&[1][..]) {
+                    return Err(invalid("invalid VIP deletion tombstone"));
+                }
+            }
+            "name" => {
+                let value = self
+                    .value
+                    .as_ref()
+                    .ok_or_else(|| invalid("VIP name cannot be absent"))?;
+                let name: String =
+                    postcard::from_bytes(value).map_err(|_| invalid("invalid VIP name field"))?;
+                VipConfig {
+                    id: self.vip.clone(),
+                    name,
+                    listeners: vec![],
+                }
+                .validate(&Limits::default())?;
+            }
+            key if key.starts_with("l/") => {
+                let parts: Vec<_> = key.split('/').collect();
+                if parts.len() != 3
+                    || parts[1].parse::<Protocol>().is_err()
+                    || parts[2].parse::<u16>().ok().is_none_or(|p| p == 0)
+                {
+                    return Err(invalid("invalid listener field key"));
+                }
+                if let Some(value) = &self.value {
+                    let l: Listener = postcard::from_bytes(value)
+                        .map_err(|_| invalid("invalid listener field"))?;
+                    if !l.backends.is_empty() || key != format!("l/{}/{}", l.protocol, l.port) {
+                        return Err(invalid("listener key does not match its value"));
+                    }
+                    VipConfig {
+                        id: self.vip.clone(),
+                        name: "validation".into(),
+                        listeners: vec![l],
+                    }
+                    .validate(&Limits::default())?;
+                }
+            }
+            key if key.starts_with("b/") => {
+                let parts: Vec<_> = key.split('/').collect();
+                if parts.len() != 6
+                    || parts[1].parse::<Protocol>().is_err()
+                    || parts[2].parse::<u16>().ok().is_none_or(|p| p == 0)
+                    || parts[3].len() != 32
+                    || hex::decode(parts[3]).is_err()
+                    || parts[4].parse::<EndpointId>().is_err()
+                    || parts[5].parse::<u16>().ok().is_none_or(|p| p == 0)
+                {
+                    return Err(invalid("invalid backend field key"));
+                }
+                if let Some(value) = &self.value {
+                    let b: Backend = postcard::from_bytes(value)
+                        .map_err(|_| invalid("invalid backend field"))?;
+                    if b.peer != parts[4] || b.port.to_string() != parts[5] {
+                        return Err(invalid("backend key does not match its value"));
+                    }
+                }
+            }
+            _ => return Err(invalid("unknown VIP field")),
+        }
+        Ok(())
+    }
+    fn rank(&self) -> (u64, [u8; 32], Vec<u8>) {
+        (self.clock, self.author, self.bytes())
+    }
+}
+/// Signed registers retained across absences, including deletion tombstones.
+#[derive(Debug, Clone, Default)]
+pub struct VipSet {
+    entries: BTreeMap<(String, String), VipEntry>,
+}
+impl VipSet {
+    /// Highest retained version genuinely signed by this author.
+    pub fn authored_version(&self, author: EndpointId) -> u64 {
+        self.entries
+            .values()
+            .filter(|e| e.author == *author.as_bytes())
+            .map(|e| e.version)
+            .max()
+            .unwrap_or(0)
+    }
+    /// Complete independently verifiable state of one service, including tombstones.
+    pub fn entries(&self, id: &str) -> Vec<VipEntry> {
+        self.entries
+            .values()
+            .filter(|e| e.vip == id)
+            .cloned()
+            .collect()
+    }
+    /// All retained service ids, including deleted services.
+    pub fn ids(&self) -> Vec<String> {
+        self.entries
+            .keys()
+            .map(|(id, _)| id.clone())
+            .collect::<BTreeSet<_>>()
+            .into_iter()
+            .collect()
+    }
+    /// Opaque optimistic concurrency token for a whole service.
+    pub fn revision(&self, id: &str) -> String {
+        let mut hash = Sha256::new();
+        for e in self.entries(id) {
+            hash.update(e.bytes());
+            hash.update(&e.signature);
+        }
+        hex::encode(hash.finalize())
+    }
+    /// Atomically verifies and merges one complete service snapshot.
+    pub fn merge(
+        &mut self,
+        network: NetworkId,
+        entries: Vec<VipEntry>,
+        limits: &Limits,
+    ) -> Result<bool> {
+        if entries.len() > limits.max_vip_fields {
+            return Err(invalid("too many VIP fields"));
+        }
+        let mut next = self.clone();
+        for e in entries {
+            e.verify(network)?;
+            let key = (e.vip.clone(), e.key.clone());
+            if next
+                .entries
+                .get(&key)
+                .is_none_or(|old| e.rank() > old.rank())
+            {
+                next.entries.insert(key, e);
+            }
+        }
+        let deleted: BTreeSet<_> = next
+            .entries
+            .values()
+            .filter(|e| e.key == "deleted" && e.value.is_some())
+            .map(|e| e.vip.clone())
+            .collect();
+        // A permanent service tombstone alone prevents every former field's replay.
