diff --git a/src/agent/cover_variants.rs b/src/agent/cover_variants.rs
index f4a78cb..81321ba 100644
--- a/src/agent/cover_variants.rs
+++ b/src/agent/cover_variants.rs
@@ -96,17 +96,11 @@ fn generate_missing_variants_sync(
             image::ExtendedColorType::Rgb8,
         );
         match result {
-            Ok(()) => crate::metrics::record_agent_cover_variant(
-                variant.name,
-                "ok",
-                start.elapsed(),
-            ),
+            Ok(()) => {
+                crate::metrics::record_agent_cover_variant(variant.name, "ok", start.elapsed())
+            }
             Err(err) => {
-                crate::metrics::record_agent_cover_variant(
-                    variant.name,
-                    "error",
-                    start.elapsed(),
-                );
+                crate::metrics::record_agent_cover_variant(variant.name, "error", start.elapsed());
                 return Err(err.into());
             }
         }
diff --git a/src/jobs/archive_cleanup.rs b/src/jobs/archive_cleanup.rs
new file mode 100644
index 0000000..04a3694
--- /dev/null
+++ b/src/jobs/archive_cleanup.rs
@@ -0,0 +1,431 @@
+use std::collections::BTreeSet;
+use std::io::ErrorKind;
+
+use sqlx::PgPool;
+
+use crate::scheduler::{Job, JobContext, JobLog};
+
+const SAMPLE_LOG_LIMIT: usize = 50;
+
+pub struct ArchiveCleanupJob;
+
+#[derive(Debug, sqlx::FromRow)]
+struct TrackFileRow {
+    track_id: i64,
+    track_title: String,
+    release_id: i64,
+    release_title: Option<String>,
+    media_file_id: Option<i64>,
+    file_type: Option<String>,
+    file_path: Option<String>,
+}
+
+#[derive(Debug)]
+struct MissingTrack {
+    track_id: i64,
+    track_title: String,
+    release_id: i64,
+    release_title: Option<String>,
+    media_file_id: Option<i64>,
+    file_path: Option<String>,
+    reason: MissingReason,
+}
+
+#[derive(Debug)]
+enum MissingReason {
+    MissingMediaRow,
+    InvalidMediaType(String),
+    EmptyPath,
+    MissingFile,
+    NotRegularFile,
+}
+
+#[derive(Debug, Default)]
+struct DeleteStats {
+    playback_states_cleared: u64,
+    playlist_entries_deleted: u64,
+    likes_deleted: u64,
+    play_history_deleted: u64,
+    popularity_history_deleted: u64,
+    scrobble_outbox_deleted: u64,
+    track_genres_deleted: u64,
+    entity_tags_deleted: u64,
+    external_ids_deleted: u64,
+    track_artists_deleted: u64,
+    tracks_deleted: u64,
+    media_files_deleted: u64,
+}
+
+#[async_trait::async_trait]
+impl Job for ArchiveCleanupJob {
+    fn name(&self) -> &'static str {
+        "archive_cleanup"
+    }
+
+    fn description(&self) -> &'static str {
+        "Clean stale archive records, starting with tracks whose audio files are missing"
+    }
+
+    fn default_cron(&self) -> &'static str {
+        // Daily at 04:45.
