fix(admin): re-acquisition settings take effect without a restart, and say why a save was refused (#3936 , #3937 )

#3936: Router() built a reacquisition.SettingsService of its own, so a save from the admin card refreshed that instance's cache while the sweeper in main.go kept serving what it loaded at boot. The card showed the new policy, the feature ran the old one, and only a restart reconciled them. main.go now hands its instance to the server (srv.ReacqSettings), as it already did for RecSettings, TagSettings and FingerprintSettings, and Router() constructs one only when that field is nil. The regression test saves through the router and reads the sweeper's instance. #3937: the card's catch tested `e instanceof Error`, but api.put throws a plain {code, message, status} object, so every reason the server gave was discarded in favour of "Couldn't save settings." It now uses errMessage, which appends the server's message for invalid_setting. Its test rejected with an Error no code path produces, so it passed throughout; it now rejects with what the client actually throws. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SQ31KQpYbStyK5y58UmPLH
test(api): pass fingerprint settings to Mount in the route-registration test (M400 #3913 )
2026-09-11 20:15:35 -04:00 · 2026-09-11 17:58:05 -04:00 · 2026-09-11 17:53:55 -04:00 · 2026-09-11 17:31:42 -04:00 · 2026-09-11 17:25:06 -04:00 · 2026-09-11 17:11:11 -04:00
195 changed files with 13598 additions and 798 deletions
@@ -6,9 +6,20 @@
 **/build
 web/build

-# Flutter mobile client — built separately on developer machines / Flutter CI.
-# Including it in the Go build context wastes ~70 files and invalidates the
-# `COPY . .` layer cache on every Flutter-only change.
+# The Android client — built by its own job, never from this context. The APK
+# reaches the image through client/, downloaded as a CI artifact, so nothing
+# here reads android/ sources.
+#
+# This block named `flutter_client/` until 2026-09-10 and lost its PATTERN when
+# that tree was deleted, leaving a comment describing an exclusion that was no
+# longer happening. android/ never took its place, so 4.1 MB of Gradle project
+# has been entering the context and busting the `COPY . .` layer on every
+# Android-only change.
+android/
+
+# Local `make build` output — an 18 MB binary the image never uses, since the
+# builder stage compiles its own.
+bin/

 # Docs and IDE noise
 docs/
@@ -26,5 +37,8 @@ docs/
 !.env.example

 # CI workflow files don't need to ship in the image.
-.forgejo/
-.github/
+#
+# This said `.forgejo/` and `.github/` — neither of which this repo has. Gitea
+# Actions reads `.gitea/`, so the one directory that actually exists was the
+# one not excluded, and every workflow edit invalidated the context.
+.gitea/
@@ -80,15 +80,12 @@ jobs:

      - name: Upload debug APK
        if: github.event_name == 'push' && github.ref == 'refs/heads/main'
-        # Mirrored action, never actions/upload-artifact. @v4+ throws
-        # GHESNotSupportedError client-side on the hostname (no server setting
-        # reaches that check), and @v3 is worse — it reports success while Gitea
-        # serves artifacts back only through the v4 API, so the upload is stored
-        # and invisible to every retrieval path. @v3 is what left 72 unreachable
-        # artifacts on this repo. Pinned by SHA because the mirror auto-syncs;
-        # full URL because DEFAULT_ACTIONS_URL sends bare owner/repo to github.com.
-        # See Scribe issues 2255 / 2270.
-        uses: https://git.fabledsword.com/bvandeusen/upload-artifact@cb8afe72b42edc798abfb8fcb556cf660d894245
+        # Stock action: it works on this forge since the runner moved to
+        # gitea/runner 3.x, which edits upload-artifact's client-side GHES refusal
+        # out of the action bundle (Scribe snippet #2271). Never @v3 — it reports
+        # success while Gitea serves artifacts back only through the v4 API, and
+        # it is what left 72 unreachable artifacts on this repo (Scribe 2270).
+        uses: actions/upload-artifact@v7
        with:
          name: minstrel-android-debug-${{ github.sha }}
          path: android/app/build/outputs/apk/debug/app-debug.apk
@@ -2,15 +2,71 @@ name: release

 # Builds and pushes the minstrel container image to the Gitea registry.
 #
-#   push to main             → :main and :latest (latest-release APK bundled)
-#   push tag vYYYY.MM.DD     → :vYYYY.MM.DD and :latest (freshly-built APK bundled)
-#   workflow_dispatch        → manual trigger (same rules based on the ref)
+#   push to dev                  → :dev            (freshly-built dev APK bundled)
+#   push to main                 → :latest + :<sha> (latest-release APK bundled)
+#   push tag vYYYY.MM.DD.HHMM    → :latest          (fresh APK bundled)
+#   workflow_dispatch            → manual trigger (same rules based on the ref)
 #
-# Release model: per-day CalVer tags (no trailing patch digit). The day's
-# tag is intentionally mutable — if a second release happens the same day,
-# move the tag with `git push -f origin vYYYY.MM.DD` and the image tag of
-# the same name gets overwritten. :latest is updated by every main push
-# AND every tag push, so it always reflects the newest blessed image.
+# That is the whole tag map, and it is family rule 145 + 147 as written.
+#
+# :<sha> on main is the ROLLBACK UNIT — every production commit addressable
+# without a release ceremony. It is minted only on main, where rollback is
+# actually worth having: merges are gated (rule 2) so they number in the dozens
+# per year, while on dev they would be one per push, forever, for a channel
+# whose entire contract is that it moves.
+#
+# There are NO :<version> image tags. This repo published :vYYYY.MM.DD.HHMM
+# until 2026-09-10 and it was the inverse of the rule on both counts — minting
+# a version tag nobody pinned while the rollback unit the rule names did not
+# exist here at all. Git and the build's own self-reported version answer
+# "which build is this"; a third name for the same thing is upkeep for a model
+# we do not run. Operator, 2026-09-10: "only things like the APK need that kind
+# of versioning for their update process."
+#
+# There is no :main either. :latest tracks main's tip with no gate between them
+# (rule 147), so a second name for the same image sends readers looking for a
+# distinction that does not exist.
+#
+# The dev channel exists so testing a build does not require shipping one.
+# Before it, the only way to get an APK onto a phone was to cut a release,
+# which made `main` the staging area by default. `:dev` carries its own
+# freshly-built APK, signed with the SAME key as release builds — a different
+# key cannot install over the stable app, so anyone crossing channels would
+# have to uninstall and lose their data.
+#
+# :dev is published ALONE, with no per-commit tag. A rolling channel is
+# rolling by definition; a commit-addressable image for it would be a
+# rollback target nobody ever pulls, kept forever. Recovery on dev is to fix
+# forward.
+#
+# Note what this repo does NOT need: a cross-repo dispatch to refresh the
+# channel when its bundled APK is rebuilt. That mechanism exists elsewhere in
+# the family because the app and the server live in separate repos. Minstrel
+# is a monorepo — one push builds the APK and the image in the same run from
+# the same commit, so the channel cannot go stale against its own artifact.
+# The requirement is satisfied structurally; copying the mechanism would add
+# a moving part to fix a problem that does not exist here.
+#
+# Release model: the tag IS the artifact's version name with a `v` in front.
+# `v2026.09.10.1432` and `2026.09.10.1432` are the same string, derived from
+# the tagged commit's UTC timestamp — so there is no mismatch to reconcile
+# between what the tag says and what the APK reports, and nothing to look up
+# when minting one.
+#
+# TAGS ARE IMMUTABLE. Never move, retarget or delete a published tag. A
+# same-day second release is not a collision — HHMM makes every tag unique
+# by construction, so the answer is simply another tag.
+#
+# This block used to say the opposite: that the per-day tag was
+# "intentionally mutable" and that a same-day re-cut should
+# `git push -f origin vYYYY.MM.DD`. That instruction is what the family
+# rulebook now forbids outright, and it has incidents behind it — moving a
+# same-day tag forward once took a published release down with it. Anyone
+# installing from a tag is holding something the tag no longer points at,
+# which is a worse failure than an extra row in the tag list.
+#
+# :latest is updated by every main push AND every tag push, so it always
+# reflects the newest blessed image.
 #
 # APK pipeline: on tag pushes the android-release job builds + signs the
 # Android APK and uploads it as a workflow artifact. The image-release
@@ -24,33 +80,37 @@ name: release
 # :latest (not just tags), a main build with no APK would silently strip
 # the in-app update channel off :latest until the next release. So on
 # non-tag builds image-release pulls the MOST RECENT release's signed APK
-# and reconstructs its exact versionName (tag + commit-count, the same
-# formula android-release bakes in) for the version sidecar — no rebuild,
-# just rebundle. Tag builds keep bundling their own freshly-built APK.
+# AND the version sidecar published beside it — the recorded values, not
+# recomputed ones — so no rebuild is needed, just a rebundle. Tag builds
+# keep bundling their own freshly-built APK.
 #
 # Android testing (lint + detekt + unit tests, debug APK upload on main)
 # lives in android.yml and runs independently on every push.

 on:
  push:
-    branches: [main]
+    branches: [main, dev]
    tags: ['v*']
    paths-ignore:
      - 'docs/**'
      - '**/*.md'
  workflow_dispatch:

-# Force-moving the per-day tag (or rapidly re-pushing to main) should
-# supersede the in-flight build — the operator explicitly wants the
-# later commit to win.
+# A rapid re-push to main should supersede the in-flight build — the
+# operator explicitly wants the later commit to win. Tags no longer enter
+# into this: they are immutable and unique, so no tag build can ever be
+# superseded by another run on the same ref.
 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

 jobs:
  android-release:
-    name: Build signed APK (tag releases only)
-    if: startsWith(github.ref, 'refs/tags/v')
+    name: Build signed APK (releases and dev)
+    # Also builds on `dev`, which is what makes a test channel possible at
+    # all. Without it the only way to get a build onto a phone was to cut a
+    # release, which quietly turns `main` into the staging area.
+    if: startsWith(github.ref, 'refs/tags/v') || github.ref == 'refs/heads/dev'
    runs-on: flutter-ci
    container:
      image: git.fabledsword.com/bvandeusen/ci-android:36
@@ -75,14 +135,18 @@ jobs:
    outputs:
      version_name: ${{ steps.ver.outputs.name }}
      version_code: ${{ steps.ver.outputs.code }}
+      channel: ${{ steps.ver.outputs.channel }}

    steps:
      - name: Checkout
        uses: actions/checkout@v4
        with:
-          # fetch-depth: 0 retrieves full history; default shallow clone
-          # would return 1 for `git rev-list --count HEAD`, breaking the
-          # iteration suffix.
+          # Full history. The version name now reads only the tip commit's
+          # timestamp, so a shallow clone would technically serve — but this
+          # job derives a value that ships to devices, and a shallow checkout
+          # changes what git-derived values resolve to WITHOUT failing. The
+          # whole failure class here is a green build carrying a wrong
+          # version, so the cheap guarantee is worth keeping.
          fetch-depth: 0

      - name: Compute release version
@@ -91,12 +155,23 @@ jobs:
        working-directory: ${{ github.workspace }}
        run: |
          set -euo pipefail
-          TAG="${GITHUB_REF#refs/tags/v}"
-          COMMIT_COUNT=$(git rev-list --count HEAD)
-          VERSION_NAME="${TAG}.${COMMIT_COUNT}"
-          echo "name=${VERSION_NAME}" >> "$GITHUB_OUTPUT"
-          echo "code=${COMMIT_COUNT}" >> "$GITHUB_OUTPUT"
-          echo "::notice::APK version: ${VERSION_NAME} (code=${COMMIT_COUNT})"
+          # The derivation lives in ci/version.sh, not here, so it can be
+          # executed by a test on every push. Anything inline in this file is
+          # unverifiable until a release is already running.
+          out="$(ci/version.sh HEAD)"
+          printf '%s\n' "${out}" >> "$GITHUB_OUTPUT"
+
+          # The channel is a property of the LANE, not of the commit, which is
+          # why it is derived here rather than in version.sh. Same commit built
+          # on dev and on main reports the same NAME and differs only here —
+          # that is the whole point of separating the two values.
+          if [ "${GITHUB_REF}" = "refs/heads/dev" ]; then
+            channel=dev
+          else
+            channel=stable
+          fi
+          echo "channel=${channel}" >> "$GITHUB_OUTPUT"
+          echo "::notice::APK $(printf '%s' "${out}" | tr '\n' ' ') channel=${channel}"

      # Checked BEFORE the expensive work, not after it. "Attach APK to gitea
      # Release" below resolves the release by tag and fails if it is absent —
@@ -108,6 +183,7 @@ jobs:
      # the release together, so this passes). A bare `git push origin vX` is the
      # case this catches.
      - name: Release must exist for this tag
+        if: startsWith(github.ref, 'refs/tags/v')
        shell: bash
        working-directory: ${{ github.workspace }}
        env:
@@ -156,13 +232,12 @@ jobs:
            -PMINSTREL_VERSION_CODE=${{ steps.ver.outputs.code }}

      - name: Upload APK as workflow artifact
-        # Mirrored action, never actions/upload-artifact — @v4+ refuses on the
-        # hostname, @v3 uploads something Gitea will never serve back. This is
-        # the producing half of a pair: image-release downloads `minstrel-apk`
-        # below with the matching download-artifact mirror. Both must stay on
-        # the v4 protocol — mixing a v3 upload with a v4 download (or the
-        # reverse) yields an empty listing, not an error. See Scribe 2255 / 2270.
-        uses: https://git.fabledsword.com/bvandeusen/upload-artifact@cb8afe72b42edc798abfb8fcb556cf660d894245
+        # Stock action (snippet #2271) — never @v3, which uploads something Gitea
+        # will never serve back. This is the producing half of a pair:
+        # image-release downloads `minstrel-apk` below. Any upload v4+ pairs with
+        # any download v4+ on this forge (every combination tested 2026-09-10,
+        # Scribe spike #3843), so the two pins need not move together.
+        uses: actions/upload-artifact@v7
        with:
          name: minstrel-apk
          path: android/app/build/outputs/apk/release/app-release.apk
@@ -171,9 +246,15 @@ jobs:
          if-no-files-found: error

      - name: Attach APK to gitea Release
+        # Tag releases only. A dev build has no Release to hang assets on and
+        # does not need one — the :dev image bundles the APK, and the server
+        # serves it from /api/client/apk like any other.
+        if: startsWith(github.ref, 'refs/tags/v')
        shell: bash
        env:
          CI_TOKEN: ${{ secrets.CI_TOKEN }}
+          VERSION_NAME: ${{ steps.ver.outputs.name }}
+          VERSION_CODE: ${{ steps.ver.outputs.code }}
        run: |
          set -euxo pipefail
          TAG="${GITHUB_REF#refs/tags/}"
@@ -181,6 +262,20 @@ jobs:
          APK_PATH="app/build/outputs/apk/release/app-release.apk"
          ls -lh "${APK_PATH}"

+          # Publish the version sidecar as a release asset next to the APK.
+          #
+          # This is what lets a later :latest build stop RECONSTRUCTING the
+          # bundled APK's version and simply read what was recorded. The
+          # ordering key in particular cannot be re-derived after the fact —
+          # it is build-time minutes, so once this job ends the value exists
+          # nowhere else. Reconstruction could only ever recover the name,
+          # and only by duplicating a formula that then has to be kept in
+          # step across two files.
+          SIDECAR_PATH="/tmp/minstrel.apk.version"
+          printf '{"name":"%s","code":%s,"channel":"stable"}\n' \
+            "${VERSION_NAME}" "${VERSION_CODE}" > "${SIDECAR_PATH}"
+          cat "${SIDECAR_PATH}"
+
          RELEASE_JSON="$(curl -fsSL \
            -H "Authorization: token ${CI_TOKEN}" \
            "https://git.fabledsword.com/api/v1/repos/${REPO}/releases/tags/${TAG}")"
@@ -202,6 +297,20 @@ jobs:
            exit 1
          fi

+          # Same treatment for the sidecar. Named `.apk.version` so the
+          # downloader's `\.apk$` match cannot pick it up by mistake.
+          SIDECAR_HTTP=$(curl -sS -L -o /tmp/upload-sidecar.out -w '%{http_code}' \
+            -H "Authorization: token ${CI_TOKEN}" \
+            -F "attachment=@${SIDECAR_PATH}" \
+            "https://git.fabledsword.com/api/v1/repos/${REPO}/releases/${RELEASE_ID}/assets?name=minstrel-${TAG}.apk.version")
+          echo "sidecar_upload_http=${SIDECAR_HTTP}"
+          cat /tmp/upload-sidecar.out || true
+          echo
+          if [ "${SIDECAR_HTTP}" -lt 200 ] || [ "${SIDECAR_HTTP}" -ge 300 ]; then
+            echo "::error::version sidecar upload returned HTTP ${SIDECAR_HTTP}"
+            exit 1
+          fi
+
  image-release:
    name: Build + push container image
    # `needs:` waits for android-release. For tag pushes android-release
@@ -222,11 +331,16 @@ jobs:
      - name: Checkout
        uses: actions/checkout@v4
        with:
-          # Full history + tags so non-tag :latest builds can resolve the
-          # latest release tag's commit count and reconstruct the bundled
-          # APK's exact versionName (see "Bundle latest release APK" below).
+          # Full history, and rule 149 names this specifically: any job that
+          # DERIVES the version name needs it, because a shallow clone changes
+          # what git-derived values resolve to WITHOUT failing — a too-low
+          # value, silently, with every lane green.
+          #
+          # This job was depth-1 while it took the version from GITHUB_REF. It
+          # now runs ci/version.sh itself, because with :<version> image tags
+          # gone the server's self-reported version is the only thing that says
+          # which build an image is.
          fetch-depth: 0
-          fetch-tags: true

      - name: Detect buildable project
        id: guard
@@ -244,21 +358,68 @@ jobs:
        if: steps.guard.outputs.ready == 'true'
        shell: bash
        run: |
-          if [[ "${GITHUB_REF}" == refs/tags/v* ]]; then
-            VERSION="${GITHUB_REF#refs/tags/}"
-            echo "args=-t ${IMAGE}:${VERSION} -t ${IMAGE}:latest" >> "$GITHUB_OUTPUT"
-            echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
-            echo "::notice::Release build: ${VERSION} + latest"
-          else
-            # Main is the protected, post-PR-merge branch. Treat it as the
-            # rolling stable channel — every main push moves :latest.
-            # Pinned consumers can target :vYYYY.MM.DD; everyone else
-            # gets the newest main.
-            echo "args=-t ${IMAGE}:main -t ${IMAGE}:latest" >> "$GITHUB_OUTPUT"
-            echo "version=main" >> "$GITHUB_OUTPUT"
-            echo "::notice::Main-branch build: :main + :latest"
+          set -euo pipefail
+
+          # THE VERSION, and it is derived the same way on every ref — the
+          # branch decides the CHANNEL, never the version (family rule 149).
+          #
+          # This used to be three different things: the literal string "main"
+          # on main, "dev" on dev, and the tag name on a tag. None of them
+          # ordered, and the first two were the same string forever — two dev
+          # images eight weeks apart were indistinguishable in the UI. That
+          # mattered little while :vYYYY.MM.DD.HHMM existed to identify a
+          # build; with version image tags gone, this IS how an operator tells
+          # which build a container is running.
+          #
+          # `sed -n s///p` rather than `grep`: it exits 0 when nothing matches,
+          # so the empty check below is actually reachable. A grep here would
+          # kill the step at the assignment under the runner's pipefail — the
+          # exact bug that took down the first main build after the version
+          # rework.
+          VERSION="$(ci/version.sh HEAD | sed -n 's/^name=//p')"
+          if [ -z "${VERSION}" ]; then
+            echo "::error::could not derive a build version from ci/version.sh"
+            exit 1
          fi