+        next.entries
+            .retain(|(id, key), _| !deleted.contains(id) || key == "deleted");
+        let ids = next.ids();
+        if ids.len().saturating_sub(deleted.len()) > limits.max_vips
+            || deleted.len() > limits.max_vip_tombstones
+            || ids
+                .iter()
+                .any(|id| next.entries(id).len() > limits.max_vip_fields)
+        {
+            return Err(invalid(
+                "VIP state capacity exceeded; deletion protection is retained",
+            ));
+        }
+        let changed = self.entries != next.entries;
+        *self = next;
+        Ok(changed)
+    }
+    /// Materialized configuration. A service tombstone permanently retires its id.
+    pub fn configs(&self) -> Vec<VipConfig> {
+        self.ids()
+            .into_iter()
+            .filter_map(|id| self.config(&id))
+            .collect()
+    }
+    fn config(&self, id: &str) -> Option<VipConfig> {
+        let fields = self.entries(id);
+        if fields
+            .iter()
+            .any(|e| e.key == "deleted" && e.value.is_some())
+        {
+            return None;
+        }
+        let name = fields
+            .iter()
+            .find(|e| e.key == "name")?
+            .value
+            .as_ref()
+            .and_then(|v| postcard::from_bytes::<String>(v).ok())?;
+        let mut listeners: Vec<Listener> = fields
+            .iter()
+            .filter(|e| e.key.starts_with("l/"))
+            .filter_map(|e| e.value.as_ref().and_then(|v| postcard::from_bytes(v).ok()))
+            .collect();
+        for l in &mut listeners {
+            l.backends = fields
+                .iter()
+                .filter(|e| {
+                    e.key
+                        .starts_with(&format!("b/{}/{}/{}/", l.protocol, l.port, l.id))
+                })
+                .filter_map(|e| e.value.as_ref().and_then(|v| postcard::from_bytes(v).ok()))
+                .collect();
+            l.backends.sort();
+        }
+        listeners.sort_by_key(|l| (l.protocol, l.port));
+        Some(VipConfig {
+            id: id.into(),
+            name,
+            listeners,
+        })
+    }
+    /// Builds a signed patch from the user's complete, version-checked draft.
+    pub fn edit(
+        &self,
+        network: NetworkId,
+        secret: &SecretKey,
+        version: u64,
+        config: &VipConfig,
+        expected: Option<&str>,
+        limits: &Limits,
+    ) -> Result<Vec<VipEntry>> {
+        config.validate(limits)?;
+        if let Some(revision) = expected {
+            if revision != self.revision(&config.id) {
+                return Err(invalid("VIP changed while editing; reload before saving"));
+            }
+        } else if !self.entries(&config.id).is_empty() {
+            return Err(invalid("VIP id already exists"));
+        }
+        if self
+            .entries
+            .get(&(config.id.clone(), "deleted".into()))
+            .is_some_and(|e| e.value.is_some())
+        {
+            return Err(invalid("VIP was deleted; create a new VIP"));
+        }
+        let before = self
+            .config(&config.id)
+            .map(|c| flatten(&c))
+            .transpose()?
+            .unwrap_or_default();
+        let after = flatten(config)?;
+        let keys: BTreeSet<_> = before.keys().chain(after.keys()).cloned().collect();
+        let changes = keys
+            .into_iter()
+            .filter_map(|key| {
+                let value = after.get(&key).cloned();
+                (before.get(&key) != value.as_ref()).then_some((key, value))
+            })
+            .collect();
+        self.sign(network, secret, version, &config.id, changes)
+    }
+    /// Builds a permanent service tombstone after checking the selected revision.
+    pub fn delete(
+        &self,
+        network: NetworkId,
+        secret: &SecretKey,
+        version: u64,
+        id: &str,
+        expected: &str,
+    ) -> Result<Vec<VipEntry>> {
+        if self.config(id).is_none() || self.revision(id) != expected {
+            return Err(invalid("VIP missing or changed while editing"));
+        }
+        self.sign(
+            network,
+            secret,
+            version,
+            id,
+            vec![("deleted".into(), Some(vec![1]))],
+        )
+    }
+    fn sign(
+        &self,
+        network: NetworkId,
+        secret: &SecretKey,
+        version: u64,
+        id: &str,
+        changes: Vec<(String, Option<Vec<u8>>)>,
+    ) -> Result<Vec<VipEntry>> {
+        let clock = self
+            .entries
+            .values()
+            .map(|e| e.clock)
+            .max()
+            .unwrap_or(0)
+            .checked_add(1)
+            .filter(|n| *n <= i64::MAX as u64)
+            .ok_or_else(|| invalid("VIP clock exhausted"))?;
+        Ok(changes
+            .into_iter()
+            .map(|(key, value)| {
+                let mut e = VipEntry {
+                    network: *network.as_bytes(),
+                    vip: id.into(),
+                    key,
+                    value,
+                    clock,
+                    author: *secret.public().as_bytes(),
+                    version,
+                    signature: vec![],
+                };
+                e.signature = secret.sign(&e.bytes()).to_bytes().to_vec();
+                e
+            })
+            .collect())
+    }
+}
+pub(crate) fn flatten(cfg: &VipConfig) -> Result<BTreeMap<String, Vec<u8>>> {
+    let mut fields = BTreeMap::new();
+    fields.insert(
+        "name".into(),
+        postcard::to_stdvec(&cfg.name).map_err(|e| invalid(e.to_string()))?,
+    );
+    for l in &cfg.listeners {
+        let mut settings = l.clone();
+        settings.backends.clear();
+        fields.insert(
+            format!("l/{}/{}", l.protocol, l.port),
+            postcard::to_stdvec(&settings).map_err(|e| invalid(e.to_string()))?,
+        );
+        for b in &l.backends {
+            fields.insert(
+                format!("b/{}/{}/{}/{}/{}", l.protocol, l.port, l.id, b.peer, b.port),
+                postcard::to_stdvec(b).map_err(|e| invalid(e.to_string()))?,
+            );
+        }
+    }
+    Ok(fields)
+}
+impl VipSet {
+    /// Allocates internal addresses deterministically against durable node claims.