+        "0 45 4 * * *"
+    }
+
+    async fn run(&self, ctx: &JobContext, log: &mut JobLog) -> anyhow::Result<()> {
+        run_missing_audio_cleanup(ctx, log).await
+    }
+}
+
+async fn run_missing_audio_cleanup(ctx: &JobContext, log: &mut JobLog) -> anyhow::Result<()> {
+    let storage_dir = ctx.config.agent_storage_dir.trim();
+    if storage_dir.is_empty() {
+        log.warn("Archive cleanup: agent_storage_dir is not configured, skipping file checks");
+        return Ok(());
+    }
+
+    let rows = sqlx::query_as::<_, TrackFileRow>(
+        r#"SELECT t.id AS track_id,
+                  t.title::text AS track_title,
+                  t.release_id,
+                  r.title::text AS release_title,
+                  mf.id AS media_file_id,
+                  mf.file_type::text AS file_type,
+                  mf.file_path::text AS file_path
+             FROM furumusic__track t
+             LEFT JOIN furumusic__release r ON r.id = t.release_id
+             LEFT JOIN furumusic__media_file mf ON mf.id = t.audio_file_id
+            ORDER BY t.id"#,
+    )
+    .fetch_all(&ctx.pool)
+    .await?;
+
+    if rows.is_empty() {
+        log.info("Archive cleanup: no tracks found");
+        return Ok(());
+    }
+
+    log.info(&format!(
+        "Archive cleanup: checking {} track audio reference(s)",
+        rows.len()
+    ));
+
+    let mut missing_tracks = Vec::new();
+    let mut skipped_io_errors = 0u64;
+
+    for row in rows {
+        let Some(media_file_id) = row.media_file_id else {
+            missing_tracks.push(MissingTrack::from_row(row, MissingReason::MissingMediaRow));
+            continue;
+        };
+
+        let file_type = row.file_type.clone();
+        match file_type.as_deref() {
+            Some("audio") => {}
+            Some(file_type) => {
+                missing_tracks.push(MissingTrack::from_row(
+                    row,
+                    MissingReason::InvalidMediaType(file_type.to_owned()),
+                ));
+                continue;
+            }
+            None => {
+                missing_tracks.push(MissingTrack::from_row(row, MissingReason::MissingMediaRow));
+                continue;
+            }
+        }
+
+        let Some(file_path) = row
+            .file_path
+            .as_deref()
+            .map(str::trim)
+            .filter(|path| !path.is_empty())
+        else {
+            missing_tracks.push(MissingTrack::from_row(row, MissingReason::EmptyPath));
+            continue;
+        };
+
+        let absolute_path = crate::media_paths::resolve_media_file_path(storage_dir, file_path);
+        match tokio::fs::metadata(&absolute_path).await {
+            Ok(meta) if meta.is_file() => {}
+            Ok(_) => {
+                missing_tracks.push(MissingTrack::from_row(row, MissingReason::NotRegularFile));
+            }
+            Err(err) if err.kind() == ErrorKind::NotFound => {
+                missing_tracks.push(MissingTrack::from_row(row, MissingReason::MissingFile));
+            }
+            Err(err) => {
+                skipped_io_errors += 1;
+                log.warn(&format!(
+                    "Archive cleanup: skipping track {} media_file_id={media_file_id}; cannot inspect {}: {err}",
+                    row.track_id,
+                    absolute_path.display()
+                ));
+            }
+        }
+    }
+
+    if missing_tracks.is_empty() {
+        log.info(&format!(
+            "Archive cleanup: all checked tracks have readable audio files; skipped_io_errors={skipped_io_errors}"
+        ));
+        return Ok(());
+    }
+
+    for (index, track) in missing_tracks.iter().take(SAMPLE_LOG_LIMIT).enumerate() {
+        log.warn(&format!(
+            "Archive cleanup: deleting stale track {} \"{}\"{}{} ({})",
+            track.track_id,
+            track.track_title,
+            track
+                .release_title
+                .as_deref()
+                .map(|title| format!(" from \"{title}\""))
+                .unwrap_or_default(),
+            track
+                .file_path
+                .as_deref()
+                .map(|path| format!(", path={path}"))
+                .unwrap_or_default(),
+            track.reason
+        ));
+        if index + 1 == SAMPLE_LOG_LIMIT && missing_tracks.len() > SAMPLE_LOG_LIMIT {
+            log.warn(&format!(
+                "Archive cleanup: suppressing per-track logs for remaining {} stale track(s)",
+                missing_tracks.len() - SAMPLE_LOG_LIMIT
+            ));
+        }
+    }
+
+    let track_ids = unique_sorted(
+        missing_tracks
+            .iter()
+            .map(|track| track.track_id)
+            .collect::<Vec<_>>(),
+    );
+    let media_file_ids = unique_sorted(
+        missing_tracks
+            .iter()
+            .filter_map(|track| track.media_file_id)
+            .collect::<Vec<_>>(),
+    );
+    let release_ids = unique_sorted(
+        missing_tracks
+            .iter()
+            .map(|track| track.release_id)
+            .collect::<Vec<_>>(),
+    );
+
+    let stats =
+        delete_tracks_and_unreferenced_audio_media(&ctx.pool, &track_ids, &media_file_ids).await?;
+    let empty_release_count = count_empty_releases(&ctx.pool, &release_ids).await?;
+
+    log.info(&format!(