+          if [[ "${GITHUB_REF}" == refs/tags/v* ]]; then
+            # A release refreshes the CHANNEL and mints nothing else.
+            #
+            # The tag build exists to produce the signed APK and attach it to
+            # the release; the image it rebuilds is the SAME SOURCE as the main
+            # build minutes earlier, differing only in which APK is baked in.
+            # Rule 145 is explicit about that case: when the same source is
+            # rebuilt with different contents, publish the moving channel tag
+            # and never a commit-addressable one.
+            #
+            # :latest must move here rather than waiting for the next main
+            # push, or the channel would carry the PREVIOUS release's APK
+            # indefinitely — a channel that cannot refresh itself (rule 146).
+            CHANNEL=stable
+            echo "args=-t ${IMAGE}:latest" >> "$GITHUB_OUTPUT"
+            echo "::notice::Release build ${VERSION}: refreshing :latest around the new APK"
+          elif [[ "${GITHUB_REF}" == "refs/heads/dev" ]]; then
+            # The rolling test channel, and :dev ALONE — deliberately no
+            # per-commit tag. A rolling channel is rolling by definition, so a
+            # commit-addressable image here would be a rollback target nobody
+            # has ever pulled, accumulating in the registry forever. Recovery
+            # on dev is to fix forward.
+            CHANNEL=dev
+            echo "args=-t ${IMAGE}:dev" >> "$GITHUB_OUTPUT"
+            echo "::notice::Dev-branch build ${VERSION}: :dev"
+          else
+            # The production line: :latest tracks main's tip (rule 147) and
+            # :<sha> is the rollback unit (rule 145). Full 40-char SHA, matching
+            # the family's other repos, so a rollback target is addressable
+            # straight from the commit anyone is reading.
+            CHANNEL=stable
+            echo "args=-t ${IMAGE}:latest -t ${IMAGE}:${GITHUB_SHA}" >> "$GITHUB_OUTPUT"
+            echo "::notice::Main-branch build ${VERSION}: :latest + :${GITHUB_SHA}"
+          fi
+
+          echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
+          echo "channel=${CHANNEL}" >> "$GITHUB_OUTPUT"
+
      - name: Registry login
        if: steps.guard.outputs.ready == 'true'
        shell: bash
@@ -267,54 +428,57 @@ jobs:
            | docker login git.fabledsword.com -u "${{ github.actor }}" --password-stdin

      - name: Download signed APK artifact
-        # Tag pushes only — android-release just produced this. Non-tag
-        # builds take the "Bundle latest release APK" path below instead.
-        if: steps.guard.outputs.ready == 'true' && startsWith(github.ref, 'refs/tags/v')
-        # Consuming half of the pair — never actions/download-artifact. Same fork,
-        # same reason: upstream's client-side GHES check rejects this hostname
-        # before it connects. bvandeusen/download-artifact mirrors
-        # code.forgejo.org/forgejo/download-artifact.
-        #
-        # SHA below is that fork's `v6` tag. Match on @actions/artifact, NOT on
-        # the action's own version number — the two actions release on unrelated
-        # cadences, and download v5 would pair a ^2.3.2 client with this file's
-        # ^4.0.0 uploader. v6 is the tag whose bundled library major (^4.0.0) is
-        # the same one proven against this instance by the upload side.
-        # Deliberately NOT v7: it moves to node24 and upstream requires runner
-        # >= 2.327.1 for it, which act_runner does not claim to satisfy.
-        # Pinned, not tagged — the mirror auto-syncs every 8h.
-        uses: https://git.fabledsword.com/bvandeusen/download-artifact@8d4e9521a5f7e5f8b6351f341f719f9f45a92a3a
+        # Tag and dev pushes — android-release just produced this. Only `main`
+        # takes the "Bundle latest release APK" path below, because it is the
+        # one ref that moves a channel without building an APK of its own.
+        if: >-
+          steps.guard.outputs.ready == 'true' &&
+          (startsWith(github.ref, 'refs/tags/v') || github.ref == 'refs/heads/dev')
+        # Consuming half of the pair: stock download-artifact, which works here for
+        # the same reason as the upload (gitea/runner 3.x edits the GHES refusal
+        # out of the bundle; snippet #2271). v8 runs on node24, which every
+        # CI-runner image carries — the runner uses the image's own node.
+        uses: actions/download-artifact@v8
        with:
          name: minstrel-apk
          path: client/

      - name: Stage bundled APK + version sidecar
-        if: steps.guard.outputs.ready == 'true' && startsWith(github.ref, 'refs/tags/v')
+        if: >-
+          steps.guard.outputs.ready == 'true' &&
+          (startsWith(github.ref, 'refs/tags/v') || github.ref == 'refs/heads/dev')
        shell: bash
        env:
-          # Pulled from android-release.outputs.version_name so the
-          # sidecar string the server hands clients matches the
-          # versionName baked into the APK they're comparing against.
+          # All three pulled from android-release's outputs so the sidecar the
+          # server hands clients matches exactly what is baked into the APK
+          # they are comparing against.
          APK_VERSION_NAME: ${{ needs.android-release.outputs.version_name }}
+          APK_VERSION_CODE: ${{ needs.android-release.outputs.version_code }}
+          APK_CHANNEL: ${{ needs.android-release.outputs.channel }}
        run: |
          set -euxo pipefail
          # The artifact lands as `app-release.apk` (the original Gradle
          # output name). The Dockerfile COPYs client/* into /app/client/
          # and the server reads minstrel.apk + minstrel.apk.version.
          mv client/app-release.apk client/minstrel.apk
-          echo "${APK_VERSION_NAME}" > client/minstrel.apk.version
+          printf '{"name":"%s","code":%s,"channel":"%s"}\n' \
+            "${APK_VERSION_NAME}" "${APK_VERSION_CODE}" "${APK_CHANNEL}" \
+            > client/minstrel.apk.version
+          cat client/minstrel.apk.version
          ls -lh client/

      - name: Bundle latest release APK (non-tag :latest builds)
        # Main pushes don't build an APK, but they DO move :latest — so
        # without this the in-app update channel would vanish from :latest
        # until the next tag. Pull the most-recent release's signed APK and
-        # reconstruct its exact versionName (${TAG#v}.$(git rev-list --count
-        # TAG) — identical to android-release's formula) so the version
-        # sidecar the server hands clients matches the installed build.
+        # the sidecar published beside it, so what the server reports is what
+        # that build actually recorded rather than something re-derived here.
        # Degrades to an empty client/ (404 update channel) — never a wrong
-        # version — if no release / APK asset / tag-count can be resolved.
-        if: steps.guard.outputs.ready == 'true' && !startsWith(github.ref, 'refs/tags/v')
+        # version — if no release or APK asset can be resolved. That
+        # degradation only actually works because the greps below carry
+        # `|| true`; under the runner's default pipefail a non-matching grep
+        # kills the step instead of falling through to the empty-case branch.
+        if: steps.guard.outputs.ready == 'true' && github.ref == 'refs/heads/main'
        shell: bash
        env:
          CI_TOKEN: ${{ secrets.CI_TOKEN }}
@@ -326,19 +490,40 @@ jobs:
          if [ -z "${REL_JSON}" ]; then
            echo "::notice::no published release — image ships without bundled APK"; exit 0
          fi
-          TAG="$(printf '%s' "${REL_JSON}" | grep -oP '"tag_name":\s*"\K[^"]+' | head -1)"
-          APK_URL="$(printf '%s' "${REL_JSON}" | grep -oP '"browser_download_url":\s*"\K[^"]+' | grep -E '\.apk$' | head -1)"
+          # `|| true` on every one of these, and it is load-bearing rather
+          # than defensive habit. The runner already invokes this shell as
+          # `bash -e -o pipefail`, so a pipeline whose grep matches NOTHING
+          # exits non-zero even though `head` succeeded — and the step dies at
+          # the assignment, before ever reaching the `if` written to handle the
+          # empty case. Every "degrades gracefully" branch below is unreachable
+          # without this.
+          TAG="$(printf '%s' "${REL_JSON}" | grep -oP '"tag_name":\s*"\K[^"]+' | head -1)" || true
+          APK_URL="$(printf '%s' "${REL_JSON}" | grep -oP '"browser_download_url":\s*"\K[^"]+' | grep -E '\.apk$' | head -1)" || true
          if [ -z "${TAG}" ] || [ -z "${APK_URL}" ]; then
            echo "::notice::latest release '${TAG:-?}' has no APK asset — image ships without bundled APK"; exit 0
          fi
-          COUNT="$(git rev-list --count "${TAG}" 2>/dev/null || true)"
-          if [ -z "${COUNT}" ]; then
-            echo "::notice::could not resolve commit count for ${TAG} (tag not fetched?) — skipping APK bundle"; exit 0
-          fi
-          VERSION_NAME="${TAG#v}.${COUNT}"
          curl -fsSL -H "Authorization: token ${CI_TOKEN}" -o client/minstrel.apk "${APK_URL}"
-          echo "${VERSION_NAME}" > client/minstrel.apk.version
-          echo "::notice::bundled release APK ${TAG} as version ${VERSION_NAME}"
+
+          # Take the version the release RECORDED rather than recomputing it.
+          # This used to re-derive the name from the tagged commit, which meant
+          # the formula lived in two files that had to be kept in step, and it
+          # could only ever recover the name — the ordering key is build-time
+          # minutes and does not exist anywhere after that build ends.
+          SIDECAR_URL="$(printf '%s' "${REL_JSON}" | grep -oP '"browser_download_url":\s*"\K[^"]+' | grep -E '\.apk\.version$' | head -1)" || true
+          if [ -n "${SIDECAR_URL}" ]; then
+            curl -fsSL -H "Authorization: token ${CI_TOKEN}" -o client/minstrel.apk.version "${SIDECAR_URL}"
+            cat client/minstrel.apk.version
+          else
+            # Releases published before sidecars were attached. Their name is
+            # still recoverable from the tag, but their ordering key genuinely
+            # is not — so it is reported ABSENT rather than guessed. A wrong
+            # key is an install the platform refuses; an absent one just tells
+            # the client to fall back to comparing names, which is exactly
+            # what those builds already do.
+            echo "::notice::release ${TAG} predates the version sidecar — bundling with name only, no ordering key"
+            printf '{"name":"%s","code":null,"channel":"stable"}\n' "${TAG#v}" > client/minstrel.apk.version
+          fi
+          echo "::notice::bundled release APK from ${TAG}"
          ls -lh client/

      - name: Build and push
@@ -346,6 +531,7 @@ jobs:
        run: |
          docker buildx build \
            --build-arg MINSTREL_VERSION="${{ steps.tags.outputs.version }}" \
+            --build-arg MINSTREL_CHANNEL="${{ steps.tags.outputs.channel }}" \
            --push ${{ steps.tags.outputs.args }} .

  # Verifies a tag release actually ended up complete, and names the specific
@@ -356,8 +542,8 @@ jobs:
  # `failure` with none executed and image-release showed `skipped`. The run was
  # red, but the *release page rendered fine*, and `main`'s own push build had
  # already moved `:latest`, so the code was deployable and nothing looked
-  # obviously wrong. The release was simply missing its APK and its immutable
-  # `:vYYYY.MM.DD` image, which is easy to skim past.
+  # obviously wrong. The release was simply missing its APK and its image,
+  # which is easy to skim past.
  #
  # This job cannot prevent that (the cause was a runner failing to launch, not
  # anything in this file). What it does is turn an incomplete release into an
@@ -408,18 +594,30 @@ jobs:
      # missing when v2026.08.07 had to be re-cut. `always()` on this job means
      # it runs even when image-release failed, so without this the guard would
      # cheerfully verify an incomplete release.
-      - name: Immutable image tag must exist
+      #
+      # This asserted `:${TAG}` — the :vYYYY.MM.DD.HHMM image — until
+      # 2026-09-10. Version image tags are no longer published (rule 145), so
+      # that assertion would now fail every release for a tag nothing mints.
+      # The rollback target it was really protecting is the :<sha> image, which
+      # main's own build published for this same commit before the tag was cut.
+      #
+      # Checking it here earns its keep twice over: it still catches an image
+      # push that silently did not happen, and it additionally proves the
+      # ORDERING — a tag cut on a commit whose main build never completed has
+      # no rollback target, and that is worth failing on rather than
+      # discovering during an incident.
+      - name: Rollback image must exist for the tagged commit
        shell: bash
        run: |
          set -euo pipefail
-          TAG="${GITHUB_REF#refs/tags/}"
          IMAGE="git.fabledsword.com/bvandeusen/minstrel"

          echo "${{ secrets.CI_TOKEN }}" \
            | docker login git.fabledsword.com -u "${{ github.actor }}" --password-stdin

-          if ! docker manifest inspect "${IMAGE}:${TAG}" > /dev/null 2>&1; then
-            echo "::error::image ${IMAGE}:${TAG} was never pushed — the release tag has no immutable image, so there is nothing to pin or roll back to. Re-run this workflow run."
+          if ! docker manifest inspect "${IMAGE}:${GITHUB_SHA}" > /dev/null 2>&1; then
+            echo "::error::image ${IMAGE}:${GITHUB_SHA} does not exist — this commit has no rollback target."
+            echo "::error::That image is published by the MAIN build of this commit, not by the tag build. If main's build never ran or failed, fix that first; a release whose commit cannot be rolled back to is the thing this check exists to refuse."
            exit 1
          fi
-          echo "::notice::image verified: ${IMAGE}:${TAG}"
+          echo "::notice::rollback target verified: ${IMAGE}:${GITHUB_SHA}"
@@ -32,6 +32,12 @@ on:
      - 'cmd/**'
      - '.golangci.yml'
      - '.gitea/workflows/test-go.yml'
+      # The release lane's own trigger is `main` + tags, so nothing it
+      # contains is exercised until a release is already running. These two
+      # entries are what let internal/server/release_version_test.go guard
+      # the version derivation on ordinary dev pushes instead.
+      - 'ci/**'
+      - '.gitea/workflows/release.yml'

 # pull_request trigger intentionally omitted — see test-web.yml for
 # the rationale (single-author repo, push covers PR-merge equivalent).
@@ -12,6 +12,11 @@
 # Test binary, built with `go test -c`
 *.test

+# `make build` output. bin/minstrel was tracked until 2026-09-10 — an 18 MB
+# binary committed by accident, last refreshed by a commit about web test
+# mocks, and re-dirtied by every local build since.
+bin/
+
 # Bundled Android APK + version sidecar (#397). Populated by CI for
 # tag releases; never committed. README in client/ explains the flow.
 client/minstrel.apk
@@ -15,17 +15,32 @@ COPY . .
 # Overwrite the committed placeholder with the freshly-built SPA assets.
 COPY --from=web /web/build ./web/build
 ENV CGO_ENABLED=0
-# Version stamping: release.yml passes the git tag via MINSTREL_VERSION
-# build-arg; local `docker build` falls back to "dev". Surfaced at
-# /healthz for operator-side image-version verification.
+# Version stamping. release.yml passes the DERIVED version name
+# (YYYY.MM.DD.HHMM) and the lane's channel; a local `docker build` falls back
+# to "dev"/"local". Both are surfaced at /healthz.
+#
+# These are two values on purpose (family rule 149): the same commit built on
+# dev and on main reports the same NAME and differs only in CHANNEL. Folding
+# the channel into the version string is what the rule forbids — the version
+# used to BE the channel word here ("main"/"dev"), which meant two dev images
+# eight weeks apart were indistinguishable.
 ARG MINSTREL_VERSION=dev
+ARG MINSTREL_CHANNEL=local
 RUN go build -trimpath \
-    -ldflags="-s -w -X 'git.fabledsword.com/bvandeusen/minstrel/internal/server.ServerVersion=${MINSTREL_VERSION}'" \
+    -ldflags="-s -w \
+      -X 'git.fabledsword.com/bvandeusen/minstrel/internal/server.ServerVersion=${MINSTREL_VERSION}' \
+      -X 'git.fabledsword.com/bvandeusen/minstrel/internal/server.ServerChannel=${MINSTREL_CHANNEL}'" \
    -o /out/minstrel ./cmd/minstrel

 FROM debian:bookworm-slim
+# ffmpeg: duration probes and the exact-tier audio hash (a SHA-256 of the
+# encoded audio packets, so no decode). libchromaprint-tools: fpcalc, the
+# acoustic fingerprint that tells the same recording at two bitrates apart
+# from two different recordings (M400). Both are baked in at build time so a
+# deployed instance never fetches either (rule 164); fpcalc is shelled out
+# rather than bound because CGO_ENABLED=0 above rules out cgo.
 RUN apt-get update \
-    && apt-get install -y --no-install-recommends ca-certificates ffmpeg \
+    && apt-get install -y --no-install-recommends ca-certificates ffmpeg libchromaprint-tools \
    && rm -rf /var/lib/apt/lists/*

 RUN groupadd --system --gid 1000 minstrel \
@@ -37,8 +37,12 @@ services:
    ports: ['4533:4533']
    volumes:
      # Your music library. Point ./music at wherever your audio files
-      # live. Mounted read-only — Minstrel never writes to your library.
-      - ./music:/music:ro
+      # live. Writable, because Minstrel deletes a file when an admin asks
+      # it to (for example, quarantine's "Delete file"). It never moves,
+      # renames or retags anything. The container runs as uid 1000, so that
+      # user needs write access to the folders. Mount it :ro to forbid even
+      # deletes: those actions then refuse, say why, and delete nothing.
+      - ./music:/music
      # Generated data: playlist cover collages, artist art, caches.
      # The path must match MINSTREL_STORAGE_DATA_DIR, which the image
      # sets to /app/data — keep this mount on /app/data or your cache
@@ -47,7 +51,7 @@ services:
    environment:
      MINSTREL_DATABASE_URL: postgres://minstrel:minstrel@db:5432/minstrel?sslmode=disable
      # Colon-separated library roots to scan; must match the container
-      # path of the read-only music mount above (/music here).
+      # path of the music mount above (/music here).
      MINSTREL_LIBRARY_SCAN_PATHS: /music
    depends_on: [db]

@@ -112,11 +116,21 @@ Most operational keys have a `MINSTREL_<SECTION>_<FIELD>` env override. Recommen

 Image tags (`git.fabledsword.com/bvandeusen/minstrel:<tag>`):

- `:latest` — the newest blessed image. Moves on every `main` push **and** every release. Recommended for most operators.
- `:vYYYY.MM.DD` — immutable per-day release tags. Pin one of these for a deployment you don't want moving under you. (Per-day CalVer — no trailing patch digit; a same-day re-cut moves the tag forward.)
- `:main` — the rolling post-merge tip. Same image as `:latest` at push time; choose it if you want to track `main` explicitly rather than the release line.
+- `:latest` — production. Tracks `main`'s tip and moves on every `main` push and every release. What most operators should run.
+- `:<commit-sha>` — the rollback unit. Every `main` push publishes one, so any production commit is addressable without a release ceremony. Immutable: a given SHA tag is never re-pushed. Pin one if you need a deployment that cannot change under you, and use it to roll back.
+- `:dev` — the rolling test channel, rebuilt on every push to `dev` and carrying its own freshly-built Android APK. Run this to try something before it ships. It moves constantly, has no per-commit tag, and its only recovery path is forward — if a `:dev` image is broken, the fix is the next push, not a rollback.

-Every `:latest` and every `:vYYYY.MM.DD` bundles the current signed Android APK, so the in-app update channel is always live. Database migrations run automatically at startup; rollbacks require restoring a Postgres dump.
+That is the whole tag map. **There are no version-numbered image tags**, and no `:main`. Git and the build's own self-reported version answer "which build is this" — the Settings page shows it, and so does `/healthz`. Release *tags* in git are still `vYYYY.MM.DD.HHMM`; they name a changelog entry and the APK attached to it, not an image.
+
+Rolling back to `:<commit-sha>` pins the **server code** at that commit — not the server-and-app pair. The Android APK is baked in at image build time, so a SHA image carries whichever app was current when that commit was built, which may be older than what `:latest` bundles now. If both halves matter, check what the image bundles rather than trusting the tag's name.
+
+Every `:latest`, `:<commit-sha>` and `:dev` bundles a signed Android APK, so the in-app update channel is always live. All are signed with the same key, so a phone can move between the stable and dev channels without uninstalling — point it at a `:dev` server and the in-app updater offers that channel's build.
+
+The app reports which channel it is on alongside its version, and decides whether an update is available using the build's ordering key rather than its displayed name — the same value Android installs by, so an offer it makes is one the platform will accept.
+
+Database migrations run automatically at startup; rollbacks require restoring a Postgres dump.
+
+Releases up to 2026-09-10 also published a `:vYYYY.MM.DD[.HHMM]` image tag. Those images still exist and still work — they are simply not extended.

 ## Specs

@@ -150,7 +164,7 @@ Two concurrent dev processes:

 - Day-to-day work happens on `dev` (or feature branches merged into `dev`).
 - `main` is **protected** — changes land via PR from `dev`.
- Releases are cut by tagging `v*` off `main`; the release workflow builds and pushes the container image to the Gitea registry.
+- Releases are cut by tagging `v*` off `main`; the release workflow builds the signed APK, attaches it to the release, and refreshes `:latest` around it.

 Task and milestone tracking: Fable (`Minstrel` project, id 12).