+    pub(crate) fn views(
+        &self,
+        range: Option<crate::state::Ipv4Range>,
+        mut taken: std::collections::HashSet<Ipv4Addr>,
+        hostnames: &BTreeSet<String>,
+        limits: &Limits,
+    ) -> Vec<VipStatus> {
+        let mut names = hostnames.clone();
+        let mut result = Vec::new();
+        for config in self.configs() {
+            let mut error = config.validate(limits).err().map(|e| e.to_string());
+            if !names.insert(config.name.clone()) {
+                error = Some("name conflicts with another VIP or node".into());
+            }
+            let address = range.and_then(|r| {
+                let usable = crate::state::allocator::usable_addresses(r);
+                if usable == 0 {
+                    return None;
+                }
+                let digest = Sha256::digest(config.id.as_bytes());
+                let seed = u64::from_be_bytes([
+                    digest[0], digest[1], digest[2], digest[3], digest[4], digest[5], digest[6],
+                    digest[7],
+                ]);
+                let bits = 32 - u32::from(r.prefix_len);
+                let mask = if bits == 32 { 0 } else { u32::MAX << bits };
+                (0..usable).find_map(|step| {
+                    let offset = (seed.wrapping_add(step) % usable + 1) as u32;
+                    let ip = Ipv4Addr::from((u32::from(r.base) & mask) | offset);
+                    taken.insert(ip).then_some(ip)
+                })
+            });
+            if address.is_none() {
+                error = Some("no free overlay address".into());
+            }
+            result.push(VipStatus {
+                revision: self.revision(&config.id),
+                config,
+                address,
+                error,
+                listeners: vec![],
+            });
+        }
+        result
+    }
+}

crates/tsunagi-core/src/storage/mod.rs

@@ -0,0 +1,603 @@
+//! Persistence, split into mandatory state and a disposable cache.
+//!
+//! | store          | contents                                               | on damage |
+//! |----------------|--------------------------------------------------------|-----------|
+//! | `state.sqlite` | device identity, network configuration, hostname        | hard error |
+//! | `cache.sqlite` | address hints and other recoverable data                | discarded and recreated |
+//!
+//! Both files are created with owner-only permissions where the platform
+//! supports it. The state directory additionally carries an ownership lock, see
+//! [`DirectoryLock`].
+//!
+//! SQLite is synchronous. Every call that touches a database therefore runs on
+//! a blocking pool via [`tokio::task::spawn_blocking`], and no database lock is
+//! ever held across a network `await`.
+
+mod cache;
+mod lock;
+mod state;
+
+pub use cache::{AddressHint, CacheOutcome, CacheStore};
+pub use lock::DirectoryLock;
+pub use state::{SCHEMA_VERSION, StateStore, StoredNetwork};
+
+use std::path::Path;
+use std::sync::{Arc, Mutex, MutexGuard};
+
+use crate::config::StoragePaths;
+use crate::error::{Error, Result};
+use crate::identity::{DeviceIdentity, NetworkId, NetworkName, NetworkSecret};
+use crate::state::SignedRecord;
+
+/// Applies the pragmas both stores share.
+fn apply_common_pragmas(conn: &rusqlite::Connection) -> rusqlite::Result<()> {
+    conn.busy_timeout(std::time::Duration::from_secs(5))?;
+    conn.pragma_update(None, "journal_mode", "WAL")?;
+    conn.pragma_update(None, "synchronous", "NORMAL")?;
+    conn.pragma_update(None, "foreign_keys", "ON")?;
+    Ok(())
+}
+
+/// Restricts a file to the current user where the platform supports it.
+#[cfg(unix)]
+fn restrict_permissions(file: &std::fs::File, path: &Path) -> Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    let perms = std::fs::Permissions::from_mode(0o600);
+    file.set_permissions(perms).map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })
+}
+
+/// On Windows, files inherit the parent directory's ACL, which for the
+/// per-user application data directory is already restricted to that user.
+#[cfg(not(unix))]
+fn restrict_permissions(_file: &std::fs::File, _path: &Path) -> Result<()> {
+    Ok(())
+}
+
+/// Restricts a file to the user that owns it.
+///
+/// Part of what the system level offers a protocol: a plugin keeping keys of
+/// its own on disk has the same obligation as the agent, and should not have
+/// to work out the platform details again to meet it.
+#[cfg(unix)]
+pub fn restrict_path_permissions(path: &Path) -> Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o600)).map_err(|source| {
+        Error::Io {
+            path: path.to_path_buf(),
+            source,
+        }
+    })
+}
+
+/// Restricts a file to the user that owns it. A no-op off Unix.