+        "Archive cleanup: deleted {} track(s), {} unreferenced audio media_file row(s); cleared playback_states={}, playlist_entries={}, likes={}, play_history={}, popularity_history={}, scrobble_outbox={}, track_genres={}, entity_tags={}, external_ids={}, track_artists={}; skipped_io_errors={skipped_io_errors}; empty_releases_left={empty_release_count}",
+        stats.tracks_deleted,
+        stats.media_files_deleted,
+        stats.playback_states_cleared,
+        stats.playlist_entries_deleted,
+        stats.likes_deleted,
+        stats.play_history_deleted,
+        stats.popularity_history_deleted,
+        stats.scrobble_outbox_deleted,
+        stats.track_genres_deleted,
+        stats.entity_tags_deleted,
+        stats.external_ids_deleted,
+        stats.track_artists_deleted,
+    ));
+
+    Ok(())
+}
+
+impl MissingTrack {
+    fn from_row(row: TrackFileRow, reason: MissingReason) -> Self {
+        Self {
+            track_id: row.track_id,
+            track_title: row.track_title,
+            release_id: row.release_id,
+            release_title: row.release_title,
+            media_file_id: row.media_file_id,
+            file_path: row.file_path,
+            reason,
+        }
+    }
+}
+
+impl std::fmt::Display for MissingReason {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::MissingMediaRow => f.write_str("missing media_file row"),
+            Self::InvalidMediaType(file_type) => write!(f, "invalid media_file type {file_type:?}"),
+            Self::EmptyPath => f.write_str("empty media file path"),
+            Self::MissingFile => f.write_str("audio file not found on disk"),
+            Self::NotRegularFile => f.write_str("audio path is not a regular file"),
+        }
+    }
+}
+
+fn unique_sorted(values: Vec<i64>) -> Vec<i64> {
+    values
+        .into_iter()
+        .collect::<BTreeSet<_>>()
+        .into_iter()
+        .collect()
+}
+
+async fn delete_tracks_and_unreferenced_audio_media(
+    pool: &PgPool,
+    track_ids: &[i64],
+    media_file_ids: &[i64],
+) -> anyhow::Result<DeleteStats> {
+    if track_ids.is_empty() {
+        return Ok(DeleteStats::default());
+    }
+
+    let mut tx = pool.begin().await?;
+    let mut stats = DeleteStats::default();
+
+    stats.playback_states_cleared = sqlx::query(
+        r#"UPDATE furumusic__playback_state
+              SET current_track_id = NULL
+            WHERE current_track_id = ANY($1)"#,
+    )
+    .bind(track_ids)
+    .execute(&mut *tx)
+    .await?
+    .rows_affected();
+
+    stats.playlist_entries_deleted =
+        delete_track_rows(&mut tx, "furumusic__playlist_track", track_ids).await?;
+    stats.likes_deleted =
+        delete_track_rows(&mut tx, "furumusic__user_liked_track", track_ids).await?;
+    stats.play_history_deleted =
+        delete_track_rows(&mut tx, "furumusic__play_history", track_ids).await?;
+    stats.popularity_history_deleted =
+        delete_track_rows(&mut tx, "furumusic__track_popularity_history", track_ids).await?;
+    stats.scrobble_outbox_deleted =
+        delete_track_rows(&mut tx, "furumusic__lastfm_scrobble_outbox", track_ids).await?;
+    stats.track_genres_deleted =
+        delete_track_rows(&mut tx, "furumusic__track_genre", track_ids).await?;
+
+    stats.entity_tags_deleted = sqlx::query(
+        r#"DELETE FROM furumusic__entity_genre_tag
+            WHERE entity_kind = 'track'
+              AND entity_id = ANY($1)"#,
+    )
+    .bind(track_ids)
+    .execute(&mut *tx)
+    .await?
+    .rows_affected();
+
+    stats.external_ids_deleted = sqlx::query(
+        r#"DELETE FROM furumusic__external_metadata_id
+            WHERE entity_kind = 'track'
+              AND entity_id = ANY($1)"#,
+    )
+    .bind(track_ids)
+    .execute(&mut *tx)
+    .await?
+    .rows_affected();
+
+    stats.track_artists_deleted =
+        delete_track_rows(&mut tx, "furumusic__track_artist", track_ids).await?;
+
+    stats.tracks_deleted = sqlx::query("DELETE FROM furumusic__track WHERE id = ANY($1)")
+        .bind(track_ids)
+        .execute(&mut *tx)
+        .await?
+        .rows_affected();
+
+    if !media_file_ids.is_empty() {
+        stats.media_files_deleted = sqlx::query(
+            r#"DELETE FROM furumusic__media_file mf
+                WHERE mf.id = ANY($1)
+                  AND mf.file_type = 'audio'
+                  AND NOT EXISTS (
+                        SELECT 1
+                          FROM furumusic__track t
+                         WHERE t.audio_file_id = mf.id
+                            OR t.cover_file_id = mf.id
+                  )
+                  AND NOT EXISTS (
+                        SELECT 1
+                          FROM furumusic__release r
+                         WHERE r.cover_file_id = mf.id
+                  )
+                  AND NOT EXISTS (
+                        SELECT 1
+                          FROM furumusic__artist a
+                         WHERE a.image_file_id = mf.id
+                  )
+                  AND NOT EXISTS (
+                        SELECT 1
+                          FROM furumusic__playlist p
+                         WHERE p.cover_file_id = mf.id
+                  )"#,
+        )
+        .bind(media_file_ids)
+        .execute(&mut *tx)
+        .await?