@@ -21,13 +21,24 @@ android {
        applicationId = "com.fabledsword.minstrel"
        minSdk = 26
        targetSdk = 36
-        // versionName / versionCode are released-build values injected by
-        // CI from the git tag + commit count. Local / debug builds fall
-        // back to "dev" so the About card reads honestly. Releases ship
-        // versionName="YYYY.MM.DD.<commits>" (e.g. "2026.06.02.142") and
-        // versionCode=<commits>, which is monotonic forever and lets the
-        // shared isVersionNewer comparator distinguish two same-day
-        // re-cuts (the iteration suffix differs).
+        // versionName / versionCode are released-build values injected by CI.
+        // Local / debug builds fall back to "dev" so the About card reads
+        // honestly.
+        //
+        // versionName is "YYYY.MM.DD.HHMM" from the COMMIT's timestamp, so
+        // every lane building this source reports the same string and the
+        // channel is the only thing that differs between them.
+        //
+        // versionCode is minutes since 2020-01-01 at BUILD time. It is the
+        // value the platform decides installs by, so it must be monotonic by
+        // construction.
+        //
+        // This comment used to say versionCode was a commit count and that it
+        // was "monotonic forever". It was neither — a commit count runs ahead
+        // on `dev`, so a dev build outranked the `main` release meant to
+        // replace it and Android refused the install as a downgrade. Worth
+        // knowing the claim was here, stated as a reassurance, while the bug
+        // it denied was live.
        val versionNameOverride =
            (project.findProperty("MINSTREL_VERSION_NAME") as String?)?.takeIf { it.isNotBlank() }
        val versionCodeOverride =
@@ -150,7 +161,6 @@ dependencies {
    implementation(libs.compose.ui)
    implementation(libs.compose.ui.graphics)
    implementation(libs.compose.material3)
-    implementation(libs.compose.ui.text.google.fonts)
    debugImplementation(libs.compose.ui.tooling)
    implementation(libs.compose.ui.tooling.preview)

@@ -15,10 +15,14 @@ import androidx.compose.material3.HorizontalDivider
 import androidx.compose.material3.MaterialTheme
 import androidx.compose.material3.OutlinedButton
 import androidx.compose.material3.Scaffold
+import androidx.compose.material3.SnackbarHost
+import androidx.compose.material3.SnackbarHostState
 import androidx.compose.material3.Text
 import androidx.compose.material3.TextButton
 import androidx.compose.runtime.Composable
+import androidx.compose.runtime.LaunchedEffect
 import androidx.compose.runtime.getValue
+import androidx.compose.runtime.remember
 import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.text.style.TextOverflow
@@ -42,6 +46,12 @@ fun AdminQuarantineScreen(
    viewModel: AdminQuarantineViewModel = hiltViewModel(),
 ) {
    val state by viewModel.uiState.collectAsStateWithLifecycle()
+    val snackbarHostState = remember { SnackbarHostState() }
+    LaunchedEffect(Unit) {
+        viewModel.transientMessages.collect { msg ->
+            snackbarHostState.showSnackbar(msg)
+        }
+    }
    Scaffold(
        contentWindowInsets = ShellContentWindowInsets,
        modifier = Modifier.fillMaxSize(),
@@ -53,6 +63,7 @@ fun AdminQuarantineScreen(
                onBack = { navController.popBackStack() },
            )
        },
+        snackbarHost = { SnackbarHost(snackbarHostState) },
    ) { inner ->
        PullToRefreshScaffold(
            onRefresh = { viewModel.refresh().join() },
@@ -10,10 +10,13 @@ import com.fabledsword.minstrel.events.EventsStream
 import com.fabledsword.minstrel.models.AdminQuarantineItemRef
 import dagger.hilt.android.lifecycle.HiltViewModel
 import kotlinx.coroutines.Job
+import kotlinx.coroutines.channels.Channel
+import kotlinx.coroutines.flow.Flow
 import kotlinx.coroutines.flow.MutableStateFlow
 import kotlinx.coroutines.flow.StateFlow
 import kotlinx.coroutines.flow.asStateFlow
 import kotlinx.coroutines.flow.filter
+import kotlinx.coroutines.flow.receiveAsFlow
 import kotlinx.coroutines.launch
 import javax.inject.Inject

@@ -34,6 +37,15 @@ class AdminQuarantineViewModel @Inject constructor(
    private val internal = MutableStateFlow<AdminQuarantineUiState>(AdminQuarantineUiState.Loading)
    val uiState: StateFlow<AdminQuarantineUiState> = internal.asStateFlow()

+    /**
+     * One-shot messages for the screen's snackbar. A failed action has to say
+     * why: the row quietly reappearing reads as a glitch, and for a Delete
+     * file refused by a read-only library it hides the one thing the
+     * operator can fix (#3918).
+     */
+    private val transientMessagesChannel = Channel<String>(Channel.BUFFERED)
+    val transientMessages: Flow<String> = transientMessagesChannel.receiveAsFlow()
+
    init {
        refresh()
        viewModelScope.launch {
@@ -86,8 +98,9 @@ class AdminQuarantineViewModel @Inject constructor(
            try {
                action(trackId)
            } catch (
-                @Suppress("TooGenericExceptionCaught", "SwallowedException") e: Throwable,
+                @Suppress("TooGenericExceptionCaught") e: Throwable,
            ) {
+                transientMessagesChannel.trySend(ErrorCopy.fromThrowable(e))
                refresh()
            }
        }
@@ -37,18 +37,35 @@ object ErrorCopy {
     * as connection failures.
     */
    fun fromThrowable(t: Throwable): String = when (t) {
-        is HttpException -> messageFor(codeFromHttp(t))
+        is HttpException -> fromHttp(t)
        is IOException -> messageFor("connection_refused")
        else -> TABLE.getValue("unknown")
    }

-    private fun codeFromHttp(e: HttpException): String {
+    /**
+     * Codes whose server message carries specifics the operator needs in
+     * order to act — which directory, which uid — that fixed copy cannot say.
+     * For these the message follows the copy (#3918). Kept to a named set on
+     * purpose: most server messages are internal detail. Mirrors web's
+     * errors.ts.
+     */
+    private val DETAIL_CODES = setOf("library_not_writable", "file_delete_failed")
+
+    private fun fromHttp(e: HttpException): String {
+        val body = bodyFromHttp(e)
+        val copy = messageFor(body.code.ifEmpty { "unknown" })
+        return if (body.code in DETAIL_CODES && body.message.isNotBlank()) {
+            "$copy ${body.message}"
+        } else {
+            copy
+        }
+    }
+
+    private fun bodyFromHttp(e: HttpException): Body {
        val raw = runCatching { e.response()?.errorBody()?.string() }.getOrNull()
-            ?: return "unknown"
-        val code = runCatching { json.decodeFromString<Envelope>(raw).error?.code }
-            .getOrNull()
-            .orEmpty()
-        return code.ifEmpty { "unknown" }
+            ?: return Body()
+        return runCatching { json.decodeFromString<Envelope>(raw).error }
+            .getOrNull() ?: Body()
    }

    private val TABLE: Map<String, String> = mapOf(
@@ -99,6 +116,8 @@ object ErrorCopy {
        "request_not_pending" to "This request is no longer pending.",
        "request_not_found" to "That request no longer exists.",
        "track_not_found" to "That track no longer exists.",
+        "library_not_writable" to "The music library isn't writable by the server.",
+        "file_delete_failed" to "The file couldn't be deleted.",
        "album_not_found" to "That album no longer exists.",
        "artist_not_found" to "That artist no longer exists.",
        "playlist_not_found" to "That playlist no longer exists.",
@@ -1,15 +1,26 @@
 package com.fabledsword.minstrel.models

 /**
- * Wire shape returned by `GET /api/client/version`. Mirrors
- * the Flutter client's `UpdateInfo`.
+ * The server-bundled APK, as reported by `GET /api/client/version`.
 *
- * `version` is the server-bundled APK version (may have a leading
- * "v" from the git tag); `apkUrl` is server-relative (e.g.
- * `/api/client/apk`); `sizeBytes` is the download size.
+ * Three values that are deliberately kept apart:
+ *
+ *  - [version] is a LABEL for people — "YYYY.MM.DD.HHMM", derived from the
+ *    build's commit, so two channels carrying the same code read the same.
+ *    Display this; never decide on it when [code] is present.
+ *  - [code] is the ORDERING KEY, and is the same value Android itself
+ *    installs by. It answers "may this be installed over that?", which the
+ *    name cannot. Null when the server predates the field.
+ *  - [channel] is a SIBLING FIELD, never a suffix inside the name. Reported
+ *    verbatim rather than validated, so an unexpected value is shown rather
+ *    than dropped.
+ *
+ * [apkUrl] is server-relative (e.g. `/api/client/apk`).
 */
 data class UpdateInfo(
    val version: String,
+    val code: Long?,
+    val channel: String?,
    val apkUrl: String,
    val sizeBytes: Long,
 )
@@ -4,12 +4,26 @@ import kotlinx.serialization.SerialName
 import kotlinx.serialization.Serializable

 /**
- * Wire shape for `GET /api/client/version`. Defaults match Flutter:
- * apk_url falls back to `/api/client/apk` if the server omits it.
+ * Wire shape for `GET /api/client/version`.
+ *
+ * `apkUrl` falls back to `/api/client/apk` if the server omits it.
+ *
+ * [code] MUST stay nullable, and this is not a style preference. The app's
+ * Json is configured with `coerceInputValues = true`, which replaces a JSON
+ * null with the declared default for a NON-nullable property — so writing
+ * `val code: Long = 0` would turn "this server reports no ordering key" into
+ * "this build's ordering key is 0", silently, with no error anywhere. A
+ * nullable type is what keeps absent distinguishable from zero, and the
+ * distinction is the whole reason the field exists.
+ *
+ * A server predating the ordering key sends neither [code] nor [channel];
+ * both arrive null and the caller falls back to comparing names.
 */
@Serializable
 data class UpdateInfoWire(
    val version: String = "",
+    val code: Long? = null,
+    val channel: String? = null,
    @SerialName("apk_url") val apkUrl: String = "/api/client/apk",
    @SerialName("size_bytes") val sizeBytes: Long = 0,
 )
@@ -109,8 +109,11 @@ private fun MiniCover(coverUrl: String, contentDescription: String) {
 * NowPlayingScreen via [onExpandClick].
 *
 * Layout (Column):
- *   - Slim seek slider at the top (4dp track)
- *   - Row: cover | title/artist column | like | prev | play/pause | next
+ *   - Slim seek slider pinned at the top (4dp track)
+ *   - Row: cover | title/artist column | like | prev | play/pause | next.
+ *     Weighted so it fills the rest of the fixed-height bar and centres its
+ *     own content; otherwise the row keeps its intrinsic 48dp and the
+ *     leftover height collects at the bottom as dead surface.
 *
 * No kebab on the mini bar (operator 2026-06-01): the full kebab
 * surface lives on NowPlayingScreen, and dropping it from the mini
@@ -164,6 +167,12 @@ fun MiniPlayer(
                durationMs = state.durationMs,
            )
            MiniRow(
+                // Take whatever the progress fill leaves. Without this the
+                // Column stacks 4dp + the row's intrinsic 48dp from the top
+                // and the remaining 28dp of an 80dp bar sits empty
+                // underneath — the content looked top-aligned rather than
+                // centred, with a dead strip above the gesture bar.
+                modifier = Modifier.weight(1f),
                track = track,
                isPlaying = state.isPlaying,
                isUpnpLoading = state.isUpnpLoading,
@@ -205,6 +214,7 @@ private fun MiniProgressFill(positionMs: Long, durationMs: Long) {
@Composable
@Suppress("LongParameterList")
 private fun MiniRow(
+    modifier: Modifier,
    track: TrackRef,
    isPlaying: Boolean,
    isUpnpLoading: Boolean,
@@ -216,7 +226,7 @@ private fun MiniRow(
    onToggleLike: () -> Unit,
 ) {
    Row(
-        modifier = Modifier
+        modifier = modifier
            .fillMaxWidth()
            .padding(horizontal = 12.dp),
        verticalAlignment = Alignment.CenterVertically,
@@ -9,7 +9,7 @@ import com.fabledsword.minstrel.update.data.ApkInstaller
 import com.fabledsword.minstrel.update.data.InstallStage
 import com.fabledsword.minstrel.update.data.UpdateRepository
 import com.fabledsword.minstrel.update.data.isBusy
-import com.fabledsword.minstrel.update.data.isVersionNewer
+import com.fabledsword.minstrel.update.data.isUpdateAvailable
 import com.fabledsword.minstrel.update.data.message
 import com.fabledsword.minstrel.update.data.stage
 import dagger.hilt.android.lifecycle.HiltViewModel
@@ -37,6 +37,10 @@ sealed interface UpdateCheckResult {

 data class AboutUiState(
    val installedVersion: String = BuildConfig.VERSION_NAME,
+    // The value the platform installs by, and therefore the one the update
+    // check must decide on. Held in state rather than read inline so a test
+    // can drive the comparison without a BuildConfig.
+    val installedCode: Long = BuildConfig.VERSION_CODE.toLong(),
    val isChecking: Boolean = false,
    val installStage: InstallStage = InstallStage.IDLE,
    val installMessage: String? = null,
@@ -45,8 +49,9 @@ data class AboutUiState(

 /**
 * Backs the About card's update controls. "Check for updates" calls
- * [UpdateRepository.getLatest], compares versus the build's
- * VERSION_NAME via [isVersionNewer], and reports the terminal state.
+ * [UpdateRepository.getLatest], compares versus this build via
+ * [isUpdateAvailable] — on the ordering key where the server reports one,
+ * on the name otherwise — and reports the terminal state.
 * When an update is available, [install] downloads the APK via
 * [ApkInstaller] and installs it — routing the user to the "install
 * unknown apps" settings page first when that permission hasn't been
@@ -66,9 +71,17 @@ class AboutCardViewModel @Inject constructor(
        viewModelScope.launch {
            internal.update { it.copy(isChecking = true, installMessage = null) }
            val installed = internal.value.installedVersion
+            val installedCode = internal.value.installedCode
            val result = runCatching { repository.getLatest() }
                .map { latest ->
-                    if (isVersionNewer(latest.version, installed)) {
+                    if (
+                        isUpdateAvailable(
+                            serverCode = latest.code,
+                            serverName = latest.version,
+                            installedCode = installedCode,
+                            installedName = installed,
+                        )
+                    ) {
                        UpdateCheckResult.UpdateAvailable(latest)
                    } else {
                        UpdateCheckResult.Latest
@@ -2,72 +2,45 @@ package com.fabledsword.minstrel.theme

 import androidx.compose.material3.Typography
 import androidx.compose.ui.text.TextStyle
+import androidx.compose.ui.text.font.Font
 import androidx.compose.ui.text.font.FontFamily
 import androidx.compose.ui.text.font.FontStyle
 import androidx.compose.ui.text.font.FontWeight
-import androidx.compose.ui.text.googlefonts.Font
-import androidx.compose.ui.text.googlefonts.GoogleFont
 import androidx.compose.ui.unit.sp
 import com.fabledsword.minstrel.R

 /**
- * Google Fonts provider — fetches font files via Play Services Fonts at
- * runtime, caches them across launches. Matches the Flutter client's
- * `google_fonts` package behaviour (no bundled .ttf files in either tree).
+ * Bundled typefaces, vendored into res/font by tools/vendor-fonts.py.
+ *
+ * These were fetched at runtime through the Play Services font provider until
+ * 2026-09-09. That is a network dependency for rendering, and a deployed
+ * instance is not guaranteed one — the provider is also absent entirely on
+ * devices without Play Services, where the app silently fell back to the
+ * platform default and stopped looking like Minstrel. Bundling costs ~0.86 MB
+ * of APK and removes both failure modes.
 *
 * Per FabledSword design system:
 *   - Fraunces — display + headline (mythic serif)
 *   - Inter — body + label (clean sans for UI text)
 *   - JetBrains Mono — technical / monospace
 * Weights are restricted to 400 (regular) and 500 (medium) only.
+ *
+ * Each res/font entry is a single static instance, not a variable font: the
+ * weight declared beside it here must match the file's own OS/2
+ * usWeightClass, which the vendoring script asserts on download.
 */
-private val GoogleFontProvider = GoogleFont.Provider(
-    providerAuthority = "com.google.android.gms.fonts",
-    providerPackage = "com.google.android.gms",
-    certificates = R.array.com_google_android_gms_fonts_certs,
-)
-
-private val FrauncesFont = GoogleFont("Fraunces")
-private val InterFont = GoogleFont("Inter")
-private val JetBrainsMonoFont = GoogleFont("JetBrains Mono")
-
 private val Fraunces = FontFamily(
-    Font(
-        googleFont = FrauncesFont,
-        fontProvider = GoogleFontProvider,
-        weight = FontWeight.W400,
-        style = FontStyle.Normal,
-    ),
-    Font(
-        googleFont = FrauncesFont,
-        fontProvider = GoogleFontProvider,
-        weight = FontWeight.W500,
-        style = FontStyle.Normal,
-    ),
+    Font(R.font.fraunces_regular, FontWeight.W400, FontStyle.Normal),
+    Font(R.font.fraunces_medium, FontWeight.W500, FontStyle.Normal),
 )

 private val Inter = FontFamily(
-    Font(
-        googleFont = InterFont,
-        fontProvider = GoogleFontProvider,
-        weight = FontWeight.W400,
-        style = FontStyle.Normal,
-    ),
-    Font(
-        googleFont = InterFont,
-        fontProvider = GoogleFontProvider,
-        weight = FontWeight.W500,
-        style = FontStyle.Normal,
-    ),
+    Font(R.font.inter_regular, FontWeight.W400, FontStyle.Normal),
+    Font(R.font.inter_medium, FontWeight.W500, FontStyle.Normal),
 )

 private val JetBrainsMono = FontFamily(
-    Font(
-        googleFont = JetBrainsMonoFont,
-        fontProvider = GoogleFontProvider,
-        weight = FontWeight.W400,
-        style = FontStyle.Normal,
-    ),
+    Font(R.font.jetbrains_mono_regular, FontWeight.W400, FontStyle.Normal),
 )

 /**
@@ -19,7 +19,8 @@ private const val POLL_INTERVAL_MS = 24 * 60 * 60 * 1000L
 /**
 * Drives the shell's soft "update available" banner. Polls
 * `/api/client/version` at launch + every 24h and, when the bundled
- * APK is strictly newer than this build, exposes its [UpdateInfo] so
+ * APK outranks this build — by ordering key where the server reports one,
+ * by name otherwise — exposes its [UpdateInfo] so
 * [com.fabledsword.minstrel.update.ui.UpdateBanner] can nudge an
 * install. Mirrors Flutter's `ClientUpdateController`.
 *
@@ -58,6 +59,13 @@ class UpdateBannerController @Inject constructor(

    private suspend fun runOnce() {
        val info = runCatching { repository.getLatest() }.getOrNull() ?: return
-        latest.value = info.takeIf { isVersionNewer(it.version, BuildConfig.VERSION_NAME) }
+        latest.value = info.takeIf {
+            isUpdateAvailable(
+                serverCode = it.code,
+                serverName = it.version,
+                installedCode = BuildConfig.VERSION_CODE.toLong(),
+                installedName = BuildConfig.VERSION_NAME,
+            )
+        }
    }
 }
@@ -21,10 +21,39 @@ class UpdateRepository @Inject constructor(retrofit: Retrofit) {

 private fun UpdateInfoWire.toDomain(): UpdateInfo = UpdateInfo(
    version = version,
+    code = code,
+    channel = channel,
    apkUrl = apkUrl,
    sizeBytes = sizeBytes,
 )