+#[cfg(not(unix))]
+pub fn restrict_path_permissions(_path: &Path) -> Result<()> {
+    Ok(())
+}
+
+/// Restricts a directory to the user that owns it.
+#[cfg(unix)]
+pub fn restrict_dir_permissions(path: &Path) -> Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o700)).map_err(|source| {
+        Error::Io {
+            path: path.to_path_buf(),
+            source,
+        }
+    })
+}
+
+/// Restricts a directory to the user that owns it. A no-op off Unix.
+#[cfg(not(unix))]
+pub fn restrict_dir_permissions(_path: &Path) -> Result<()> {
+    Ok(())
+}
+
+/// Creates a directory and every parent, restricted to this user.
+pub fn create_dir(path: &Path) -> Result<()> {
+    std::fs::create_dir_all(path).map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })?;
+    restrict_dir_permissions(path)
+}
+
+/// Async facade over both stores, holding the directory ownership lock.
+///
+/// Cloning shares the same underlying connections and the same lock.
+#[derive(Debug, Clone)]
+pub struct Storage {
+    inner: Arc<Inner>,
+}
+
+#[derive(Debug)]
+struct Inner {
+    state: Mutex<StateStore>,
+    cache: Mutex<Option<CacheStore>>,
+    cache_outcome: CacheOutcome,
+    paths: StoragePaths,
+    lock: Mutex<Option<DirectoryLock>>,
+}
+
+impl Storage {
+    /// Opens both stores and takes the ownership lock on the state directory.
+    ///
+    /// Fails with [`Error::StateLocked`] if another live agent owns the state
+    /// directory, and with [`Error::StateCorrupted`] if the mandatory state is
+    /// unusable. A broken cache is silently discarded and reported through
+    /// [`Storage::cache_outcome`].
+    pub fn open(paths: &StoragePaths) -> Result<Self> {
+        create_dir(&paths.state_dir)?;
+        create_dir(&paths.cache_dir)?;
+
+        let lock = DirectoryLock::acquire(paths.lock_file())?;
+        let state = StateStore::open(paths.state_db())?;
+        let (cache, cache_outcome) = CacheStore::open_or_reset(paths.cache_db())?;
+
+        Ok(Self {
+            inner: Arc::new(Inner {
+                state: Mutex::new(state),
+                cache: Mutex::new(Some(cache)),
+                cache_outcome,
+                paths: paths.clone(),
+                lock: Mutex::new(Some(lock)),
+            }),
+        })
+    }
+
+    /// What happened to the cache when the agent started.
+    pub fn cache_outcome(&self) -> &CacheOutcome {
+        &self.inner.cache_outcome
+    }
+
+    /// The configured paths.
+    pub fn paths(&self) -> &StoragePaths {
+        &self.inner.paths
+    }
+
+    fn lock_state(&self) -> MutexGuard<'_, StateStore> {
+        match self.inner.state.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    fn lock_cache(&self) -> MutexGuard<'_, Option<CacheStore>> {
+        match self.inner.cache.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    /// Runs a closure against the mandatory state store on the blocking pool.
+    async fn with_state<T, F>(&self, f: F) -> Result<T>
+    where
+        T: Send + 'static,
+        F: FnOnce(&StateStore) -> Result<T> + Send + 'static,
+    {
+        let inner = Arc::clone(&self.inner);
+        tokio::task::spawn_blocking(move || {
+            let guard = match inner.state.lock() {
+                Ok(guard) => guard,
+                Err(poisoned) => poisoned.into_inner(),
+            };
+            f(&guard)
+        })
+        .await
+        .map_err(|err| Error::Storage(format!("state task failed: {err}")))?
+    }
+
+    /// Runs a closure against the cache, tolerating an unavailable cache.
+    ///
+    /// If the cache has been disabled because it misbehaved, the closure is
+    /// skipped and `default` is returned.
+    async fn with_cache<T, F>(&self, default: T, f: F) -> T
+    where
+        T: Send + 'static,
+        F: FnOnce(&CacheStore) -> Result<T> + Send + 'static,
+    {
+        let inner = Arc::clone(&self.inner);
+        let joined = tokio::task::spawn_blocking(move || {
+            let guard = match inner.cache.lock() {
+                Ok(guard) => guard,
+                Err(poisoned) => poisoned.into_inner(),
+            };
+            match guard.as_ref() {
+                Some(cache) => f(cache),
+                None => Err(Error::Storage("cache is unavailable".into())),
+            }
+        })
+        .await;
+
+        match joined {
+            Ok(Ok(value)) => value,
+            Ok(Err(err)) => {
+                tracing::debug!(%err, "cache operation failed; continuing without it");
+                default
+            }
+            Err(err) => {
+                tracing::debug!(%err, "cache task failed; continuing without it");
+                default
+            }
+        }
+    }
+
+    /// Loads or creates the persistent device identity.
+    pub async fn device_identity(&self) -> Result<DeviceIdentity> {
+        self.with_state(|state| state.load_or_create_device_identity())
+            .await
+    }
+
+    /// Lists configured networks.