+        .rows_affected();
+    }
+
+    tx.commit().await?;
+    Ok(stats)
+}
+
+async fn delete_track_rows(
+    tx: &mut sqlx::Transaction<'_, sqlx::Postgres>,
+    table: &str,
+    track_ids: &[i64],
+) -> anyhow::Result<u64> {
+    let sql = format!("DELETE FROM {table} WHERE track_id = ANY($1)");
+    Ok(sqlx::query(&sql)
+        .bind(track_ids)
+        .execute(&mut **tx)
+        .await?
+        .rows_affected())
+}
+
+async fn count_empty_releases(pool: &PgPool, release_ids: &[i64]) -> anyhow::Result<i64> {
+    if release_ids.is_empty() {
+        return Ok(0);
+    }
+
+    let count = sqlx::query_scalar::<_, i64>(
+        r#"SELECT COUNT(*)
+             FROM furumusic__release r
+            WHERE r.id = ANY($1)
+              AND NOT EXISTS (
+                    SELECT 1
+                      FROM furumusic__track t
+                     WHERE t.release_id = r.id
+              )"#,
+    )
+    .bind(release_ids)
+    .fetch_one(pool)
+    .await?;
+
+    Ok(count)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn unique_sorted_deduplicates_ids() {
+        assert_eq!(unique_sorted(vec![3, 1, 3, 2, 1]), vec![1, 2, 3]);
+    }
+
+    #[test]
+    fn missing_reason_display_is_stable() {
+        assert_eq!(
+            MissingReason::InvalidMediaType("cover_art".to_owned()).to_string(),
+            "invalid media_file type \"cover_art\""
+        );
+        assert_eq!(
+            MissingReason::MissingFile.to_string(),
+            "audio file not found on disk"
+        );
+    }
+}
diff --git a/src/jobs/inbox_discover.rs b/src/jobs/inbox_discover.rs
index 7c66f36..21e65ab 100644
--- a/src/jobs/inbox_discover.rs
+++ b/src/jobs/inbox_discover.rs
@@ -128,7 +128,11 @@ impl Job for InboxDiscoverJob {
                             v
                         }
                         Err(e) => {
-                            crate::metrics::record_agent_file_hash(hash_start.elapsed(), 0, "error");
+                            crate::metrics::record_agent_file_hash(
+                                hash_start.elapsed(),
+                                0,
+                                "error",
+                            );
                             log.warn(&format!("Failed to hash {}: {e}", file_path.display()));
                             continue;
                         }
diff --git a/src/jobs/inbox_process.rs b/src/jobs/inbox_process.rs
index e892ce8..69557b7 100644
--- a/src/jobs/inbox_process.rs
+++ b/src/jobs/inbox_process.rs
@@ -494,7 +494,12 @@ async fn process_folder_batch(
         .await
         {
             Ok(Ok(results)) => {
-                crate::metrics::record_agent_rag("artist", "ok", rag_start.elapsed(), results.len());
+                crate::metrics::record_agent_rag(
+                    "artist",
+                    "ok",
+                    rag_start.elapsed(),
+                    results.len(),
+                );
                 for a in results {
                     if !all_similar_artists
                         .iter()
@@ -525,7 +530,12 @@ async fn process_folder_batch(
         .await
         {
             Ok(Ok(results)) => {
-                crate::metrics::record_agent_rag("release", "ok", rag_start.elapsed(), results.len());
+                crate::metrics::record_agent_rag(
+                    "release",
+                    "ok",
+                    rag_start.elapsed(),
+                    results.len(),
+                );
                 for r in results {
                     if !all_similar_releases
                         .iter()
diff --git a/src/jobs/mod.rs b/src/jobs/mod.rs
index 903675a..19067b6 100644
--- a/src/jobs/mod.rs
+++ b/src/jobs/mod.rs
@@ -1,3 +1,4 @@
+pub mod archive_cleanup;
 pub mod artwork_backfill;
 pub mod inbox_discover;
 pub mod inbox_process;
diff --git a/src/main.rs b/src/main.rs
index b6ccb99..4d75fca 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -52,6 +52,7 @@ fn build_registry() -> Arc<JobRegistry> {
     registry.register(jobs::inbox_discover::InboxDiscoverJob);