+/**
+ * True when [server] should be offered over the installed build.
+ *
+ * **Decide on the ordering key whenever the server sends one.** That is the
+ * same value Android's package installer compares, so an offer made this way
+ * implies an install the platform will actually accept. The app used to
+ * compare NAMES while the platform installed by `versionCode`, with nothing
+ * keeping the two orderings consistent — so it could offer a build Android
+ * then refused as a downgrade, or stay quiet about one it would have taken.
+ *
+ * Name comparison survives only as the fallback for a server that predates
+ * the field. A null code means "this server cannot tell me" — never "zero" —
+ * because treating absent as zero would rank every such server as infinitely
+ * old and offer its build to everyone, forever.
+ */
+fun isUpdateAvailable(
+    serverCode: Long?,
+    serverName: String,
+    installedCode: Long,
+    installedName: String,
+): Boolean =
+    if (serverCode != null) {
+        serverCode > installedCode
+    } else {
+        isVersionNewer(serverName, installedName)
+    }
+
 /**
 * True when [server] is strictly newer than [installed]. Mirrors
 * Flutter's `isVersionNewer` — splits both strings on `.`, parses
@@ -1,16 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<!-- Google Fonts provider certificate hashes for downloadable fonts via
-     androidx.compose.ui.text.googlefonts.GoogleFont.Provider. Standard
-     values published by Google; copied verbatim from the AndroidX docs. -->
-<resources>
-    <array name="com_google_android_gms_fonts_certs">
-        <item>@array/com_google_android_gms_fonts_certs_dev</item>
-        <item>@array/com_google_android_gms_fonts_certs_prod</item>
-    </array>
-    <string-array name="com_google_android_gms_fonts_certs_dev">
-        <item>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</item>
-    </string-array>
-    <string-array name="com_google_android_gms_fonts_certs_prod">
-        <item>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</item>
-    </string-array>
-</resources>
@@ -0,0 +1,60 @@
+package com.fabledsword.minstrel.api
+
+import okhttp3.MediaType.Companion.toMediaType
+import okhttp3.ResponseBody.Companion.toResponseBody
+import org.junit.jupiter.api.Assertions.assertEquals
+import org.junit.jupiter.api.Test
+import retrofit2.HttpException
+import retrofit2.Response
+import java.io.IOException
+
+class ErrorCopyTest {
+    private fun httpError(status: Int, body: String): HttpException =
+        HttpException(
+            Response.error<Unit>(status, body.toResponseBody("application/json".toMediaType())),
+        )
+
+    @Test
+    fun libraryNotWritableAppendsTheServerDetail() {
+        val detail = "Minstrel runs as uid 1000, gid 1000 and cannot delete from /music/A " +
+            "(read-only file system). The library mount must be writable by that user. " +
+            "Nothing was deleted."
+        val e = httpError(409, """{"error":{"code":"library_not_writable","message":"$detail"}}""")
+
+        assertEquals(
+            "${ErrorCopy.messageFor("library_not_writable")} $detail",
+            ErrorCopy.fromThrowable(e),
+        )
+    }
+
+    @Test
+    fun detailCodeWithoutAMessageShowsTheCopyAlone() {
+        val e = httpError(409, """{"error":{"code":"library_not_writable","message":""}}""")
+
+        assertEquals(ErrorCopy.messageFor("library_not_writable"), ErrorCopy.fromThrowable(e))
+    }
+
+    // Server messages are usually internal detail; appending them for every
+    // code would leak driver errors into snackbars. This pins the scope.
+    @Test
+    fun otherCodesNeverCarryTheServerMessage() {
+        val e = httpError(404, """{"error":{"code":"track_not_found","message":"pgx: no rows"}}""")
+
+        assertEquals(ErrorCopy.messageFor("track_not_found"), ErrorCopy.fromThrowable(e))
+    }
+
+    @Test
+    fun anUnparseableBodyFallsBackToUnknown() {
+        val e = httpError(500, "not json")
+
+        assertEquals(ErrorCopy.messageFor("unknown"), ErrorCopy.fromThrowable(e))
+    }
+
+    @Test
+    fun transportFailureMapsToConnectionRefused() {
+        assertEquals(
+            ErrorCopy.messageFor("connection_refused"),
+            ErrorCopy.fromThrowable(IOException("refused")),
+        )
+    }
+}
@@ -0,0 +1,125 @@
+package com.fabledsword.minstrel.theme
+
+import org.junit.jupiter.api.Test
+import java.io.File
+import kotlin.test.assertEquals
+import kotlin.test.assertTrue
+
+/**
+ * Guards that the typefaces ship inside the APK instead of being fetched at
+ * runtime.
+ *
+ * Until 2026-09-09 these were resolved through the Play Services font
+ * provider. That needs a network the deployed app is not guaranteed, and a
+ * provider that devices without Play Services do not have at all. Both
+ * failures are silent — text just renders in the platform default, which
+ * reads as a styling regression rather than a missing dependency.
+ *
+ * Expectations are read out of Typography.kt itself rather than hardcoded, so
+ * this cannot drift away from what the app actually declares: adding a face
+ * without vendoring its file fails here, and so does changing a declared
+ * weight without refetching the matching static instance.
+ */
+class BundledFontsTest {
+    @Test
+    fun `typography builds its families from bundled resources`() {
+        val source = typographySource()
+        assertTrue(
+            source.contains("R.font."),
+            "Typography.kt should build its families from res/font resources",
+        )
+        FORBIDDEN.forEach { symbol ->
+            assertTrue(
+                !source.contains(symbol),
+                "Typography.kt must not reference $symbol — fonts are bundled, not fetched",
+            )
+        }
+    }
+
+    @Test
+    fun `every declared face is vendored as TrueType at its declared weight`() {
+        val declared = FACE_PATTERN.findAll(typographySource()).toList()
+        assertTrue(
+            declared.isNotEmpty(),
+            "no Font(R.font.…, FontWeight.W…) declarations found — the guard would pass vacuously",
+        )
+
+        declared.forEach { match ->
+            val (name, weight) = match.destructured
+            val file = File(appDir(), "src/main/res/font/$name.ttf")
+            assertTrue(file.isFile, "res/font/$name.ttf is missing — run tools/vendor-fonts.py")
+
+            val bytes = file.readBytes()
+            assertTrue(
+                bytes.copyOfRange(0, TTF_MAGIC.size).contentEquals(TTF_MAGIC),
+                "$name.ttf is not TrueType — res/font cannot load a woff2 or an eot",
+            )
+            // The decisive check. Google's css2 endpoint silently collapses a
+            // multi-weight request to 400 for legacy clients, so Medium can
+            // come back as Regular: a valid TrueType file that renders at the
+            // wrong weight everywhere. usWeightClass is the only field that
+            // tells the two apart.
+            assertEquals(
+                weight.toInt(),
+                weightClass(bytes),
+                "$name.ttf carries a different OS/2 usWeightClass than the FontWeight declared beside it",
+            )
+        }
+    }
+
+    /** Typography.kt with comments removed, so prose naming the forbidden
+     *  symbols cannot satisfy — or trip — the absence check above. */
+    private fun typographySource(): String =
+        File(appDir(), TYPOGRAPHY)
+            .readText()
+            .replace(BLOCK_COMMENT, "")
+            .replace(LINE_COMMENT, "")
+
+    /** Gradle's working directory for tests is the module dir, but don't rely
+     *  on it: walk up until the module is found, and say so if it isn't. */
+    private fun appDir(): File {
+        var dir: File? = File("").absoluteFile
+        while (dir != null) {
+            if (File(dir, TYPOGRAPHY).isFile) return dir
+            if (File(dir, "app/$TYPOGRAPHY").isFile) return File(dir, "app")
+            dir = dir.parentFile
+        }
+        error("could not locate the app module from ${File("").absolutePath}")
+    }
+
+    private fun weightClass(bytes: ByteArray): Int {
+        val tables = readU16(bytes, NUM_TABLES)
+        for (i in 0 until tables) {
+            val record = TABLE_DIRECTORY + i * TABLE_RECORD
+            if (String(bytes, record, TAG_LENGTH, Charsets.US_ASCII) == "OS/2") {
+                return readU16(bytes, readU32(bytes, record + OFFSET_FIELD) + WEIGHT_FIELD)
+            }
+        }
+        error("no OS/2 table in the font")
+    }
+
+    private fun readU16(bytes: ByteArray, at: Int): Int =
+        ((bytes[at].toInt() and BYTE_MASK) shl Byte.SIZE_BITS) or (bytes[at + 1].toInt() and BYTE_MASK)
+
+    private fun readU32(bytes: ByteArray, at: Int): Int =
+        (readU16(bytes, at) shl Short.SIZE_BITS) or readU16(bytes, at + 2)
+
+    private companion object {
+        const val TYPOGRAPHY = "src/main/java/com/fabledsword/minstrel/theme/Typography.kt"
+
+        val FORBIDDEN = listOf("GoogleFont", "googlefonts")
+        val FACE_PATTERN = Regex("""R\.font\.(\w+)\s*,\s*FontWeight\.W(\d+)""")
+        val BLOCK_COMMENT = Regex("""/\*[\s\S]*?\*/""")
+        val LINE_COMMENT = Regex("""//.*""")
+        val TTF_MAGIC = byteArrayOf(0x00, 0x01, 0x00, 0x00)
+
+        // Offsets into the TrueType table directory, per the OpenType spec.
+        const val NUM_TABLES = 4
+        const val TABLE_DIRECTORY = 12
+        const val TABLE_RECORD = 16
+        const val TAG_LENGTH = 4
+        const val OFFSET_FIELD = 8
+        const val WEIGHT_FIELD = 4
+        const val BYTE_MASK = 0xFF
+    }
+}
@@ -0,0 +1,163 @@
+package com.fabledsword.minstrel.update.data
+
+import org.junit.jupiter.api.Test
+import kotlin.test.assertFalse
+import kotlin.test.assertTrue
+
+/**
+ * The update channel had no tests at all before this. That is worth saying
+ * out loud, because the thing it decides — whether anyone is ever offered an
+ * update — fails silently in both directions: an update nobody is offered
+ * looks exactly like being up to date, and nobody files a bug about a prompt
+ * they never saw.
+ */
+class UpdateVersioningTest {
+    @Test
+    fun `decides on the ordering key when the server reports one`() {
+        assertTrue(
+            isUpdateAvailable(
+                serverCode = 3523847, serverName = "2026.09.10.1432",
+                installedCode = 3519456, installedName = "2026.09.09.1828",
+            ),
+        )
+        assertFalse(
+            isUpdateAvailable(
+                serverCode = 3519456, serverName = "2026.09.09.1828",
+                installedCode = 3523847, installedName = "2026.09.10.1432",
+            ),
+        )
+    }
+
+    @Test
+    fun `an equal ordering key is not an update`() {
+        assertFalse(
+            isUpdateAvailable(
+                serverCode = 3523847, serverName = "2026.09.10.1432",
+                installedCode = 3523847, installedName = "2026.09.10.1432",
+            ),
+        )
+    }
+
+    /**
+     * The property the whole rework exists for: the offer must agree with what
+     * the platform will actually install. Where the two disagree, the ordering
+     * key wins, because that is the value Android compares.
+     */
+    @Test
+    fun `the ordering key wins even when the name disagrees`() {
+        // Name looks older, key is newer — e.g. an older commit rebuilt later.
+        assertTrue(
+            isUpdateAvailable(
+                serverCode = 9_000_000, serverName = "2020.01.01.0000",
+                installedCode = 1, installedName = "2099.12.31.2359",
+            ),
+        )
+        // Name looks newer, key is not. Offering this would be offering an
+        // install the platform then refuses as a downgrade.
+        assertFalse(
+            isUpdateAvailable(
+                serverCode = 1, serverName = "2099.12.31.2359",
+                installedCode = 9_000_000, installedName = "2020.01.01.0000",
+            ),
+        )
+    }
+
+    @Test
+    fun `falls back to the name when the server reports no ordering key`() {
+        assertTrue(
+            isUpdateAvailable(
+                serverCode = null, serverName = "2026.09.10.1432",
+                installedCode = 3519456, installedName = "2026.09.09.1828",
+            ),
+        )
+        assertFalse(
+            isUpdateAvailable(
+                serverCode = null, serverName = "2026.09.09.1828",
+                installedCode = 3519456, installedName = "2026.09.10.1432",
+            ),
+        )
+    }
+
+    /**
+     * A null code must never be read as zero. Zero would rank every
+     * older server as infinitely behind and offer its build to everyone,
+     * forever — so this asserts the fallback runs instead of a comparison
+     * against 0 succeeding by accident.
+     */
+    @Test
+    fun `a null ordering key is absent, not zero`() {
+        // installedCode is 0 here: if null coerced to 0, "0 > 0" would be
+        // false and this would wrongly report no update despite a newer name.
+        assertTrue(
+            isUpdateAvailable(
+                serverCode = null, serverName = "2026.09.10.1432",
+                installedCode = 0, installedName = "2026.09.09.1828",
+            ),
+        )
+    }
+
+    /**
+     * The recorded migration constraint, pinned so it cannot be forgotten:
+     * the old scheme's fourth segment was a commit count (~1895), the new
+     * one is HHMM. Across a day boundary the date decides and all is well.
+     */
+    @Test
+    fun `new-scheme name outranks an old-scheme name on a later day`() {
+        assertTrue(isVersionNewer("2026.09.10.1432", "2026.09.09.1895"))
+    }
+
+    /**
+     * ...but on the SAME day the comparison comes down to HHMM against a
+     * commit count, and any build before ~19:00 UTC reads as older. This is
+     * why the first new-scheme release had to be cut on a later calendar day.
+     * Asserting the trap so nobody "fixes" it by accident.
+     */
+    @Test
+    fun `same-day new-scheme name can read older than an old-scheme name`() {
+        assertFalse(isVersionNewer("2026.09.09.1828", "2026.09.09.1895"))
+    }
+
+    @Test
+    fun `name comparison degrades per segment rather than discarding`() {
+        // The string is still compared rather than rejected outright: an
+        // earlier segment decides and the unparseable tail never matters.
+        assertTrue(isVersionNewer("2026.09.10.1432-dev", "2026.09.09.1828"))
+
+        // A shorter name pads with zeros instead of being refused.
+        assertTrue(isVersionNewer("2026.09.10", "2026.09.09.9999"))
+        assertFalse(isVersionNewer("2026.09.10", "2026.09.10.0"))
+    }
+
+    /**
+     * What "costs that segment's precision" actually means, and it is worth
+     * pinning because it is a real edge rather than a nicety: when the
+     * unparseable segment is the DECIDING one, it reads as 0 and loses. So a
+     * `-dev` suffixed build compares as older than an unsuffixed one from the
+     * same minute.
+     *
+     * That is the correct behaviour for a degrading parser — it is bounded
+     * loss rather than a discarded string — but it is exactly why the channel
+     * belongs in its own field and never in the name.
+     */
+    @Test
+    fun `an unparseable deciding segment reads as zero and loses`() {
+        assertFalse(isVersionNewer("2026.09.10.1432-dev", "2026.09.10.1000"))
+    }
+
+    /**
+     * Both sides unparseable (branch-name builds) falls back to string
+     * inequality, so a dev build still surfaces rather than comparing equal
+     * and going silent.
+     */
+    @Test
+    fun `two unparseable names fall back to string inequality`() {
+        assertTrue(isVersionNewer("main", "dev"))
+        assertFalse(isVersionNewer("dev", "dev"))
+    }
+
+    @Test
+    fun `a leading v is ignored on either side`() {
+        assertTrue(isVersionNewer("v2026.09.10.1432", "2026.09.09.1828"))
+        assertFalse(isVersionNewer("v2026.09.10.1432", "v2026.09.10.1432"))
+    }
+}
@@ -53,7 +53,6 @@ compose-ui-graphics = { module = "androidx.compose.ui:ui-graphics" }
 compose-ui-tooling = { module = "androidx.compose.ui:ui-tooling" }
 compose-ui-tooling-preview = { module = "androidx.compose.ui:ui-tooling-preview" }
 compose-material3 = { module = "androidx.compose.material3:material3" }
-compose-ui-text-google-fonts = { module = "androidx.compose.ui:ui-text-google-fonts" }
 hilt-android = { module = "com.google.dagger:hilt-android", version.ref = "hilt" }
 hilt-compiler = { module = "com.google.dagger:hilt-compiler", version.ref = "hilt" }
 room-runtime = { module = "androidx.room:room-runtime", version.ref = "room" }
@@ -62,57 +62,30 @@ None.
 - **Go toolchain pin.** `go.mod` is on `go 1.25.0` because `golang.org/x/crypto v0.51.0` declares 1.25 as its minimum. `ci-go:1.26` satisfies this with headroom. Future `x/crypto` bumps that move the Go floor should be paired with an image-tag bump in this file + the workflows.
 - **In-app update channel — `needs:`, not polling.** `release.yml`'s `image-release` job declares `needs: [android-release]`, so on tag pushes the signed APK is guaranteed present before the image build starts — no polling window, no race. (The old cross-workflow polling against `flutter.yml` is gone with that workflow.) On non-tag `main` pushes `android-release` is skipped and `image-release` instead pulls the most recent release's APK and reconstructs its exact `versionName`, so `:latest` never ships without an update channel. It degrades to an empty `client/` — never a wrong version — if no release, asset, or tag commit-count can be resolved.
 - **Cache server reachability.** `test-web.yml` does NOT use `cache: 'npm'` on `actions/setup-node` — the Gitea Actions cache server isn't reachable from this runner's container network and `setup-node` was burning ~4m41s on ETIMEDOUT before failing open. With the migration to `ci-go:1.26`, `setup-node` is removed entirely (Node is in the image). The cache concern reappears if a future change re-introduces a network-dependent action.
- **Artifacts — use the mirrored actions, never `actions/{upload,download}-artifact`.**
+- **Artifacts — stock `actions/upload-artifact@v7` and `actions/download-artifact@v8`; never `@v3`.**
  ```yaml
-  uses: https://git.fabledsword.com/bvandeusen/upload-artifact@cb8afe72b42edc798abfb8fcb556cf660d894245
-  uses: https://git.fabledsword.com/bvandeusen/download-artifact@8d4e9521a5f7e5f8b6351f341f719f9f45a92a3a
+  uses: actions/upload-artifact@v7
+  uses: actions/download-artifact@v8
  ```
-  Upstream's `@v4+` cannot work against this instance and no server-side change
-  will help: `isGhes()` rejects any hostname that isn't `github.com` /
-  `*.ghe.com` / `*.localhost` and throws before it opens a connection, so the
-  server is never asked what it supports. `@v3` is worse — it reports success,
-  and Gitea then serves artifacts back only through the v4 API
-  (`content_encoding = application/zip`), so a v3 upload is stored but invisible
-  to every retrieval path. A green job producing nothing retrievable; that is how
-  72 unreachable artifacts accumulated on this repo. Scribe issues 2255 / 2270.
+  Stock works on this forge since the runner moved to gitea/runner 3.x, which
+  edits the actions' client-side `isGhes()` refusal out of their bundles. Proven
+  on 2026-09-10 for upload v4–v7 and download v4–v8 (Scribe spike #3843). Until
+  then this repo pinned SHA mirrors of the Forgejo project's forks, because
+  upstream threw on the hostname before it opened a connection (Scribe 2255).

-  Both are pull mirrors of the Forgejo project's forks
-  (`code.forgejo.org/forgejo/{upload,download}-artifact`, one commit on upstream
-  disabling that check), mirrored so CI depends on commits we hold and pinned by
-  SHA because the mirrors auto-sync every 8h — a moved upstream tag would
-  otherwise silently change what runs.
+  `@v3` is still broken: it reports success, and Gitea serves artifacts back only
+  through the v4 API (`content_encoding = application/zip`), so a v3 upload is
+  stored but invisible to every retrieval path. That is how 72 unreachable
+  artifacts accumulated on this repo (Scribe 2270).

-  **Match the pins on `@actions/artifact`, not on the actions' own version
-  numbers.** The two actions release on unrelated cadences, so equal version
-  numbers do NOT mean a compatible pair — upload `v5` bundles `@actions/artifact`
-  ^4.0.0 while download `v5` bundles ^2.3.2. The pins above are upload **v5** and
-  download **v6**, which is the pairing that puts ^4.0.0 on both sides. This
-  matters because `release.yml` is a producer/consumer pair — `android-release`
-  uploads `minstrel-apk`, `image-release` downloads it — and a protocol mismatch
-  across it yields an empty listing rather than an error, exactly the silent
-  failure this entry exists to prevent.
-
-  | tag | `@actions/artifact` | runtime |
-  |---|---|---|
-  | upload v4 | ^2.1.1 | node20 |
-  | **upload v5** ← pinned | **^4.0.0** | node20 |
-  | download v4 | ^2.1.1 | node20 |
-  | download v5 | ^2.3.2 | node20 |
-  | **download v6** ← pinned | **^4.0.0** | node20 |
-  | download v7 | ^5.0.0 | **node24** |
-
-  The only true protocol break in this history was **v3 → v4** (upstream:
-  "Downloading artifacts that were created from `actions/upload-artifact@v3` and
-  below are not supported"); v4-and-up are one family. Later majors are mostly
-  ergonomics and runtime — upload v4 forbids re-uploading a name and caps a job
-  at 500 artifacts; download v5 made by-ID extraction match by-name.
-
-  **Do not jump the download pin to v7.** That major is a runner requirement, not
-  a feature change: it moves to `runs.using: node24` and upstream states it
-  "requires a minimum Actions Runner version of 2.327.1 … if you are using
-  self-hosted runners, ensure they are updated before upgrading." act_runner is
-  not GitHub's runner and makes no such version claim, so node24 is unverified
-  here. Everything currently pinned is node20.
+  **Pairing no longer needs managing.** This entry used to pin upload v5 against
+  download v6 so both bundled `@actions/artifact` ^4.0.0, warning that a mismatch
+  across `release.yml`'s producer/consumer pair would list empty. Tested, and not
+  true on this instance: every download major v4–v8 read the artifacts of every
+  upload major v4–v7, by name and by pattern (CI-runner run 6312). The only real
+  protocol break is v3 → v4. node24 is no longer a concern either — every
+  CI-runner image carries Node 24 and the runner runs actions with the image's
+  `node`.