+    pub async fn list_networks(&self) -> Result<Vec<StoredNetwork>> {
+        self.with_state(|state| state.list_networks()).await
+    }
+
+    /// Stores or updates a network configuration.
+    pub async fn upsert_network(
+        &self,
+        network_id: NetworkId,
+        name: NetworkName,
+        secret: NetworkSecret,
+        auto_start: bool,
+    ) -> Result<()> {
+        self.with_state(move |state| state.upsert_network(network_id, &name, &secret, auto_start))
+            .await
+    }
+
+    /// Persists local broadcast participation for one configured network.
+    pub async fn set_broadcast(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.with_state(move |state| state.set_broadcast(network_id, enabled))
+            .await
+    }
+
+    /// Persists the ordered protocol selection off the async executor.
+    pub async fn set_protocols(&self, network_id: NetworkId, protocols: Vec<String>) -> Result<()> {
+        self.with_state(move |state| state.set_protocols(network_id, &protocols))
+            .await
+    }
+
+    /// Persists whether this agent offers itself as an exit node in a network.
+    pub async fn set_exit_node(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.with_state(move |state| state.set_exit_node(network_id, enabled))
+            .await
+    }
+
+    /// Persists the member this device sends its internet traffic through.
+    pub async fn set_exit_via(&self, network_id: NetworkId, via: Option<[u8; 32]>) -> Result<()> {
+        self.with_state(move |state| state.set_exit_via(network_id, via.as_ref()))
+            .await
+    }
+
+    /// Updates the auto-start flag of a network.
+    pub async fn set_auto_start(&self, network_id: NetworkId, auto_start: bool) -> Result<()> {
+        self.with_state(move |state| state.set_auto_start(network_id, auto_start))
+            .await
+    }
+
+    /// Configured networks in the user's most recently selected order.
+    pub async fn network_start_order(&self) -> Result<Vec<NetworkId>> {
+        self.with_state(|state| {
+            let configured = state.list_networks()?;
+            let text = state
+                .get_setting("network_start_order")?
+                .unwrap_or_default();
+            let mut order = Vec::new();
+            for word in text.split(',') {
+                if let Ok(id) = word.parse::<NetworkId>()
+                    && configured.iter().any(|n| n.network_id == id)
+                    && !order.contains(&id)
+                {
+                    order.push(id);
+                }
+            }
+            Ok(order)
+        })
+        .await
+    }
+
+    /// Prefers this network at startup without disabling suspended networks.
+    pub async fn prefer_network(&self, network: NetworkId) -> Result<()> {
+        self.with_state(move |state| state.prefer_network(network))
+            .await
+    }
+
+    /// Removes a network configuration and its cached hints.
+    pub async fn remove_network(&self, network_id: NetworkId) -> Result<()> {
+        self.with_state(move |state| {
+            state.remove_network(network_id)?;
+            state.forget_signed_records(network_id)?;
+            state.forget_retirements(network_id)?;
+            state.forget_peer_hostnames(network_id)
+        })
+        .await?;
+        self.with_cache((), move |cache| cache.forget_network(network_id))
+            .await;
+        Ok(())
+    }
+
+    /// Reads the stored hostname.
+    pub async fn hostname(&self) -> Result<Option<String>> {
+        self.with_state(|state| state.hostname()).await
+    }
+
+    /// Writes the stored hostname.
+    pub async fn set_hostname(&self, hostname: String) -> Result<()> {
+        self.with_state(move |state| state.set_hostname(&hostname))
+            .await
+    }
+
+    /// Remembers the hostname a member announced, so it can still be named
+    /// once it is away.
+    pub async fn remember_peer_hostname(
+        &self,
+        network_id: NetworkId,
+        endpoint_id: [u8; 32],
+        hostname: String,
+    ) -> Result<()> {
+        self.with_state(move |state| {
+            state.remember_peer_hostname(network_id, &endpoint_id, &hostname)
+        })
+        .await
+    }
+
+    /// The hostnames members of a network last announced.
+    pub async fn peer_hostnames(&self, network_id: NetworkId) -> Result<Vec<([u8; 32], String)>> {
+        self.with_state(move |state| state.peer_hostnames(network_id))
+            .await
+    }
+
+    /// Records an address hint. Failures are non-fatal.
+    pub async fn record_hint(
+        &self,
+        network_id: NetworkId,
+        endpoint_id: [u8; 32],
+        addr: String,
+        max_per_peer: usize,
+    ) {
+        self.with_cache((), move |cache| {
+            cache.record_hint(network_id, &endpoint_id, &addr, max_per_peer)
+        })
+        .await;
+    }
+
+    /// Loads every signed record known for a network.
+    pub async fn signed_records(&self, network_id: NetworkId) -> Result<Vec<SignedRecord>> {
+        self.with_state(move |state| state.signed_records(network_id))
+            .await
+    }
+
+    /// Loads mandatory signed VIP fields with a strict capacity bound.
+    pub async fn vip_fields(
+        &self,
+        network: NetworkId,
+        max: usize,
+    ) -> Result<Vec<crate::vip::VipEntry>> {
+        self.with_state(move |state| state.vip_fields(network, max))
+            .await
+    }
+    /// Stores a complete merged service snapshot and optional own counter.