     registry.register(jobs::inbox_process::InboxProcessJob);
     registry.register(jobs::inbox_process::FileProcessJob);
+    registry.register(jobs::archive_cleanup::ArchiveCleanupJob);
     registry.register(jobs::artwork_backfill::ArtworkBackfillJob);
     registry.register(jobs::metadata_backfill::MetadataBackfillJob);
     registry.register(jobs::lastfm_popularity::LastfmPopularityJob);
diff --git a/src/player/mod.rs b/src/player/mod.rs
index 5ef8277..fd4261f 100644
--- a/src/player/mod.rs
+++ b/src/player/mod.rs
@@ -804,6 +804,10 @@ fn normalize_device_id(raw: &str) -> Option<String> {
 }
 
 fn device_name_from_user_agent(user_agent: Option<&str>) -> String {
+    if let Some(name) = native_device_name_from_user_agent(user_agent) {
+        return name;
+    }
+
     let ua = user_agent.unwrap_or_default().to_ascii_lowercase();
     let browser = if ua.contains("edg/") || ua.contains("edgios/") || ua.contains("edga/") {
         "Edge"
@@ -838,8 +842,47 @@ fn device_name_from_user_agent(user_agent: Option<&str>) -> String {
     format!("{browser} on {os}")
 }
 
+fn native_device_name_from_user_agent(user_agent: Option<&str>) -> Option<String> {
+    let raw = user_agent?.trim();
+    for token in raw.split_ascii_whitespace() {
+        let Some((product, version)) = token.split_once('/') else {
+            continue;
+        };
+        if !product.eq_ignore_ascii_case("FurumiAndroid") {
+            continue;
+        }
+
+        let version = sanitize_user_agent_version(version);
+        return Some(match version.as_deref() {
+            Some(version) => format!("Furumi Android {version}"),
+            None => "Furumi Android".to_string(),
+        });
+    }
+    None
+}
+
+fn sanitize_user_agent_version(version: &str) -> Option<String> {
+    let version = version
+        .chars()
+        .take(32)
+        .filter(|ch| ch.is_ascii_alphanumeric() || matches!(ch, '.' | '-' | '_'))
+        .collect::<String>();
+    if version.is_empty() {
+        None
+    } else {
+        Some(version)
+    }
+}
+
 fn device_kind_from_user_agent(user_agent: Option<&str>) -> &'static str {
     let ua = user_agent.unwrap_or_default().to_ascii_lowercase();
+    if ua.contains("furumiandroid/") {
+        return if ua.contains("tablet") || (ua.contains("android") && !ua.contains("mobile")) {
+            "tablet"
+        } else {
+            "phone"
+        };
+    }
     if ua.contains("iphone") || (ua.contains("android") && ua.contains("mobile")) {
         "phone"
     } else if ua.contains("ipad") || ua.contains("tablet") || ua.contains("android") {
@@ -849,6 +892,33 @@ fn device_kind_from_user_agent(user_agent: Option<&str>) -> &'static str {
     }
 }
 
+#[cfg(test)]
+mod device_tests {
+    use super::*;
+
+    #[test]
+    fn detects_furumi_android_native_client() {
+        let user_agent = Some("FurumiAndroid/1.0 Android Mobile");
+
+        assert_eq!(
+            device_name_from_user_agent(user_agent),
+            "Furumi Android 1.0"
+        );
+        assert_eq!(device_kind_from_user_agent(user_agent), "phone");
+    }
+
+    #[test]
+    fn keeps_browser_fallback_for_generic_android_user_agents() {
+        let user_agent = Some("Mozilla/5.0 Android Mobile");
+
+        assert_eq!(
+            device_name_from_user_agent(user_agent),
+            "Browser on Android"
+        );
+        assert_eq!(device_kind_from_user_agent(user_agent), "phone");
+    }
+}
+
 #[derive(Debug, sqlx::FromRow)]
 struct LastfmAccountApiRow {
     session_key: String,
diff --git a/src/scheduler/mod.rs b/src/scheduler/mod.rs
index f3a1640..c775324 100644
--- a/src/scheduler/mod.rs
+++ b/src/scheduler/mod.rs
@@ -1372,6 +1372,7 @@ async fn run_scheduled_job(
     if !live_config.agent_enabled
         && job_name != "lastfm_popularity"
         && job_name != "lastfm_scrobble"
+        && job_name != "archive_cleanup"
         && job_name != "artwork_backfill"
     {
         tracing::warn!(job = job_name, "Skipping: agent_enabled=false");