  Upload steps set `if-no-files-found: error` rather than the default `warn`, so
  an upload that matches nothing fails its own job instead of failing the
@@ -0,0 +1,159 @@
+#!/usr/bin/env bash
+#
+# Derives the three values a build is stamped with, and the tag that names it.
+#
+#   name=YYYY.MM.DD.HHMM   label for people, from the timestamp of the newest
+#                          commit that CHANGED SOMETHING SHIPPED (see SHIPPED)
+#   code=<int>             ordering key, minutes since 2020-01-01 at BUILD time
+#   tag=v<name>            what a release of this commit must be called
+#
+# Usage:  ci/version.sh [<commit-ish>]        (default HEAD)
+#
+# This exists as a script rather than inline workflow YAML for one reason:
+# release.yml only runs on `main` and on tags, so anything living inside it is
+# unverifiable until a release is already happening — which is the worst
+# possible moment to discover the version is wrong, because the failure mode
+# is silent (an update nobody is offered looks exactly like being current).
+# As a script it can be executed by a test on every push instead.
+#
+# The two clocks are deliberate and are NOT interchangeable:
+#
+#   The NAME answers "is this the same code?" — so it must read identically on
+#   every lane that builds this commit. Commit time does that; build time
+#   prints two different strings for one thing.
+#
+#   The CODE answers "may this be installed over that?" — so it must be
+#   monotonic BY CONSTRUCTION. Build time is; commit time is not (rebuild an
+#   older commit and it goes down, which on a phone is a refused install), and
+#   a commit COUNT is worse still, because it runs ahead on `dev` and inverts
+#   against `main`.
+set -euo pipefail
+
+readonly EPOCH_2020=1577836800 # 2020-01-01T00:00:00Z
+readonly REF="${1:-HEAD}"
+
+# Both clocks are overridable so a test can pin them. Nothing but tests should
+# set these — the defaults are the real derivation.
+# The paths that do NOT ship, in either artifact. Everything else counts.
+#
+# A DENYLIST, and the direction is the whole point. As an allowlist, the list
+# has to be updated by whoever adds a directory and nothing fails if they
+# don't — so the failure mode is a changed artifact keeping its old version,
+# silently, on a green run. That is a build lying about what it is. Inverted,
+# new content counts by default and the only way to wrongly EXCLUDE something
+# is to name it here deliberately.
+#
+# The two error directions are not symmetric, which is why this is not taste:
+#   wrongly excluded → changed artifact, unchanged version. A silent lie.
+#   wrongly included → version moves when nothing shipped. Cosmetic noise in
+#                      a string nobody sorts.
+#
+# THIS REPO SHIPS TWO ARTIFACTS FROM ONE DERIVATION, and that is why the list
+# is shorter than it looks like it should be. The server image ships cmd/,
+# internal/, shared/, web/, config.example.yaml and client/; the APK ships
+# android/. Neither ships the other's sources — but excluding android/ here
+# would stop an Android-only commit from moving the APK's OWN version, which
+# is the dangerous direction. So this is the union: exclude only what ships in
+# NEITHER, and accept that an Android commit also nudges the server's reported
+# version. Over-inclusion across the two, which is the harmless direction.
+#
+# The family's other repos (roundtable / roundtable-android) each keep a
+# tighter list because they are separate repos with one artifact apiece. Do
+# not copy theirs onto this one.
+readonly SHIPPED=(
+  .
+  ':!.gitea'          # CI workflows — including this script's own caller
+  ':!ci'              # CI scripts — including this script
+  ':!docs'
+  ':!tools'           # asset/font generators; their OUTPUT ships, they do not
+  ':!deploy'          # test-database bootstrap SQL
+  ':!bin'             # local `make build` output
+  ':!*.md'
+  ':!Makefile'
+  ':!docker-compose.yml'
+  ':!.env.example'
+  ':!.gitignore'
+  ':!.dockerignore'
+  ':!renovate.json'
+  ':!.golangci.yml'
+
+  # TESTS DO NOT SHIP, so they must not re-version an artifact.
+  #
+  # Named as globs rather than a directory because this repo has no tests/
+  # tree to exclude: Go tests sit inline beside the code they cover, and the
+  # web suite sits beside its modules. `go build` drops *_test.go outright and
+  # the Vite build never imports a .test.ts, so neither reaches an artifact.
+  #
+  # A commit touching a test AND its source still moves the version — the
+  # source path matches on its own. Only a test-ONLY commit is inert, which is
+  # the whole intent.
+  #
+  # Patterns match what exists today and nothing speculative: there are no
+  # .spec.* files, no __tests__/ directories and no androidTest/ tree. If any
+  # appear they will re-version until named here, which is the harmless
+  # direction and the reason this list is a denylist.
+  ':!*_test.go'                   # 158 files, inline beside the code
+  ':!*.test.ts'                   # 114 files
+  ':!*.test.js'
+  ':!android/app/src/test'        # JVM unit tests; no androidTest tree exists
+  ':!web/vitest.config.ts'        # test-harness config, not build config
+  ':!web/vitest.setup.ts'
+)
+
+commit_epoch="${MINSTREL_COMMIT_EPOCH:-}"
+if [ -z "${commit_epoch}" ]; then
+  commit_epoch="$(git log --format=%ct -1 "${REF}" -- "${SHIPPED[@]}")"
+  # Loudly, on purpose. A silent fallback here is the landmine this whole
+  # script exists to avoid: a plausible-looking version that is quietly wrong,
+  # on a green run. Realistically this means a shallow clone (no commit in
+  # range touches the shipped set) rather than a repo of pure CI config.
+  if [ -z "${commit_epoch}" ]; then
+    echo "version.sh: no commit under '${REF}' touches the shipped file set — shallow clone? (needs fetch-depth: 0)" >&2
+    exit 1
+  fi
+fi
+now_epoch="${MINSTREL_NOW_EPOCH:-$(date -u +%s)}"
+
+if ! name="$(date -u -d "@${commit_epoch}" +%Y.%m.%d.%H%M 2>/dev/null)"; then
+  echo "version.sh: could not read a commit timestamp from '${commit_epoch}'" >&2
+  exit 1
+fi
+
+if ! [ "${now_epoch}" -eq "${now_epoch}" ] 2>/dev/null; then
+  echo "version.sh: build timestamp '${now_epoch}' is not a number" >&2
+  exit 1
+fi
+code=$(( (now_epoch - EPOCH_2020) / 60 ))
+
+# Assert the shape here, at the source. A malformed name builds, signs and
+# publishes perfectly happily; it only surfaces later as an update channel
+# that has quietly stopped offering anything.
+if [[ ! "${name}" =~ ^[0-9]{4}\.[0-9]{2}\.[0-9]{2}\.[0-9]{4}$ ]]; then
+  echo "version.sh: name '${name}' is not YYYY.MM.DD.HHMM" >&2
+  exit 1
+fi
+
+# A non-positive key means the build clock is set before 2020, and every
+# comparison downstream would be nonsense.
+if [ "${code}" -le 0 ]; then
+  echo "version.sh: ordering key '${code}' is not positive — build clock wrong?" >&2
+  exit 1
+fi
+
+# Android's versionCode is a signed 32-bit int and the platform refuses an APK
+# whose code exceeds it. At ~525k minutes a year this is four thousand years
+# away in normal operation, so the realistic cause is a build machine with a
+# badly wrong clock — which produces a code that is not merely too large but
+# also unreachably high, permanently blocking every real build that follows
+# from ever outranking it. Cheaper to refuse the build than to discover that
+# from a phone that will not update.
+readonly VERSION_CODE_CEILING=2147483647
+if [ "${code}" -gt "${VERSION_CODE_CEILING}" ]; then
+  echo "version.sh: ordering key '${code}' exceeds versionCode's int32 ceiling — build clock wrong?" >&2
+  exit 1
+fi
+
+# KEY=VALUE, which is also exactly $GITHUB_OUTPUT's format.
+echo "name=${name}"
+echo "code=${code}"
+echo "tag=v${name}"
@@ -122,7 +122,15 @@ func run() error {
 	}
 	defer pool.Close()

-	scanner := library.New(pool, logger, cfg.Library.ScanPaths)
+	// Fingerprinting settings (M400 #3913): one instance, shared by the scanner,
+	// the fingerprint backfill, the duplicate sweep and the admin API, so a save
+	// reaches all of them without a restart. A load failure is logged, not fatal:
+	// the service falls back to the shipped defaults.
+	fpSettings, fpErr := library.NewFingerprintSettingsService(ctx, pool)
+	if fpErr != nil {
+		logger.Warn("fingerprint settings: using defaults", "err", fpErr)
+	}
+	scanner := library.New(pool, logger, cfg.Library.ScanPaths, fpSettings)

 	contact := cfg.Library.ContactEmail
 	if contact == "" {
@@ -214,6 +222,17 @@ func run() error {
 	// SQL, no external calls; empty on single-user servers.
 	go coplay.NewWorker(pool, logger.With("component", "coplay")).Run(ctx)

+	// Fingerprint backfill (M400 #3908): fingerprints the tracks the scan never
+	// will — everything imported before fingerprinting existed, and rows derived
+	// by an older method. A worker of its own rather than a scan stage; see
+	// internal/library/fingerprint_backfill.go for why.
+	go library.NewFingerprintBackfillWorker(pool, logger.With("component", "fingerprint_backfill"), fpSettings).Run(ctx)
+
+	// Duplicate sweep (M400 #3910): proposes groups of tracks holding one
+	// recording, from the fingerprints above. Sweeps only when fingerprints have
+	// changed since the last sweep.
+	go library.NewDuplicateSweepWorker(pool, logger.With("component", "duplicate_sweep"), fpSettings).Run(ctx)
+
 	// Start the tag-enrichment worker (#1490). Reconciles the compiled-in
 	// tag providers with tag_provider_settings, bumps the sources version if
 	// the provider set changed (re-opening settled rows), then drains tracks
@@ -357,6 +376,10 @@ func run() error {
 	srv.PlaylistScheduler = playlistScheduler
 	srv.RecSettings = recSettings
 	srv.TagSettings = tagSettings
+	srv.FingerprintSettings = fpSettings
+	// The sweeper above holds this same instance, so a save from the admin
+	// card changes what it does on its next tick (#3936).
+	srv.ReacqSettings = reacqSettings
 	srv.StreamSecret = cfg.StreamSecret
 	httpServer := &http.Server{
 		Addr:              cfg.Server.Address,
@@ -1,10 +1,12 @@
 package api

 import (
+	"context"
 	"net/http"

 	"git.fabledsword.com/bvandeusen/minstrel/internal/apierror"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
 )