+    pub async fn put_vip_fields(
+        &self,
+        network: NetworkId,
+        entries: Vec<crate::vip::VipEntry>,
+        own: Option<([u8; 32], u64)>,
+    ) -> Result<()> {
+        self.with_state(move |state| state.put_vip_fields(network, &entries, own))
+            .await
+    }
+    /// Loads verified durable cleanup proofs off the executor.
+    pub async fn retirements(
+        &self,
+        network: NetworkId,
+        max: usize,
+    ) -> Result<Vec<crate::state::SignedRetirement>> {
+        self.with_state(move |state| state.retirements(network, max))
+            .await
+    }
+
+    /// Persists a received cleanup proof before applying it.
+    pub async fn put_retirement(&self, marker: crate::state::SignedRetirement) -> Result<()> {
+        self.with_state(move |state| state.put_retirement(&marker))
+            .await
+    }
+
+    /// Commits our cleanup and author counter together before announcement.
+    pub async fn publish_own_retirement(
+        &self,
+        marker: crate::state::SignedRetirement,
+    ) -> Result<()> {
+        self.with_state(move |state| state.publish_own_retirement(&marker))
+            .await
+    }
+
+    /// Stores a record received from another replica.
+    pub async fn put_signed_record(&self, record: SignedRecord) -> Result<()> {
+        self.with_state(move |state| state.put_signed_record(&record))
+            .await
+    }
+
+    /// Stores one of this agent's own records and bumps its counter in one
+    /// transaction, which must happen before the record is announced.
+    pub async fn publish_own_record(&self, record: SignedRecord) -> Result<()> {
+        self.with_state(move |state| state.publish_own_record(&record))
+            .await
+    }
+
+    /// The highest version this agent has ever published for a network.
+    pub async fn own_record_version(
+        &self,
+        network_id: NetworkId,
+        author: iroh::EndpointId,
+    ) -> Result<u64> {
+        self.with_state(move |state| state.own_record_version(network_id, author))
+            .await
+    }
+
+    /// Reads cached address hints. Returns an empty list if the cache is gone.
+    pub async fn hints_for_network(&self, network_id: NetworkId) -> Vec<AddressHint> {
+        self.with_cache(Vec::new(), move |cache| cache.hints_for_network(network_id))
+            .await
+    }
+
+    /// Synchronously reads the hostname. Used only during startup.
+    pub(crate) fn hostname_blocking(&self) -> Result<Option<String>> {
+        self.lock_state().hostname()
+    }
+
+    /// Whether the cache is currently usable.
+    pub fn cache_healthy(&self) -> bool {
+        self.lock_cache().is_some()
+    }
+
+    /// Releases the state directory ownership lock.
+    ///
+    /// Called by [`crate::Agent::shutdown`] so that a cleanly stopped agent
+    /// leaves its directory immediately claimable by another instance. The
+    /// databases stay open and readable, but this handle no longer owns the
+    /// directory and must not be used to write after this point.
+    pub fn release_ownership_lock(&self) {
+        let mut guard = match self.inner.lock.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        guard.take();
+    }
+}
+
+/// What a wipe covers: every entry in the two directories.
+///
+/// Reported before anything is removed, and again afterwards, because a
+/// command whose whole job is to destroy state has to be able to say exactly
+/// what it is about to destroy.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct WipePlan {
+    /// Entries in the state directory, which is the mandatory state and
+    /// anything a plugin keeps beside it.
+    pub state: Vec<std::path::PathBuf>,
+    /// Entries in the cache directory.
+    pub cache: Vec<std::path::PathBuf>,
+}
+
+impl WipePlan {
+    /// Whether there is anything to remove at all.
+    pub fn is_empty(&self) -> bool {
+        self.state.is_empty() && self.cache.is_empty()
+    }
+
+    /// Every entry, state first.
+    pub fn entries(&self) -> impl Iterator<Item = &std::path::Path> {
+        self.state
+            .iter()
+            .chain(self.cache.iter())
+            .map(std::path::PathBuf::as_path)
+    }
+}
+
+/// Lists what [`wipe`] would remove, without removing anything.
+pub fn wipe_plan(paths: &StoragePaths) -> Result<WipePlan> {
+    Ok(WipePlan {
+        state: removable(&paths.state_dir, &paths.state_db())?,
+        cache: removable(&paths.cache_dir, &paths.cache_db())?,
+    })
+}
+
+/// Removes everything in both directories, leaving this device a stranger.
+///
+/// The device identity, every network it belongs to, every signed record and
+/// everything a plugin kept beside them. The next start generates a new
+/// identity and knows nothing, which is the point: after this there is no
+/// membership left to be surprised by.
+///
+/// What it is **not** is a goodbye. Other members keep the claims this device
+/// signed, because signed state has no expiry and there is nobody left here
+/// to sign a release. Leaving each network first with
+/// [`crate::Agent::leave_network`] is what frees the address and the name.
+///
+/// Refuses while an agent owns the state directory, and refuses a directory
+/// with no `state.sqlite` in it — a wipe pointed at the wrong place by a
+/// mistyped `--state-dir` would otherwise remove somebody's documents.