 // coverageRollupResp is the wire shape for GET /api/admin/library/coverage.
@@ -36,3 +38,46 @@ func (h *handlers) handleGetLibraryCoverage(w http.ResponseWriter, r *http.Reque
 		PendingNoMbid: row.PendingNoMbid,
 	})
 }
+
+// fingerprintCoverageResp is the wire shape for GET /api/admin/library/fingerprints.
+// fingerprinted + rejected + pending = total. Missing tracks are not counted:
+// there is no file to fingerprint. Enabled travels with the counts because with
+// fingerprinting off (#3913) pending never shrinks, and a gauge that implies
+// progress would be promising work nothing is doing.
+type fingerprintCoverageResp struct {
+	Total         int64 `json:"total"`
+	Fingerprinted int64 `json:"fingerprinted"`
+	Rejected      int64 `json:"rejected"`
+	Pending       int64 `json:"pending"`
+	Enabled       bool  `json:"enabled"`
+}
+
+// fingerprintCoverage reads the gauge against the current settings: a print at
+// another length counts as pending, because the backfill will re-derive it.
+func (h *handlers) fingerprintCoverage(ctx context.Context) (fingerprintCoverageResp, error) {
+	cfg := h.fingerprintSettings.Get()
+	row, err := library.FingerprintCoverage(ctx, h.pool, cfg)
+	if err != nil {
+		return fingerprintCoverageResp{}, err
+	}
+	return fingerprintCoverageResp{
+		Total:         row.Total,
+		Fingerprinted: row.Fingerprinted,
+		Rejected:      row.Rejected,
+		Pending:       row.Pending,
+		Enabled:       cfg.Enabled,
+	}, nil
+}
+
+// handleGetFingerprintCoverage implements GET /api/admin/library/fingerprints:
+// how far the fingerprint backfill (#3908) has got. The backfill is its own
+// worker spanning many passes, with no scan run to attach a tally to, so its
+// progress is read live here. Always 200; zeros on an empty library.
+func (h *handlers) handleGetFingerprintCoverage(w http.ResponseWriter, r *http.Request) {
+	cov, err := h.fingerprintCoverage(r.Context())
+	if err != nil {
+		writeErrWithLog(w, h.logger, "admin: get fingerprint coverage", apierror.InternalMsg("lookup failed", err))
+		return
+	}
+	writeJSON(w, http.StatusOK, cov)
+}
@@ -0,0 +1,316 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+
+	"github.com/go-chi/chi/v5"
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
+)
+
+// duplicateMemberView is one copy in a proposed duplicate group. LikeCount and
+// PlayCount span every user: the report is admin-only, and what a copy carries
+// is the fact the operator weighs when choosing which to keep.
+type duplicateMemberView struct {
+	TrackID     string `json:"track_id"`
+	Title       string `json:"title"`
+	ArtistName  string `json:"artist_name"`
+	AlbumID     string `json:"album_id"`
+	AlbumTitle  string `json:"album_title"`
+	FilePath    string `json:"file_path"`
+	FileFormat  string `json:"file_format"`
+	FileSize    int64  `json:"file_size"`
+	DurationSec int32  `json:"duration_sec"`
+	AddedAt     string `json:"added_at"`
+	LikeCount   int64  `json:"like_count"`
+	PlayCount   int64  `json:"play_count"`
+}
+
+// duplicateGroupView is one proposal. SurvivorTrackID and SurvivorReason are
+// the copy the report proposes keeping and the rule that chose it
+// (library.ProposeSurvivor) — a default the merge (#3911) lets the operator
+// override.
+type duplicateGroupView struct {
+	ID                string                `json:"id"`
+	Tier              string                `json:"tier"`
+	WorstBitErrorRate *float32              `json:"worst_bit_error_rate"`
+	DetectedAt        string                `json:"detected_at"`
+	SurvivorTrackID   string                `json:"survivor_track_id"`
+	SurvivorReason    string                `json:"survivor_reason"`
+	Members           []duplicateMemberView `json:"members"`
+}
+
+// duplicateSweepView is the latest sweep. State is "never" when none has run,
+// which is what lets the page tell an empty report apart from a sweep that
+// found nothing.
+type duplicateSweepView struct {
+	State            string  `json:"state"`
+	StartedAt        *string `json:"started_at"`
+	FinishedAt       *string `json:"finished_at"`
+	Candidates       *int32  `json:"candidates"`
+	GroupsFound      *int32  `json:"groups_found"`
+	OversizeClusters *int32  `json:"oversize_clusters"`
+	ErrorMessage     *string `json:"error_message"`
+}
+
+// adminDuplicatesResponse is the paged report. Total counts groups.
+type adminDuplicatesResponse struct {
+	Sweep        duplicateSweepView      `json:"sweep"`
+	Fingerprints fingerprintCoverageResp `json:"fingerprints"`
+	Total        int64                   `json:"total"`
+	Limit        int                     `json:"limit"`
+	Offset       int                     `json:"offset"`
+	Groups       []duplicateGroupView    `json:"groups"`
+}
+
+// handleListDuplicates implements GET /api/admin/library/duplicates (#3912).
+//
+// Read-only. The sweep's state and the fingerprint backfill's progress travel
+// with the groups because an empty report means three different things — still
+// fingerprinting, never swept, or swept and clean — and the page has to say which.
+func (h *handlers) handleListDuplicates(w http.ResponseWriter, r *http.Request) {
+	limit, offset, err := parsePaging(r.URL.Query())
+	if err != nil {
+		writeAdminJSONErr(w, http.StatusBadRequest, "invalid_paging")
+		return
+	}
+	ctx := r.Context()
+	q := dbq.New(h.pool)
+
+	sweep := duplicateSweepView{State: "never"}
+	last, err := q.GetLatestDuplicateSweep(ctx)
+	switch {
+	case err == nil:
+		sweep = duplicateSweepViewOf(last)
+	case !errors.Is(err, pgx.ErrNoRows):
+		h.logger.Error("admin: latest duplicate sweep", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+
+	cov, err := h.fingerprintCoverage(ctx)
+	if err != nil {
+		h.logger.Error("admin: fingerprint coverage", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+	total, err := q.CountPendingDuplicateGroups(ctx)
+	if err != nil {
+		h.logger.Error("admin: count duplicate groups", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+	rows, err := q.ListPendingDuplicateGroupMembers(ctx, dbq.ListPendingDuplicateGroupMembersParams{
+		PageLimit: int32(limit), PageOffset: int32(offset),
+	})
+	if err != nil {
+		h.logger.Error("admin: list duplicate groups", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+
+	writeJSON(w, http.StatusOK, adminDuplicatesResponse{
+		Sweep:        sweep,
+		Fingerprints: cov,
+		Total:        total,
+		Limit:        limit,
+		Offset:       offset,
+		Groups:       foldDuplicateGroups(rows),
+	})
+}
+
+func duplicateSweepViewOf(s dbq.DuplicateSweep) duplicateSweepView {
+	v := duplicateSweepView{
+		State:            "running",
+		Candidates:       s.Candidates,
+		GroupsFound:      s.GroupsFound,
+		OversizeClusters: s.OversizeClusters,
+		ErrorMessage:     s.ErrorMessage,
+	}
+	started := formatTimestamp(s.StartedAt)
+	v.StartedAt = &started
+	if s.FinishedAt.Valid {
+		finished := formatTimestamp(s.FinishedAt)
+		v.FinishedAt = &finished
+		v.State = "finished"
+	}
+	return v
+}
+
+// foldDuplicateGroups folds the one-row-per-member query result into groups and
+// proposes each group's survivor. It relies on the query ordering members of a
+// group together, so a run-length fold is enough and the page order holds.
+func foldDuplicateGroups(rows []dbq.ListPendingDuplicateGroupMembersRow) []duplicateGroupView {
+	groups := make([]duplicateGroupView, 0, 8)
+	var candidates [][]library.SurvivorCandidate
+	for _, row := range rows {
+		id := uuidToString(row.GroupID)
+		if n := len(groups); n == 0 || groups[n-1].ID != id {
+			groups = append(groups, duplicateGroupView{
+				ID:                id,
+				Tier:              row.Tier,
+				WorstBitErrorRate: row.WorstBitErrorRate,
+				DetectedAt:        formatTimestamp(row.DetectedAt),
+			})
+			candidates = append(candidates, nil)
+		}
+		n := len(groups) - 1
+		trackID := uuidToString(row.TrackID)
+		groups[n].Members = append(groups[n].Members, duplicateMemberView{
+			TrackID:     trackID,
+			Title:       row.Title,
+			ArtistName:  row.ArtistName,
+			AlbumID:     uuidToString(row.AlbumID),
+			AlbumTitle:  row.AlbumTitle,
+			FilePath:    row.FilePath,
+			FileFormat:  row.FileFormat,
+			FileSize:    row.FileSize,
+			DurationSec: row.DurationMs / 1000,
+			AddedAt:     formatTimestamp(row.AddedAt),
+			LikeCount:   row.LikeCount,
+			PlayCount:   row.PlayCount,
+		})
+		candidates[n] = append(candidates[n], library.SurvivorCandidate{
+			TrackID: trackID, FileFormat: row.FileFormat, FileSize: row.FileSize, AddedAt: row.AddedAt.Time,
+		})
+	}
+	for i := range groups {
+		groups[i].SurvivorTrackID, groups[i].SurvivorReason = library.ProposeSurvivor(candidates[i])
+	}
+	return groups
+}
+
+// handleRunDuplicateSweep implements POST /api/admin/library/duplicates/sweep:
+// 202 when a sweep starts, 409 sweep_in_progress when one is already running.
+// The sweep outlives the request, so it runs on a background context, as
+// handleTriggerScan's scan does.
+func (h *handlers) handleRunDuplicateSweep(w http.ResponseWriter, _ *http.Request) {
+	// Runs whatever the sweep interval says: the interval paces the automatic
+	// sweep, and an operator pressing the button has already decided.
+	started, err := library.TryStartDuplicateSweep(
+		context.Background(), h.pool, h.logger.With("source", "manual"), h.fingerprintSettings.Get(),
+	)
+	if err != nil {
+		h.logger.Error("admin: start duplicate sweep", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+	if !started {
+		writeAdminJSONErr(w, http.StatusConflict, "sweep_in_progress")
+		return
+	}
+	writeJSON(w, http.StatusAccepted, map[string]bool{"started": true})
+}
+
+// handleDismissDuplicateGroup implements POST
+// /api/admin/library/duplicates/{id}/dismiss: "these are not duplicates". The
+// sweep keeps the dismissal and will not propose that set of tracks again. 404
+// duplicate_group_not_pending when the group was already resolved or is gone.
+func (h *handlers) handleDismissDuplicateGroup(w http.ResponseWriter, r *http.Request) {
+	id, ok := parseUUID(chi.URLParam(r, "id"))
+	if !ok {
+		writeAdminJSONErr(w, http.StatusBadRequest, "invalid_id")
+		return
+	}
+	n, err := dbq.New(h.pool).DismissDuplicateGroup(r.Context(), id)
+	if err != nil {
+		h.logger.Error("admin: dismiss duplicate group", "err", err)
+		writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		return
+	}
+	if n == 0 {
+		writeAdminJSONErr(w, http.StatusNotFound, "duplicate_group_not_pending")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]string{"status": "dismissed"})
+}
+
+// mergeDuplicateRequest chooses the copy to keep. An empty survivor_track_id
+// keeps the report's proposal.
+type mergeDuplicateRequest struct {
+	SurvivorTrackID string `json:"survivor_track_id"`
+	Unmonitor       bool   `json:"unmonitor"`
+}
+
+// mergeDuplicateResponse reports what the merge removed. RemovedPaths are files
+// deleted from disk; the operator reads them to know exactly what went.
+type mergeDuplicateResponse struct {
+	SurvivorTrackID       string   `json:"survivor_track_id"`
+	RemovedPaths          []string `json:"removed_paths"`
+	LidarrUnmonitorFailed *bool    `json:"lidarr_unmonitor_failed,omitempty"`
+}
+
+// mergeRequestBodyLimit bounds the request body. It holds one id and a flag.
+const mergeRequestBodyLimit = 1 << 16
+
+// handleMergeDuplicateGroup implements POST /api/admin/library/duplicates/{id}/merge
+// (#3911): keep one copy, move the others' likes, plays and playlist entries onto
+// it, and delete the others' files and rows.
+//
+// Errors:
+//   - 409 library_not_writable / 500 file_delete_failed when a file could not be
+//     removed — nothing was changed (fileRemoveAPIError)
+//   - 404 duplicate_group_not_pending when the group was already resolved
+//   - 400 survivor_not_in_group, invalid_id, invalid_body
+func (h *handlers) handleMergeDuplicateGroup(w http.ResponseWriter, r *http.Request) {
+	admin, ok := requireUser(w, r)
+	if !ok {
+		return
+	}
+	groupID, ok := parseUUID(chi.URLParam(r, "id"))
+	if !ok {
+		writeAdminJSONErr(w, http.StatusBadRequest, "invalid_id")
+		return
+	}
+	var body mergeDuplicateRequest
+	if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, mergeRequestBodyLimit)).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
+		writeAdminJSONErr(w, http.StatusBadRequest, "invalid_body")
+		return
+	}
+	var survivorID pgtype.UUID // invalid: keep the proposal
+	if body.SurvivorTrackID != "" {
+		if survivorID, ok = parseUUID(body.SurvivorTrackID); !ok {
+			writeAdminJSONErr(w, http.StatusBadRequest, "invalid_id")
+			return
+		}
+	}
+
+	res, unmonitorFailed, err := h.tracks.MergeDuplicates(r.Context(), groupID, survivorID, admin.ID, body.Unmonitor)
+	if err != nil {
+		if apiErr, ok := fileRemoveAPIError(err); ok {
+			logFileRemoveFailure(h.logger, apiErr, "group_id", uuidToString(groupID))
+			writeErr(w, apiErr)
+			return
+		}
+		switch {
+		case errors.Is(err, library.ErrDuplicateGroupNotPending):
+			writeAdminJSONErr(w, http.StatusNotFound, "duplicate_group_not_pending")
+		case errors.Is(err, library.ErrSurvivorNotInGroup):
+			writeAdminJSONErr(w, http.StatusBadRequest, "survivor_not_in_group")
+		default:
+			h.logger.Error("admin: merge duplicate group", "group_id", uuidToString(groupID), "err", err)
+			writeAdminJSONErr(w, http.StatusInternalServerError, "server_error")
+		}
+		return
+	}
+
+	resp := mergeDuplicateResponse{
+		SurvivorTrackID: uuidToString(res.Survivor.TrackID),
+		RemovedPaths:    make([]string, 0, len(res.Removed)),
+	}
+	for _, c := range res.Removed {
+		resp.RemovedPaths = append(resp.RemovedPaths, c.FilePath)
+	}
+	if body.Unmonitor && unmonitorFailed {
+		failed := true
+		resp.LidarrUnmonitorFailed = &failed
+	}
+	writeJSON(w, http.StatusOK, resp)
+}
@@ -0,0 +1,65 @@
+package api
+
+import (
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+)
+
+func dupUUID(b byte) pgtype.UUID {
+	var u pgtype.UUID
+	u.Bytes[15] = b
+	u.Valid = true
+	return u
+}
+
+func dupTS(t time.Time) pgtype.Timestamptz { return pgtype.Timestamptz{Time: t, Valid: true} }
+
+// Rows arrive one per member, members of a group together. The fold must keep
+// groups apart, keep the query's order, and propose each group's survivor from
+// its own members only.
+func TestFoldDuplicateGroups(t *testing.T) {
+	older := time.Date(2025, 1, 1, 0, 0, 0, 0, time.UTC)
+	newer := older.Add(48 * time.Hour)
+	ber := float32(0.04)
+	rows := []dbq.ListPendingDuplicateGroupMembersRow{
+		// Group 1: identical audio, sizes tie, the older copy should be kept.
+		{GroupID: dupUUID(1), Tier: "exact", DetectedAt: dupTS(newer), TrackID: dupUUID(10),
+			Title: "WWW", FileFormat: "mp3", FileSize: 6_900_000, DurationMs: 215_400, AddedAt: dupTS(newer), PlayCount: 3},
+		{GroupID: dupUUID(1), Tier: "exact", DetectedAt: dupTS(newer), TrackID: dupUUID(11),
+			Title: "WWW", FileFormat: "mp3", FileSize: 6_900_000, DurationMs: 215_400, AddedAt: dupTS(older), LikeCount: 1},
+		// Group 2: the same recording, FLAC against MP3.
+		{GroupID: dupUUID(2), Tier: "acoustic", WorstBitErrorRate: &ber, DetectedAt: dupTS(older), TrackID: dupUUID(20),
+			Title: "Lovesick", FileFormat: "mp3", FileSize: 9_000_000, DurationMs: 198_000, AddedAt: dupTS(older)},
+		{GroupID: dupUUID(2), Tier: "acoustic", WorstBitErrorRate: &ber, DetectedAt: dupTS(older), TrackID: dupUUID(21),
+			Title: "Lovesick", FileFormat: "flac", FileSize: 30_000_000, DurationMs: 198_000, AddedAt: dupTS(newer)},
+	}
+
+	got := foldDuplicateGroups(rows)
+	if len(got) != 2 {
+		t.Fatalf("folded %d groups, want 2", len(got))
+	}
+
+	g1, g2 := got[0], got[1]
+	if g1.ID != uuidToString(dupUUID(1)) || len(g1.Members) != 2 || g1.WorstBitErrorRate != nil {
+		t.Fatalf("group 1 = %+v, want the exact pair with no score", g1)
+	}
+	if g1.SurvivorTrackID != uuidToString(dupUUID(11)) || g1.SurvivorReason != "in the library longest" {
+		t.Errorf("group 1 survivor = (%s, %q), want the older copy", g1.SurvivorTrackID, g1.SurvivorReason)
+	}
+	if g1.Members[0].DurationSec != 215 || g1.Members[0].PlayCount != 3 || g1.Members[1].LikeCount != 1 {
+		t.Errorf("group 1 members lost their facts: %+v", g1.Members)
+	}
+
+	if g2.Tier != "acoustic" || g2.WorstBitErrorRate == nil || *g2.WorstBitErrorRate != ber {
+		t.Fatalf("group 2 = %+v, want the acoustic pair with its score", g2)
+	}
+	// Chosen from group 2's own members: a survivor leaking across groups is
+	// exactly what a wrong fold boundary would produce.
+	if g2.SurvivorTrackID != uuidToString(dupUUID(21)) || g2.SurvivorReason != "lossless (flac)" {
+		t.Errorf("group 2 survivor = (%s, %q), want the FLAC copy", g2.SurvivorTrackID, g2.SurvivorReason)
+	}
+}
@@ -0,0 +1,67 @@
+package api
+
+import (
+	"encoding/json"
+	"errors"
+	"net/http"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/apierror"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
+)
+
+// fingerprintSettingsBody is the wire shape for GET and PUT
+// /api/admin/library/fingerprint-settings (M400 #3913). The threshold travels as
+// the bit-error rate the matcher uses; the card presents it as a match percentage.
+type fingerprintSettingsBody struct {
+	Enabled                 bool    `json:"enabled"`
+	ChromaprintLengthSec    int32   `json:"chromaprint_length_sec"`
+	AcousticMaxBitErrorRate float64 `json:"acoustic_max_bit_error_rate"`
+	BackfillConcurrency     int32   `json:"backfill_concurrency"`
+	SweepIntervalHours      int32   `json:"sweep_interval_hours"`
+}
+
+func fingerprintSettingsBodyOf(s library.FingerprintSettings) fingerprintSettingsBody {
+	return fingerprintSettingsBody{
+		Enabled:                 s.Enabled,
+		ChromaprintLengthSec:    s.ChromaprintLengthSec,
+		AcousticMaxBitErrorRate: s.AcousticMaxBitErrorRate,
+		BackfillConcurrency:     s.BackfillConcurrency,
+		SweepIntervalHours:      s.SweepIntervalHours,
+	}
+}
+
+// handleGetFingerprintSettings implements GET /api/admin/library/fingerprint-settings.
+func (h *handlers) handleGetFingerprintSettings(w http.ResponseWriter, _ *http.Request) {
+	writeJSON(w, http.StatusOK, fingerprintSettingsBodyOf(h.fingerprintSettings.Get()))
+}
+
+// handleUpdateFingerprintSettings implements PUT /api/admin/library/fingerprint-settings.
+//
+// A whole-row write. A body that leaves a field out decodes it as zero, which no
+// field accepts, so a partial save is refused rather than zeroing what it omitted.
+// The saved settings reach the scanner and both workers at once: they share the
+// service instance.
+func (h *handlers) handleUpdateFingerprintSettings(w http.ResponseWriter, r *http.Request) {
+	var req fingerprintSettingsBody
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeErr(w, apierror.BadRequest("invalid_body", "malformed JSON"))
+		return
+	}
+	saved, err := h.fingerprintSettings.Set(r.Context(), library.FingerprintSettings{
+		Enabled:                 req.Enabled,
+		ChromaprintLengthSec:    req.ChromaprintLengthSec,
+		AcousticMaxBitErrorRate: req.AcousticMaxBitErrorRate,
+		BackfillConcurrency:     req.BackfillConcurrency,
+		SweepIntervalHours:      req.SweepIntervalHours,
+	})
+	if err != nil {
+		// Validation mirrors migration 0061's CHECKs and names the field.
+		if errors.Is(err, library.ErrFingerprintSettingOutOfRange) {
+			writeErr(w, apierror.BadRequest("invalid_setting", err.Error()))
+			return
+		}
+		writeErrWithLog(w, h.logger, "admin fingerprint settings: update failed", apierror.Internal(err))
+		return
+	}
+	writeJSON(w, http.StatusOK, fingerprintSettingsBodyOf(saved))
+}
@@ -0,0 +1,55 @@
+package api
+
+import (
+	"encoding/json"
+	"io"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
+)
+
+func TestGetFingerprintSettings_ServesDefaultsWithoutAService(t *testing.T) {
+	h := &handlers{logger: slog.New(slog.NewTextHandler(io.Discard, nil))}
+	rec := httptest.NewRecorder()
+	h.handleGetFingerprintSettings(rec, httptest.NewRequest(http.MethodGet, "/api/admin/library/fingerprint-settings", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200", rec.Code)
+	}
+	var got fingerprintSettingsBody
+	if err := json.NewDecoder(rec.Body).Decode(&got); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if want := fingerprintSettingsBodyOf(library.DefaultFingerprintSettings); got != want {
+		t.Fatalf("body = %+v, want the defaults %+v", got, want)
+	}
+}
+
+func TestUpdateFingerprintSettings_Rejects(t *testing.T) {
+	h := &handlers{logger: slog.New(slog.NewTextHandler(io.Discard, nil))}
+	for name, tc := range map[string]struct {
+		body     string
+		code     string
+		mentions string
+	}{
+		"a value out of range, naming the field": {
+			body:     `{"enabled":true,"chromaprint_length_sec":5,"acoustic_max_bit_error_rate":0.15,"backfill_concurrency":2,"sweep_interval_hours":1}`,
+			code:     "invalid_setting",
+			mentions: "chromaprint_length_sec",
+		},
+		// A partial body would otherwise zero every field it left out.
+		"a body missing fields": {body: `{"enabled":false}`, code: "invalid_setting"},
+		"malformed JSON":        {body: `{"enabled":`, code: "invalid_body"},
+	} {
+		rec := httptest.NewRecorder()
+		h.handleUpdateFingerprintSettings(rec, httptest.NewRequest(
+			http.MethodPut, "/api/admin/library/fingerprint-settings", strings.NewReader(tc.body)))
+		body := rec.Body.String()
+		if rec.Code != http.StatusBadRequest || !strings.Contains(body, `"`+tc.code+`"`) || !strings.Contains(body, tc.mentions) {
+			t.Errorf("%s: status %d body %s; want 400 %s mentioning %q", name, rec.Code, body, tc.code, tc.mentions)
+		}
+	}
+}
@@ -133,6 +133,13 @@ func (h *handlers) handleDeleteQuarantineFile(w http.ResponseWriter, r *http.Req
 	}
 	action, err := h.lidarrQuarantine.DeleteFile(r.Context(), id, admin.ID)
 	if err != nil {
+		// Written in the enveloped shape, not writeAdminJSONErr's bare code: the
+		// message is the part that tells the operator which directory and uid.
+		if apiErr, ok := fileRemoveAPIError(err); ok {
+			logFileRemoveFailure(h.logger, apiErr, "track_id", uuidToString(id))
+			writeErr(w, apiErr)
+			return
+		}
 		switch {
 		case errors.Is(err, lidarrquarantine.ErrTrackNotFound):
 			writeAdminJSONErr(w, http.StatusNotFound, "track_not_found")
@@ -69,7 +69,7 @@ func installQuarantineClientFn(t *testing.T, h *handlers) {
 		}
 		return lidarr.NewClient(c.BaseURL, c.APIKey)
 	}
-	h.lidarrQuarantine = lidarrquarantine.NewService(h.pool, cfg, clientFn)
+	h.lidarrQuarantine = lidarrquarantine.NewService(h.pool, cfg, clientFn, h.dataDir)
 }

 // flagDirect bypasses the HTTP handler to seed a quarantine row via the
@@ -97,14 +97,16 @@ type tuningSnapshot struct {
 func (h *handlers) tuningSnapshot() tuningSnapshot {
 	var out tuningSnapshot
 	out.Profiles = map[string]weightsResp{
-		recsettings.ScopeRadio:    weightsRespFrom(h.recSettings.Weights(recsettings.ScopeRadio)),
-		recsettings.ScopeDailyMix: weightsRespFrom(h.recSettings.Weights(recsettings.ScopeDailyMix)),
+		recsettings.ScopeRadio:     weightsRespFrom(h.recSettings.Weights(recsettings.ScopeRadio)),
+		recsettings.ScopeDailyMix:  weightsRespFrom(h.recSettings.Weights(recsettings.ScopeDailyMix)),
+		recsettings.ScopeSongsLike: weightsRespFrom(h.recSettings.Weights(recsettings.ScopeSongsLike)),
 	}
 	out.Taste = tasteRespFrom(h.recSettings.Taste())
 	out.Discover = discoverRespFrom(h.recSettings.Discover())
 	out.Shipped.Profiles = map[string]weightsResp{
-		recsettings.ScopeRadio:    weightsRespFrom(recsettings.ShippedRadioWeights()),
-		recsettings.ScopeDailyMix: weightsRespFrom(recsettings.ShippedDailyMixWeights()),
+		recsettings.ScopeRadio:     weightsRespFrom(recsettings.ShippedRadioWeights()),
+		recsettings.ScopeDailyMix:  weightsRespFrom(recsettings.ShippedDailyMixWeights()),
+		recsettings.ScopeSongsLike: weightsRespFrom(recsettings.ShippedSongsLikeWeights()),
 	}
 	out.Shipped.Taste = tasteRespFrom(recsettings.ShippedTasteTuning())
 	out.Shipped.Discover = discoverRespFrom(recsettings.ShippedDiscoverTuning())
@@ -23,15 +23,17 @@ type removeTrackResponse struct {

 // handleRemoveTrack implements DELETE /api/admin/tracks/{id}?unmonitor=true|false.
 //
-// Admin-only (gated by auth.RequireAdmin on the /admin route group). Always
-// deletes the file + DB row and runs the album/artist cascade tidy-up. When
+// Admin-only (gated by auth.RequireAdmin on the /admin route group). Deletes the
+// file, then the DB row, and runs the album/artist cascade tidy-up — and deletes
+// nothing at all when the file cannot be removed (#3918). When
 // unmonitor=true and the track has an mbid, also calls Lidarr.UnmonitorTrack
 // — failure there is non-fatal (the destructive part already completed) and
 // surfaces as `lidarr_unmonitor_failed: true` in the success envelope.
 //
 // Per spec §5, Lidarr-side errors during the unmonitor step do NOT map to
-// wire error codes; the only error codes this handler emits are not_found,
-// server_error, plus the auth codes the middleware emits upstream.
+// wire error codes. The codes this handler emits are not_found,
+// library_not_writable (409) and file_delete_failed when the file could not be
+// removed, server_error, plus the auth codes the middleware emits upstream.
 func (h *handlers) handleRemoveTrack(w http.ResponseWriter, r *http.Request) {
 	idStr := chi.URLParam(r, "id")
 	trackID, ok := parseUUID(idStr)
@@ -66,6 +68,11 @@ func (h *handlers) handleRemoveTrack(w http.ResponseWriter, r *http.Request) {
 			writeErr(w, &apierror.Error{Status: http.StatusNotFound, Code: "not_found", Message: "track not found"})
 			return
 		}
+		if apiErr, ok := fileRemoveAPIError(err); ok {
+			logFileRemoveFailure(h.logger, apiErr, "track_id", idStr)
+			writeErr(w, apiErr)
+			return
+		}
 		h.logger.Error("api: remove track failed", "err", err, "track_id", idStr)
 		writeErr(w, apierror.InternalMsg("remove failed", err))
 		return
@@ -24,6 +24,7 @@ import (
 	"git.fabledsword.com/bvandeusen/minstrel/internal/playevents"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/playlists"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/reacquisition"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/recommendation"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/recsettings"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/tags"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/tracks"
@@ -32,29 +33,31 @@ import (
 // Mount attaches /api/* handlers to r. Public endpoints (login) are outside
 // RequireUser; everything else is gated by the middleware. The events writer
 // is shared with the Subsonic mount so /rest/scrobble feeds the same store.
-func Mount(r chi.Router, pool *pgxpool.Pool, logger *slog.Logger, events *playevents.Writer, recCfg config.RecommendationConfig, recSettings *recsettings.Service, lidarrCfg *lidarrconfig.Service, lidarrReqs *lidarrrequests.Service, lidarrQuar *lidarrquarantine.Service, tracksSvc *tracks.Service, playlistsSvc *playlists.Service, coverEnricher *coverart.Enricher, coverSettings *coverart.SettingsService, tagSettings *tags.SettingsService, scanner *library.Scanner, scanCfg library.RunScanConfig, dataDir string, sender mailer.Sender, bus *eventbus.Bus, playlistScheduler *playlists.Scheduler, streamSecret []byte, netSettings *netsettings.Service, reacqSettings *reacquisition.SettingsService) {
+func Mount(r chi.Router, pool *pgxpool.Pool, logger *slog.Logger, events *playevents.Writer, recCfg config.RecommendationConfig, recSettings *recsettings.Service, lidarrCfg *lidarrconfig.Service, lidarrReqs *lidarrrequests.Service, lidarrQuar *lidarrquarantine.Service, tracksSvc *tracks.Service, playlistsSvc *playlists.Service, coverEnricher *coverart.Enricher, coverSettings *coverart.SettingsService, tagSettings *tags.SettingsService, scanner *library.Scanner, scanCfg library.RunScanConfig, dataDir string, sender mailer.Sender, bus *eventbus.Bus, playlistScheduler *playlists.Scheduler, streamSecret []byte, netSettings *netsettings.Service, reacqSettings *reacquisition.SettingsService, fpSettings *library.FingerprintSettingsService) {
 	rng := rand.New(rand.NewSource(rand.Int63()))
 	h := &handlers{
 		pool: pool, logger: logger, events: events, recCfg: recCfg,
-		recSettings:       recSettings,
-		rng:               rng.Float64,
-		lidarrCfg:         lidarrCfg,
-		lidarrRequests:    lidarrReqs,
-		lidarrQuarantine:  lidarrQuar,
-		tracks:            tracksSvc,
-		playlists:         playlistsSvc,
-		coverart:          coverEnricher,
-		coverSettings:     coverSettings,
-		tagSettings:       tagSettings,
-		scanner:           scanner,
-		scanCfg:           scanCfg,
-		dataDir:           dataDir,
-		mailer:            sender,
-		eventbus:          bus,
-		playlistScheduler: playlistScheduler,
-		streamSecret:      streamSecret,
-		netSettings:       netSettings,
-		reacqSettings:     reacqSettings,
+		recSettings:         recSettings,
+		rng:                 rng.Float64,
+		lidarrCfg:           lidarrCfg,
+		lidarrRequests:      lidarrReqs,
+		lidarrQuarantine:    lidarrQuar,
+		tracks:              tracksSvc,
+		playlists:           playlistsSvc,
+		coverart:            coverEnricher,
+		coverSettings:       coverSettings,
+		tagSettings:         tagSettings,
+		scanner:             scanner,
+		scanCfg:             scanCfg,
+		dataDir:             dataDir,
+		mailer:              sender,
+		eventbus:            bus,
+		playlistScheduler:   playlistScheduler,
+		streamSecret:        streamSecret,
+		netSettings:         netSettings,
+		reacqSettings:       reacqSettings,
+		fingerprintSettings: fpSettings,
+		librarySize:         recommendation.NewLibrarySize(nil),
 	}

 	r.Route("/api", func(api chi.Router) {
@@ -213,6 +216,16 @@ func Mount(r chi.Router, pool *pgxpool.Pool, logger *slog.Logger, events *playev
 				admin.Get("/library/missing", h.handleListMissingTracks)

 				admin.Get("/library/coverage", h.handleGetLibraryCoverage)
+				admin.Get("/library/fingerprints", h.handleGetFingerprintCoverage)
+				admin.Get("/library/fingerprint-settings", h.handleGetFingerprintSettings)
+				admin.Put("/library/fingerprint-settings", h.handleUpdateFingerprintSettings)
+				// Duplicates report (#3912): proposals from the duplicate sweep, a
+				// trigger to sweep now, dismissal, and the merge (#3911), which deletes
+				// the removed copies' files after moving their history onto the kept one.
+				admin.Get("/library/duplicates", h.handleListDuplicates)
+				admin.Post("/library/duplicates/sweep", h.handleRunDuplicateSweep)
+				admin.Post("/library/duplicates/{id}/dismiss", h.handleDismissDuplicateGroup)
+				admin.Post("/library/duplicates/{id}/merge", h.handleMergeDuplicateGroup)

 				admin.Get("/invites", h.handleListInvites)
 				admin.Post("/invites", h.handleCreateInvite)
@@ -268,12 +281,16 @@ func Mount(r chi.Router, pool *pgxpool.Pool, logger *slog.Logger, events *playev
 }

 type handlers struct {
-	pool              *pgxpool.Pool
-	logger            *slog.Logger
-	events            *playevents.Writer
-	recCfg            config.RecommendationConfig
-	recSettings       *recsettings.Service
-	rng               func() float64
+	pool        *pgxpool.Pool
+	logger      *slog.Logger
+	events      *playevents.Writer
+	recCfg      config.RecommendationConfig
+	recSettings *recsettings.Service
+	rng         func() float64
+	// librarySize memoises the track count that sizes the candidate pool
+	// (#3880). Held here rather than counted per request: the count is a
+	// full table scan, and library size only moves when a scan runs.
+	librarySize       *recommendation.LibrarySize
 	lidarrCfg         *lidarrconfig.Service
 	lidarrRequests    *lidarrrequests.Service
 	lidarrQuarantine  *lidarrquarantine.Service
@@ -292,6 +309,10 @@ type handlers struct {
 	// missing files (milestone #290) — grace window, backoff, attempt caps.
 	// Cached in the service, so the admin card reads it without a query.
 	reacqSettings *reacquisition.SettingsService
+	// fingerprintSettings is the fingerprinting policy (M400 #3913), the same
+	// instance the scanner and the fingerprint workers read, so a save from the
+	// admin card reaches them without a restart. Nil serves the defaults.
+	fingerprintSettings *library.FingerprintSettingsService
 	// netSettings caches the trusted reverse-proxy depth read by the auth
 	// middleware on every request and edited from the admin network card.
 	netSettings *netsettings.Service
@@ -65,7 +65,7 @@ func testHandlers(t *testing.T) (*handlers, *pgxpool.Pool) {
 	}
 	lidarrCfg := lidarrconfig.New(pool)
 	lidarrReqs := lidarrrequests.NewService(pool, lidarrCfg, nil, nil)
-	lidarrQuar := lidarrquarantine.NewService(pool, lidarrCfg, nil)
+	lidarrQuar := lidarrquarantine.NewService(pool, lidarrCfg, nil, "")
 	// tracks.Service has no Lidarr unmonitorer in tests by default; the
 	// admin-tracks tests below override h.tracks via installTracksLidarrStub
 	// when they need a stubbed Lidarr.
@@ -6,19 +6,21 @@ package api
 // /app/client/ at image build time.
 //
 // Both endpoints are authenticated — the bandwidth cost of the APK
-// (~30-60 MB) makes anonymous access an abuse vector. The Flutter
-// client's polling only fires after login (banner mounts in the post-
-// login shell), so this gate is invisible to the actual update flow.
+// (~30-60 MB) makes anonymous access an abuse vector. The client only
+// polls after login, so this gate is invisible to the actual update flow.
 //
 // /api/client/apk additionally rate-limits per user to a single
 // download every 60s. Real install flows fire one download per
 // update; anything tighter is scripted/abusive.
 //
 // Returns 404 gracefully when the APK isn't present (dev environments,
-// pre-CI-wiring); the Flutter client treats 404 as "no update channel
-// available."
+// pre-CI-wiring); the client treats 404 as "no update channel available."
+//
+// (These paragraphs said "the Flutter client" until 2026-09-10. That client
+// was deleted in v2026.08.18 — the Android app is the only one now.)

 import (
+	"encoding/json"
 	"errors"
 	"net/http"
 	"os"
@@ -84,8 +86,36 @@ func clientAPKAllowDownload(userID string, now time.Time) time.Duration {
 	return 0
 }

+// clientVersionSidecar is the JSON written beside the bundled APK by
+// release.yml. It carries three values that are deliberately separate:
+//
+//   - Name is a LABEL for people, "YYYY.MM.DD.HHMM" from the commit's
+//     timestamp. Two channels carrying the same code report the same name.
+//   - Code is the ORDERING KEY, minutes since 2020-01-01 at build time, and
+//     is the value Android itself installs by. It answers "may this be
+//     installed over that?" — the name never does.
+//   - Channel is a SIBLING FIELD, never a suffix inside the name.
+//
+// JSON rather than a positional line on purpose. The obvious growth path for
+// the old one-value file was "<name> <code>", which a first-space split
+// silently mangles the moment a third field appears: the code stops parsing,
+// and the reader falls back to name comparison WITHOUT erroring.
+type clientVersionSidecar struct {
+	Name string `json:"name"`
+	// Pointer, not int64: absent must stay distinguishable from zero. An
+	// artifact published before codes were recorded genuinely has no code —
+	// zero would claim it is infinitely old rather than unknown.
+	Code    *int64 `json:"code"`
+	Channel string `json:"channel"`
+}
+
 type clientVersionResponse struct {
-	Version   string `json:"version"`
+	Version string `json:"version"`
+	// omitempty on both: the client must be able to tell "this server does
+	// not report a code" from "this build's code is 0", because those call
+	// for different behaviour on the other end.
+	Code      *int64 `json:"code,omitempty"`
+	Channel   string `json:"channel,omitempty"`
 	APKURL    string `json:"apk_url"`
 	SizeBytes int64  `json:"size_bytes"`
 }
@@ -117,8 +147,25 @@ func (h *handlers) handleClientVersion(w http.ResponseWriter, _ *http.Request) {
 		return
 	}

+	var sidecar clientVersionSidecar
+	if err := json.Unmarshal(versionBytes, &sidecar); err != nil {
+		// Fail LOUDLY rather than serving a blank version. The failure mode
+		// this avoids is the one that never gets reported: if an unreadable
+		// sidecar produced an empty name, every client would compare against
+		// nothing, conclude it was current, and go quiet — "I cannot read
+		// this" and "there is nothing newer" would be the same answer.
+		writeErrWithLog(w, h.logger, "client_version: sidecar is not valid JSON", err)
+		return
+	}
+	if sidecar.Name == "" {
+		http.Error(w, `{"error":{"code":"bad_client_version","message":"version sidecar has no name"}}`, http.StatusInternalServerError)
+		return
+	}
+
 	writeJSON(w, http.StatusOK, clientVersionResponse{
-		Version:   strings.TrimSpace(string(versionBytes)),
+		Version:   strings.TrimSpace(sidecar.Name),
+		Code:      sidecar.Code,
+		Channel:   strings.TrimSpace(sidecar.Channel),
 		APKURL:    "/api/client/apk",
 		SizeBytes: stat.Size(),
 	})
@@ -77,18 +77,32 @@ func TestClientVersion_404WhenAPKButNoVersion(t *testing.T) {
 	}
 }

-func TestClientVersion_200WithBothFiles(t *testing.T) {
+// writeClientAssets stages an APK plus a raw sidecar body, and returns the
+// APK's size so callers can assert size_bytes without recomputing it.
+func writeClientAssets(t *testing.T, sidecar string) int64 {
+	t.Helper()
 	dir := withClientAPKDir(t)
 	body := []byte("fake apk content")
 	if err := os.WriteFile(filepath.Join(dir, clientAPKFilename), body, 0o644); err != nil {
 		t.Fatal(err)
 	}
-	if err := os.WriteFile(filepath.Join(dir, clientVersionFile), []byte("v2026.05.10\n"), 0o644); err != nil {
+	if err := os.WriteFile(filepath.Join(dir, clientVersionFile), []byte(sidecar), 0o644); err != nil {
 		t.Fatal(err)
 	}
+	return int64(len(body))
+}
+
+func getClientVersion(t *testing.T) *httptest.ResponseRecorder {
+	t.Helper()
 	h := &handlers{logger: slog.New(slog.NewTextHandler(io.Discard, nil))}
 	rr := httptest.NewRecorder()
 	h.handleClientVersion(rr, httptest.NewRequest(http.MethodGet, "/api/client/version", nil))
+	return rr
+}
+
+func TestClientVersion_200WithBothFiles(t *testing.T) {
+	size := writeClientAssets(t, `{"name":"2026.09.10.1432","code":3523847,"channel":"stable"}`+"\n")
+	rr := getClientVersion(t)
 	if rr.Code != http.StatusOK {
 		t.Fatalf("want 200, got %d (body: %s)", rr.Code, rr.Body.String())
 	}
@@ -96,14 +110,69 @@ func TestClientVersion_200WithBothFiles(t *testing.T) {
 	if err := json.Unmarshal(rr.Body.Bytes(), &resp); err != nil {
 		t.Fatal(err)
 	}
-	if resp.Version != "v2026.05.10" {
-		t.Errorf("version: want trimmed v2026.05.10, got %q", resp.Version)
+	if resp.Version != "2026.09.10.1432" {
+		t.Errorf("version: want 2026.09.10.1432, got %q", resp.Version)
+	}
+	if resp.Code == nil {
+		t.Fatal("code: want 3523847, got absent — the client decides on this, so absent means it silently falls back to name comparison")
+	}
+	if *resp.Code != 3523847 {
+		t.Errorf("code: want 3523847, got %d", *resp.Code)
+	}
+	if resp.Channel != "stable" {
+		t.Errorf("channel: want stable, got %q", resp.Channel)
 	}
 	if resp.APKURL != "/api/client/apk" {
 		t.Errorf("apk_url: want /api/client/apk, got %q", resp.APKURL)
 	}
-	if resp.SizeBytes != int64(len(body)) {
-		t.Errorf("size_bytes: want %d, got %d", len(body), resp.SizeBytes)
+	if resp.SizeBytes != size {
+		t.Errorf("size_bytes: want %d, got %d", size, resp.SizeBytes)
+	}
+}
+
+// A release published before ordering keys were recorded has a name and
+// genuinely no code. That must arrive as ABSENT, not as 0 — zero would claim
+// the build is infinitely old and offer an update to everyone forever.
+func TestClientVersion_CodeAbsentIsOmittedNotZero(t *testing.T) {
+	writeClientAssets(t, `{"name":"2026.09.09","code":null,"channel":"stable"}`)
+	rr := getClientVersion(t)
+	if rr.Code != http.StatusOK {
+		t.Fatalf("want 200, got %d (body: %s)", rr.Code, rr.Body.String())
+	}
+	var resp clientVersionResponse
+	if err := json.Unmarshal(rr.Body.Bytes(), &resp); err != nil {
+		t.Fatal(err)
+	}
+	if resp.Code != nil {
+		t.Errorf("code: want absent, got %d", *resp.Code)
+	}
+	// The wire must omit the key entirely, so a client can distinguish
+	// "this server reports no code" from "this build's code is 0".
+	var raw map[string]any
+	if err := json.Unmarshal(rr.Body.Bytes(), &raw); err != nil {
+		t.Fatal(err)
+	}
+	if _, present := raw["code"]; present {
+		t.Errorf("code key should be omitted entirely, body was %s", rr.Body.String())
+	}
+}
+
+// The failure this guards is the one nobody reports: if an unreadable sidecar
+// produced an empty version, every client would compare against nothing,
+// decide it was current, and go quiet. "I cannot read this" and "there is
+// nothing newer" must not be the same answer.
+func TestClientVersion_MalformedSidecarErrorsRatherThanReportingNothing(t *testing.T) {
+	for _, sidecar := range []string{
+		"2026.09.10.1432",              // the OLD plain-text format
+		`{"name":"x",`,                 // truncated JSON
+		`{"code":123,"channel":"dev"}`, // valid JSON, no name
+		"",
+	} {
+		writeClientAssets(t, sidecar)
+		rr := getClientVersion(t)
+		if rr.Code == http.StatusOK {
+			t.Errorf("sidecar %q: want an error status, got 200 with body %s", sidecar, rr.Body.String())
+		}
 	}
 }

@@ -0,0 +1,58 @@
+package api
+
+import (
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/http"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/apierror"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
+)
+
+// fileRemoveAPIError answers a delete that could not reach the track's file
+// (#3918). Both delete endpoints use it, so the operator gets the same
+// explanation from the admin remove-track action and from quarantine's Delete
+// file.
+//
+// The unwritable case is a 409 rather than a 500 because nothing is broken: the
+// request conflicts with how the library is mounted, and the fix is the
+// operator's. The message names the directory — removal writes to the parent,
+// not the file — and the uid/gid the process runs as, which is the half of a
+// permission problem invisible from the host. Every case says nothing was
+// deleted, because that is exactly what the operator will be worried about.
+func fileRemoveAPIError(err error) (*apierror.Error, bool) {
+	var fre *library.FileRemoveError
+	if !errors.As(err, &fre) {
+		return nil, false
+	}
+	if fre.NotWritable() {
+		return &apierror.Error{
+			Status: http.StatusConflict,
+			Code:   "library_not_writable",
+			Message: fmt.Sprintf(
+				"Minstrel runs as uid %d, gid %d and cannot delete from %s (%s). "+
+					"The library mount must be writable by that user. Nothing was deleted.",
+				fre.UID, fre.GID, fre.Dir(), fre.Reason()),
+			Cause: err,
+		}, true
+	}
+	return &apierror.Error{
+		Status:  http.StatusInternalServerError,
+		Code:    "file_delete_failed",
+		Message: fmt.Sprintf("Could not delete %s (%s). Nothing was deleted.", fre.Path, fre.Reason()),
+		Cause:   err,
+	}, true
+}
+
+// logFileRemoveFailure records a delete that could not reach its file. An
+// unwritable library is an environment fact the operator can fix, so it is a
+// Warn; anything else is a real fault.
+func logFileRemoveFailure(logger *slog.Logger, apiErr *apierror.Error, attrs ...any) {
+	attrs = append(attrs, "code", apiErr.Code, "err", apiErr.Cause)
+	if apiErr.Status == http.StatusConflict {
+		logger.Warn("api: track file could not be deleted", attrs...)
+		return
+	}
+	logger.Error("api: track file could not be deleted", attrs...)
+}
@@ -0,0 +1,96 @@
+package api
+
+import (
+	"errors"
+	"fmt"
+	"io/fs"
+	"net/http"
+	"strings"
+	"syscall"
+	"testing"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/library"
+)
+
+const removeTestPath = "/music/Moe Shop/WWW (2020)/01 - WWW.mp3"
+
+// removeFailure builds the error a delete service returns when the file would
+// not go, wrapped the way lidarrquarantine.DeleteFile and tracks.RemoveTrack
+// wrap it — the mapping has to see through that.
+func removeFailure(errno syscall.Errno) error {
+	return fmt.Errorf("delete file: %w", &library.FileRemoveError{
+		Path: removeTestPath, UID: 1000, GID: 1000,
+		Err: &fs.PathError{Op: "remove", Path: removeTestPath, Err: errno},
+	})
+}
+
+func TestFileRemoveAPIError(t *testing.T) {
+	cases := []struct {
+		name       string
+		errno      syscall.Errno
+		wantStatus int
+		wantCode   string
+		wantIn     []string
+	}{
+		{
+			name: "read-only mount", errno: syscall.EROFS,
+			wantStatus: http.StatusConflict, wantCode: "library_not_writable",
+			wantIn: []string{"uid 1000, gid 1000", "/music/Moe Shop/WWW (2020)", "read-only file system", "Nothing was deleted"},
+		},
+		{
+			name: "permission denied", errno: syscall.EACCES,
+			wantStatus: http.StatusConflict, wantCode: "library_not_writable",
+			wantIn: []string{"permission denied", "Nothing was deleted"},
+		},
+		{
+			name: "operation not permitted", errno: syscall.EPERM,
+			wantStatus: http.StatusConflict, wantCode: "library_not_writable",
+			wantIn: []string{"operation not permitted"},
+		},
+		{
+			name: "i/o error", errno: syscall.EIO,
+			wantStatus: http.StatusInternalServerError, wantCode: "file_delete_failed",
+			wantIn: []string{removeTestPath, "input/output error", "Nothing was deleted"},
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			apiErr, ok := fileRemoveAPIError(removeFailure(tc.errno))
+			if !ok {
+				t.Fatal("a wrapped *library.FileRemoveError was not recognised")
+			}
+			if apiErr.Status != tc.wantStatus || apiErr.Code != tc.wantCode {
+				t.Fatalf("got %d %s, want %d %s", apiErr.Status, apiErr.Code, tc.wantStatus, tc.wantCode)
+			}
+			for _, want := range tc.wantIn {
+				if !strings.Contains(apiErr.Message, want) {
+					t.Errorf("message %q lacks %q", apiErr.Message, want)
+				}
+			}
+		})
+	}
+}
+
+// The unwritable answer must name the DIRECTORY. Removal needs write access to
+// the parent, so a message naming the file would send the operator to fix the
+// wrong permissions. The directory is a prefix of the file path, which is why a
+// plain "contains the directory" check could never catch that regression.
+func TestFileRemoveAPIError_NotWritableNamesTheDirectoryNotTheFile(t *testing.T) {
+	apiErr, _ := fileRemoveAPIError(removeFailure(syscall.EROFS))
+	if strings.Contains(apiErr.Message, "01 - WWW.mp3") {
+		t.Fatalf("message names the file rather than its directory: %q", apiErr.Message)
+	}
+}
+
+func TestFileRemoveAPIError_IgnoresOtherErrors(t *testing.T) {
+	for name, err := range map[string]error{
+		"nil":         nil,
+		"plain error": errors.New("delete track: connection reset"),
+		"path error":  &fs.PathError{Op: "remove", Path: removeTestPath, Err: syscall.EROFS},
+		"not found":   library.ErrTrackNotFound,
+	} {
+		if _, ok := fileRemoveAPIError(err); ok {
+			t.Errorf("%s: mapped as a file-remove failure", name)
+		}
+	}
+}
@@ -465,7 +465,7 @@ func TestRoutesRegisteredInMount(t *testing.T) {
 	r := chi.NewRouter()
 	w := playevents.NewWriter(h.pool, slog.New(slog.NewTextHandler(io.Discard, nil)),
 		30*time.Minute, 0.5, 30000)
-	Mount(r, h.pool, h.logger, w, config.RecommendationConfig{RadioSize: 50, RadioSizeMax: 200, RecentlyPlayedHours: 1}, h.recSettings, h.lidarrCfg, h.lidarrRequests, h.lidarrQuarantine, h.tracks, h.playlists, h.coverart, h.coverSettings, h.tagSettings, h.scanner, h.scanCfg, h.dataDir, nil, eventbus.New(), nil, nil, h.netSettings, nil)
+	Mount(r, h.pool, h.logger, w, config.RecommendationConfig{RadioSize: 50, RadioSizeMax: 200, RecentlyPlayedHours: 1}, h.recSettings, h.lidarrCfg, h.lidarrRequests, h.lidarrQuarantine, h.tracks, h.playlists, h.coverart, h.coverSettings, h.tagSettings, h.scanner, h.scanCfg, h.dataDir, nil, eventbus.New(), nil, nil, h.netSettings, nil, nil)