+pub fn wipe(paths: &StoragePaths) -> Result<WipePlan> {
+    let plan = wipe_plan(paths)?;
+    // Held across the removal, so an agent cannot start into a directory
+    // that is half gone. The lock file is itself removed at the end: on
+    // Unix the lock lives on the open handle, not on the name.
+    let lock = DirectoryLock::acquire(paths.lock_file())?;
+
+    for entry in plan.entries() {
+        remove_entry(entry)?;
+    }
+    drop(lock);
+    // Created by acquiring the lock a moment ago, so it is ours to remove
+    // and nothing is left behind claiming the directory.
+    let _ = std::fs::remove_file(paths.lock_file());
+    Ok(plan)
+}
+
+/// Every entry of `dir`, once `marker` proves it is one of ours.
+///
+/// An empty or missing directory is nothing to remove rather than an error:
+/// wiping twice must be as ordinary as wiping once.
+fn removable(dir: &Path, marker: &Path) -> Result<Vec<std::path::PathBuf>> {
+    if !dir.exists() {
+        return Ok(Vec::new());
+    }
+    let mut entries = Vec::new();
+    let read = std::fs::read_dir(dir).map_err(|source| Error::Io {
+        path: dir.to_path_buf(),
+        source,
+    })?;
+    for entry in read {
+        let entry = entry.map_err(|source| Error::Io {
+            path: dir.to_path_buf(),
+            source,
+        })?;
+        entries.push(entry.path());
+    }
+    if entries.is_empty() {
+        return Ok(entries);
+    }
+
+    // The marker is what tells a state directory from a directory that
+    // merely got named as one. The lock file counts too: an agent that
+    // crashed before writing anything still leaves that.
+    let ours = marker.exists() || dir.join("state.lock").exists();
+    if !ours {
+        return Err(Error::Storage(format!(
+            "{} does not look like a tsunagi directory: no {} in it. Nothing was removed.",
+            dir.display(),
+            marker.file_name().map_or_else(
+                || marker.display().to_string(),
+                |name| name.to_string_lossy().into_owned()
+            )
+        )));
+    }
+    entries.sort();
+    Ok(entries)
+}
+
+fn remove_entry(path: &Path) -> Result<()> {
+    let metadata = match std::fs::symlink_metadata(path) {
+        Ok(metadata) => metadata,
+        // Gone between the plan and the removal is the outcome asked for.
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(()),
+        Err(source) => {
+            return Err(Error::Io {
+                path: path.to_path_buf(),
+                source,
+            });
+        }
+    };
+    let removed = if metadata.is_dir() {
+        std::fs::remove_dir_all(path)
+    } else {
+        std::fs::remove_file(path)
+    };
+    removed.map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })
+}

crates/tsunagi-core/src/task.rs

@@ -0,0 +1,74 @@
+//! Stopping the tasks an agent started.
+//!
+//! Shutdown is cooperative first and bounded always: every loop watches a
+//! cancellation token, and anything still running when its grace period ends
+//! is aborted. Nothing an agent owns may outlive the agent, and no peer may
+//! hold shutdown open by refusing to read.
+//!
+//! Public because a protocol crate spawns tasks of its own and is held to
+//! the same rule: the agent gives each plugin a grace period, but abandoning
+//! the future it is awaiting does not stop the task behind it.
+
+use std::time::Duration;
+
+use tokio::task::JoinHandle;
+
+/// How long a task gets to notice cancellation before it is cut off.
+///
+/// Winding down is cooperative first: a task in the middle of a write to a
+/// peer gets a moment to finish it. After that it is aborted, because a peer
+/// that stops reading must not be able to hold shutdown open.
+pub const TASK_GRACE: Duration = Duration::from_secs(5);
+
+/// Waits for `task` to finish, aborting it if it outlasts `grace`.
+///
+/// Returns whether it finished on its own. `what` names it in the warning,
+/// which is the only signal that something was cut off rather than stopped.
+pub async fn wind_down(task: JoinHandle<()>, grace: Duration, what: &str) -> bool {
+    let abort = task.abort_handle();
+    if tokio::time::timeout(grace, task).await.is_err() {
+        tracing::warn!(
+            task = what,
+            grace_secs = grace.as_secs_f32(),
+            "task did not wind down in time; aborting it"
+        );
+        abort.abort();
+        return false;
+    }
+    true
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn a_task_that_stops_on_its_own_is_not_aborted() {
+        let (tx, mut rx) = tokio::sync::mpsc::channel::<()>(1);
+        let task = tokio::spawn(async move {
+            tx.send(()).await.ok();
+        });
+        assert!(wind_down(task, Duration::from_secs(5), "cooperative").await);
+        assert_eq!(rx.recv().await, Some(()), "it ran to completion");
+    }
+
+    #[tokio::test]
+    async fn a_task_that_ignores_the_grace_is_aborted() {
+        // A stand-in for a task blocked writing to a peer that stopped
+        // reading: it will never finish by itself.
+        let (tx, mut rx) = tokio::sync::mpsc::channel::<()>(1);
+        let task = tokio::spawn(async move {
+            let _held = tx;
+            std::future::pending::<()>().await;
+        });
+
+        let finished = wind_down(task, Duration::from_millis(50), "wedged").await;
+        assert!(!finished, "it had to be cut off");
+
+        // Aborted for real: the task is gone, so what it held is dropped.
+        let closed = tokio::time::timeout(Duration::from_secs(5), rx.recv()).await;
+        assert_eq!(closed, Ok(None), "the aborted task dropped its sender");
+    }
+}

crates/tsunagi-core/src/testing.rs

@@ -1,26 +1,31 @@
-//! Shared helpers for the integration tests.
+//! A harness for testing against a real agent.
 //!
-//! Every test uses real iroh endpoints on loopback, its own temporary SQLite
-//! files and independent agent instances. Only discovery is substituted; iroh,
-//! the handshake, message passing and persistent storage are not.
+//! Behind the `testing` feature, because it is only of use to a test and
+//! pulls in `tempfile` and a tracing subscriber. It lives here rather than
+//! being copied into each crate's `tests/` directory: a protocol plugin in
+//! its own crate needs exactly the same harness, and two copies of it would
+//! drift.
 //!
-//! Synchronisation is always "wait for a specific event or condition, under one
-//! overall deadline", never a fixed multi-second sleep.
+//! Everything here builds agents that touch nothing outside a temporary
+//! directory: loopback only, no relay, no address publication.
 
-#![allow(dead_code, clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+// A harness fails a test loudly and at once; that is the whole job. The
+// crate forbids these elsewhere for the opposite reason — nothing off the
+// network may panic — and this module never sees anything off the network.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
 use std::future::Future;
 use std::path::Path;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
+use crate::agent::Event;
+use crate::config::{AgentConfig, StoragePaths, TransportPolicy};
+use crate::discovery::SharedMemoryDiscovery;
+use crate::identity::{NetworkName, NetworkSecret};
+use crate::{Agent, Result};
 use tempfile::TempDir;
 use tokio::sync::broadcast::error::RecvError;
-use tsunagi::agent::Event;
-use tsunagi::config::{AgentConfig, StoragePaths, TransportPolicy};
-use tsunagi::discovery::SharedMemoryDiscovery;
-use tsunagi::identity::{NetworkName, NetworkSecret};
-use tsunagi::{Agent, Result};
 
 /// Installs a tracing subscriber when `TSUNAGI_TEST_LOG` is set.
 ///
@@ -50,7 +55,7 @@ const POLL_INTERVAL: Duration = Duration::from_millis(25);
 /// port mapping, so the suite needs neither the internet nor privileges.
 /// Timeouts are shortened so that failure paths finish quickly.
 pub fn local_config(dir: &Path) -> AgentConfig {
-    let limits = tsunagi::Limits {
+    let limits = crate::Limits {
         dial_timeout: Duration::from_millis(1500),
         handshake_timeout: Duration::from_secs(5),
         ..Default::default()
@@ -69,7 +74,9 @@ pub fn config_with(dir: &Path, discovery: &SharedMemoryDiscovery) -> AgentConfig
 
 /// A temporary directory plus the agent running on it.
 pub struct TestAgent {
+    /// The directory it keeps its state in, removed when this is dropped.
     pub dir: TempDir,
+    /// The agent itself.
     pub agent: Agent,
 }
 
@@ -148,14 +155,22 @@ where
 {
     let deadline = Instant::now() + DEADLINE;
     loop {
-        if let Some(value) = probe().await {
-            return value;
-        }
+        let remaining = deadline
+            .checked_duration_since(Instant::now())
+            .unwrap_or_default();
         assert!(
-            Instant::now() < deadline,
+            !remaining.is_zero(),
             "timed out waiting for condition: {what}"
         );
-        tokio::time::sleep(POLL_INTERVAL).await;
+
+        // The probe is bounded too. A deadline checked only between probes is
+        // no deadline at all: one call into a wedged agent that never answers
+        // would hang the test process for ever instead of failing it.
+        match tokio::time::timeout(remaining, probe()).await {
+            Ok(Some(value)) => return value,
+            Ok(None) => tokio::time::sleep(POLL_INTERVAL).await,
+            Err(_) => panic!("timed out waiting for condition: {what}"),
+        }
     }
 }
 
@@ -170,7 +185,7 @@ pub async fn settle() {
 /// Waits until `agent` has `count` authenticated peers in `network`.
 pub async fn wait_for_peers(
     agent: &Agent,
-    network: tsunagi::NetworkId,
+    network: crate::NetworkId,
     count: usize,
 ) -> Vec<iroh::EndpointId> {
     wait_until(&format!("{count} peers in {network}"), || async move {
@@ -183,3 +198,20 @@ pub async fn wait_for_peers(
     })
     .await
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test(start_paused = true)]
+    #[should_panic(expected = "timed out waiting for condition")]
+    async fn a_probe_that_never_answers_fails_the_deadline_instead_of_hanging() {
+        // Virtual time, so the deadline arrives at once rather than in thirty
+        // real seconds — and the deadline does arrive, which is the point: a
+        // probe that never returns used to hang the process instead.
+        wait_until::<(), _, _>("an answer that never comes", || async {
+            std::future::pending::<Option<()>>().await
+        })
+        .await;
+    }
+}