 	paths := []string{
 		"/api/artists",
@@ -484,6 +484,9 @@ func TestRoutesRegisteredInMount(t *testing.T) {
 		// wired.
 		"/api/admin/library/missing",
 		"/api/admin/library/reacquisition",
+		"/api/admin/library/fingerprints",
+		"/api/admin/library/fingerprint-settings",
+		"/api/admin/library/duplicates",
 	}
 	for _, p := range paths {
 		req := httptest.NewRequest(http.MethodGet, p, nil)
@@ -87,10 +87,24 @@ func (h *handlers) handleRadio(w http.ResponseWriter, r *http.Request) {
 	currentVec.DeviceClass = latestDeviceClass(r.Context(), q, user.ID, h.logger)

 	exclude := parseExcludeParam(r.URL.Query().Get("exclude"))
-	limits := recommendation.DefaultCandidateSourceLimits()
+	// Size the pool to the library (#3880). A fixed ~170 candidates samples a
+	// shrinking fraction of a growing collection, which is what made the
+	// recommendations feel less relevant as the library grew. Degrades to the
+	// base limits if the count is unavailable — never fails the request over a
+	// sizing hint.
+	librarySize := h.librarySize.Get(r.Context(), func(ctx context.Context) (int64, error) {
+		return recommendation.CountLibraryTracks(ctx, q)
+	})
+	limits := recommendation.ScaleForLibrary(
+		recommendation.DefaultCandidateSourceLimits(), librarySize,
+	)
 	candidates, err := recommendation.LoadCandidatesFromSimilarity(
 		r.Context(), q, user.ID, seedID,
 		h.recCfg.RecentlyPlayedHours, currentVec, exclude, limits,
+		// A fresh seed per request (#3889): radio is a new session each time
+		// and SHOULD draw differently. The system mixes are the surfaces that
+		// promise repeatability; this is not one of them.
+		strconv.FormatInt(time.Now().UnixNano(), 36),
 	)
 	if err != nil {
 		h.logger.Warn("api: radio: similarity-pool failed; falling back to whole-library", "err", err)
@@ -108,7 +122,15 @@ func (h *handlers) handleRadio(w http.ResponseWriter, r *http.Request) {
 	// Scoring weights come from the DB-backed tuning lab (#1250) —
 	// read per request so an admin change takes effect live.
 	weights := h.recSettings.Weights(recsettings.ScopeRadio)
-	picks := recommendation.Shuffle(candidates, weights, time.Now().UTC(), h.rng, limit-1)
+	// Diversity caps (#3882). Radio had none while every sibling surface did,
+	// which is how a whole session could come back from one artist. Scaled to
+	// the requested length so a 20-track radio and a 200-track one are capped
+	// alike; Shuffle relaxes them rather than returning a short radio.
+	//
+	// limit-1 because the seed track occupies the first slot and is prepended
+	// below — the caps govern the tracks that FOLLOW it.
+	caps := recommendation.RadioDiversityCaps(limit - 1)
+	picks := recommendation.Shuffle(candidates, weights, time.Now().UTC(), h.rng, limit-1, caps)

 	out := make([]TrackRef, 0, len(picks)+1)
 	out = append(out, trackRefFrom(track, album.Title, artist.Name))
@@ -55,6 +55,11 @@ const (
 	// exercised.
 	ActionSessionRevoke       Action = "session_revoke"
 	ActionSessionRevokeOthers Action = "session_revoke_others"
+
+	// Duplicate merge (#3911). Irreversible: a copy's file and row are removed
+	// and its history moved onto the copy kept. The metadata names both, so the
+	// log can answer "where did that file go" long after the report is gone.
+	ActionDuplicateMerge Action = "duplicate_merge"
 )

 // Write inserts one audit_log row. metadata is marshaled as JSON;
@@ -168,6 +168,7 @@ func TestWrite_AllActionConstantsArePersisted(t *testing.T) {
 		audit.ActionTokenRegenerate,
 		audit.ActionForgotPasswordInit,
 		audit.ActionPasswordResetByEmail,
+		audit.ActionDuplicateMerge,
 	}
 	for _, a := range actions {
 		if err := audit.Write(context.Background(), pool, nilUUID, nilUUID, a, nil); err != nil {
@@ -0,0 +1,457 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: duplicates.sql
+
+package dbq
+
+import (
+	"context"
+
+	"github.com/jackc/pgx/v5/pgtype"
+)
+
+const addDuplicateGroupMember = `-- name: AddDuplicateGroupMember :exec
+INSERT INTO duplicate_group_members (group_id, track_id)
+VALUES ($1, $2)
+ON CONFLICT DO NOTHING
+`
+
+type AddDuplicateGroupMemberParams struct {
+	GroupID pgtype.UUID
+	TrackID pgtype.UUID
+}
+
+func (q *Queries) AddDuplicateGroupMember(ctx context.Context, arg AddDuplicateGroupMemberParams) error {
+	_, err := q.db.Exec(ctx, addDuplicateGroupMember, arg.GroupID, arg.TrackID)
+	return err
+}
+
+const countPendingDuplicateGroups = `-- name: CountPendingDuplicateGroups :one
+SELECT count(*)::bigint
+  FROM duplicate_groups g
+ WHERE g.status = 'pending'
+   AND (SELECT count(*) FROM duplicate_group_members m WHERE m.group_id = g.id) >= 2
+`
+
+// Proposals awaiting review. A group left with one member — its other tracks
+// deleted since the sweep — is no proposal at all and is not counted; the next
+// sweep retires it.
+func (q *Queries) CountPendingDuplicateGroups(ctx context.Context) (int64, error) {
+	row := q.db.QueryRow(ctx, countPendingDuplicateGroups)
+	var column_1 int64
+	err := row.Scan(&column_1)
+	return column_1, err
+}
+
+const deleteStalePendingDuplicateGroups = `-- name: DeleteStalePendingDuplicateGroups :execrows
+DELETE FROM duplicate_groups g
+ WHERE g.status = 'pending'
+   AND g.last_seen_sweep_id IS DISTINCT FROM $1
+   AND (g.last_seen_sweep_id IS NULL
+        OR (SELECT s.started_at FROM duplicate_sweeps s WHERE s.id = g.last_seen_sweep_id)
+           < (SELECT s.started_at FROM duplicate_sweeps s WHERE s.id = $1))
+`
+
+// A pending proposal this sweep did not find again no longer describes the
+// library: a member was re-fingerprinted, merged away or went missing. Dismissed
+// groups are kept regardless — they are the memory of a decision.
+//
+// Only proposals last confirmed by an EARLIER sweep go. Should two sweeps ever
+// overlap (a manual trigger racing the worker), neither may delete what the other
+// has just found.
+func (q *Queries) DeleteStalePendingDuplicateGroups(ctx context.Context, sweepID pgtype.UUID) (int64, error) {
+	result, err := q.db.Exec(ctx, deleteStalePendingDuplicateGroups, sweepID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const dismissDuplicateGroup = `-- name: DismissDuplicateGroup :execrows
+UPDATE duplicate_groups
+   SET status = 'dismissed', resolved_at = now()
+ WHERE id = $1 AND status = 'pending'
+`
+
+// "These are not duplicates." Only a pending group can be dismissed; zero rows
+// means it was already resolved or no longer exists.
+func (q *Queries) DismissDuplicateGroup(ctx context.Context, id pgtype.UUID) (int64, error) {
+	result, err := q.db.Exec(ctx, dismissDuplicateGroup, id)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const finishDuplicateSweep = `-- name: FinishDuplicateSweep :exec
+UPDATE duplicate_sweeps
+   SET finished_at       = now(),
+       candidates        = $1,
+       groups_found      = $2,
+       oversize_clusters = $3,
+       error_message     = NULLIF($4::text, '')
+ WHERE id = $5
+`
+
+type FinishDuplicateSweepParams struct {
+	Candidates       *int32
+	GroupsFound      *int32
+	OversizeClusters *int32
+	ErrorMessage     string
+	ID               pgtype.UUID
+}
+
+func (q *Queries) FinishDuplicateSweep(ctx context.Context, arg FinishDuplicateSweepParams) error {
+	_, err := q.db.Exec(ctx, finishDuplicateSweep,
+		arg.Candidates,
+		arg.GroupsFound,
+		arg.OversizeClusters,
+		arg.ErrorMessage,
+		arg.ID,
+	)
+	return err
+}
+
+const getInFlightDuplicateSweep = `-- name: GetInFlightDuplicateSweep :one
+SELECT id, started_at
+  FROM duplicate_sweeps
+ WHERE finished_at IS NULL
+ ORDER BY started_at DESC
+ LIMIT 1
+`
+
+type GetInFlightDuplicateSweepRow struct {
+	ID        pgtype.UUID
+	StartedAt pgtype.Timestamptz
+}
+
+// The guard against two sweeps at once: "in flight" is finished_at IS NULL.
+func (q *Queries) GetInFlightDuplicateSweep(ctx context.Context) (GetInFlightDuplicateSweepRow, error) {
+	row := q.db.QueryRow(ctx, getInFlightDuplicateSweep)
+	var i GetInFlightDuplicateSweepRow
+	err := row.Scan(&i.ID, &i.StartedAt)
+	return i, err
+}
+
+const getLatestDuplicateSweep = `-- name: GetLatestDuplicateSweep :one
+SELECT id, started_at, finished_at, candidates, groups_found, oversize_clusters, error_message
+  FROM duplicate_sweeps
+ ORDER BY started_at DESC
+ LIMIT 1
+`
+
+func (q *Queries) GetLatestDuplicateSweep(ctx context.Context) (DuplicateSweep, error) {
+	row := q.db.QueryRow(ctx, getLatestDuplicateSweep)
+	var i DuplicateSweep
+	err := row.Scan(
+		&i.ID,
+		&i.StartedAt,
+		&i.FinishedAt,
+		&i.Candidates,
+		&i.GroupsFound,
+		&i.OversizeClusters,
+		&i.ErrorMessage,
+	)
+	return i, err
+}
+
+const getLatestFingerprintComputedAt = `-- name: GetLatestFingerprintComputedAt :one
+SELECT max(computed_at)::timestamptz AS latest FROM track_fingerprints
+`
+
+// Whether a sweep has anything new to look at: fingerprints written since the
+// last sweep started.
+func (q *Queries) GetLatestFingerprintComputedAt(ctx context.Context) (pgtype.Timestamptz, error) {
+	row := q.db.QueryRow(ctx, getLatestFingerprintComputedAt)
+	var latest pgtype.Timestamptz
+	err := row.Scan(&latest)
+	return latest, err
+}
+
+const listDismissedDuplicateMemberSets = `-- name: ListDismissedDuplicateMemberSets :many
+SELECT g.id, array_agg(m.track_id ORDER BY m.track_id)::uuid[] AS track_ids
+  FROM duplicate_groups g
+  JOIN duplicate_group_members m ON m.group_id = g.id
+ WHERE g.status = 'dismissed'
+ GROUP BY g.id
+`
+
+type ListDismissedDuplicateMemberSetsRow struct {
+	ID       pgtype.UUID
+	TrackIds []pgtype.UUID
+}
+
+// What the operator has already said are not duplicates. A new proposal whose
+// every member sat together in one of these is not proposed again.
+func (q *Queries) ListDismissedDuplicateMemberSets(ctx context.Context) ([]ListDismissedDuplicateMemberSetsRow, error) {
+	rows, err := q.db.Query(ctx, listDismissedDuplicateMemberSets)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListDismissedDuplicateMemberSetsRow
+	for rows.Next() {
+		var i ListDismissedDuplicateMemberSetsRow
+		if err := rows.Scan(&i.ID, &i.TrackIds); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const listDuplicateCandidates = `-- name: ListDuplicateCandidates :many
+SELECT t.id, t.duration_ms, f.audio_stream_sha256, f.chromaprint
+  FROM tracks t
+  JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL
+   AND f.fingerprint_version >= $1
+   AND f.chromaprint IS NOT NULL
+   -- Only chromaprints taken at the current length: prints at two lengths are not
+   -- comparable, and after a length change the backfill is still re-deriving the
+   -- rest (#3913).
+   AND f.chromaprint_length_sec = $2
+   AND (t.duration_ms, t.id) > ($3::integer, $4::uuid)
+ ORDER BY t.duration_ms, t.id
+ LIMIT $5
+`
+
+type ListDuplicateCandidatesParams struct {
+	CurrentVersion       int16
+	ChromaprintLengthSec int32
+	AfterDurationMs      int32
+	AfterID              pgtype.UUID
+	PageLimit            int32
+}
+
+type ListDuplicateCandidatesRow struct {
+	ID                pgtype.UUID
+	DurationMs        int32
+	AudioStreamSha256 []byte
+	Chromaprint       []int32
+}
+
+// The acoustic tier's input, one page at a time in (duration_ms, id) order so the
+// sweep holds only a sliding window of durations. Tracks without a chromaprint
+// cannot be compared acoustically and are left out; any exact duplicates among
+// them come from ListExactDuplicateHashes.
+func (q *Queries) ListDuplicateCandidates(ctx context.Context, arg ListDuplicateCandidatesParams) ([]ListDuplicateCandidatesRow, error) {
+	rows, err := q.db.Query(ctx, listDuplicateCandidates,
+		arg.CurrentVersion,
+		arg.ChromaprintLengthSec,
+		arg.AfterDurationMs,
+		arg.AfterID,
+		arg.PageLimit,
+	)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListDuplicateCandidatesRow
+	for rows.Next() {
+		var i ListDuplicateCandidatesRow
+		if err := rows.Scan(
+			&i.ID,
+			&i.DurationMs,
+			&i.AudioStreamSha256,
+			&i.Chromaprint,
+		); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const listExactDuplicateHashes = `-- name: ListExactDuplicateHashes :many
+SELECT f.audio_stream_sha256,
+       array_agg(t.id ORDER BY t.id)::uuid[] AS track_ids
+  FROM track_fingerprints f
+  JOIN tracks t ON t.id = f.track_id
+ WHERE t.missing_since IS NULL
+   AND f.fingerprint_version >= $1
+   AND f.audio_stream_sha256 IS NOT NULL
+ GROUP BY f.audio_stream_sha256
+HAVING count(*) > 1
+`
+
+type ListExactDuplicateHashesRow struct {
+	AudioStreamSha256 []byte
+	TrackIds          []pgtype.UUID
+}
+
+// The exact tier, library-wide in one pass: identical encoded audio shared by
+// more than one present track.
+func (q *Queries) ListExactDuplicateHashes(ctx context.Context, currentVersion int16) ([]ListExactDuplicateHashesRow, error) {
+	rows, err := q.db.Query(ctx, listExactDuplicateHashes, currentVersion)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListExactDuplicateHashesRow
+	for rows.Next() {
+		var i ListExactDuplicateHashesRow
+		if err := rows.Scan(&i.AudioStreamSha256, &i.TrackIds); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const listPendingDuplicateGroupMembers = `-- name: ListPendingDuplicateGroupMembers :many
+WITH page AS (
+    SELECT g.id, g.tier, g.worst_bit_error_rate, g.detected_at
+      FROM duplicate_groups g
+     WHERE g.status = 'pending'
+       AND (SELECT count(*) FROM duplicate_group_members m WHERE m.group_id = g.id) >= 2
+     ORDER BY g.detected_at DESC, g.id
+     LIMIT $2 OFFSET $1
+)
+SELECT p.id                  AS group_id,
+       p.tier,
+       p.worst_bit_error_rate,
+       p.detected_at,
+       t.id                  AS track_id,
+       t.title,
+       artists.name          AS artist_name,
+       albums.id             AS album_id,
+       albums.title          AS album_title,
+       t.file_path,
+       t.file_format,
+       t.file_size,
+       t.duration_ms,
+       t.added_at,
+       (SELECT count(*) FROM general_likes l WHERE l.track_id = t.id)::bigint AS like_count,
+       (SELECT count(*) FROM play_events e WHERE e.track_id = t.id)::bigint   AS play_count
+  FROM page p
+  JOIN duplicate_group_members m ON m.group_id = p.id
+  JOIN tracks t  ON t.id = m.track_id
+  JOIN albums    ON albums.id  = t.album_id
+  JOIN artists   ON artists.id = t.artist_id
+ ORDER BY p.detected_at DESC, p.id, t.id
+`
+
+type ListPendingDuplicateGroupMembersParams struct {
+	PageOffset int32
+	PageLimit  int32
+}
+
+type ListPendingDuplicateGroupMembersRow struct {
+	GroupID           pgtype.UUID
+	Tier              string
+	WorstBitErrorRate *float32
+	DetectedAt        pgtype.Timestamptz
+	TrackID           pgtype.UUID
+	Title             string
+	ArtistName        string
+	AlbumID           pgtype.UUID
+	AlbumTitle        string
+	FilePath          string
+	FileFormat        string
+	FileSize          int64
+	DurationMs        int32
+	AddedAt           pgtype.Timestamptz
+	LikeCount         int64
+	PlayCount         int64
+}
+
+// One page of proposals, newest first, flattened to one row per member so the
+// handler folds them without a query per group. What each copy carries — likes
+// and plays from every user — is here because it is what the operator weighs
+// when deciding which copy to keep.
+func (q *Queries) ListPendingDuplicateGroupMembers(ctx context.Context, arg ListPendingDuplicateGroupMembersParams) ([]ListPendingDuplicateGroupMembersRow, error) {
+	rows, err := q.db.Query(ctx, listPendingDuplicateGroupMembers, arg.PageOffset, arg.PageLimit)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListPendingDuplicateGroupMembersRow
+	for rows.Next() {
+		var i ListPendingDuplicateGroupMembersRow
+		if err := rows.Scan(
+			&i.GroupID,
+			&i.Tier,
+			&i.WorstBitErrorRate,
+			&i.DetectedAt,
+			&i.TrackID,
+			&i.Title,
+			&i.ArtistName,
+			&i.AlbumID,
+			&i.AlbumTitle,
+			&i.FilePath,
+			&i.FileFormat,
+			&i.FileSize,
+			&i.DurationMs,
+			&i.AddedAt,
+			&i.LikeCount,
+			&i.PlayCount,
+		); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const startDuplicateSweep = `-- name: StartDuplicateSweep :one
+INSERT INTO duplicate_sweeps DEFAULT VALUES RETURNING id, started_at
+`
+
+type StartDuplicateSweepRow struct {
+	ID        pgtype.UUID
+	StartedAt pgtype.Timestamptz
+}
+
+func (q *Queries) StartDuplicateSweep(ctx context.Context) (StartDuplicateSweepRow, error) {
+	row := q.db.QueryRow(ctx, startDuplicateSweep)
+	var i StartDuplicateSweepRow
+	err := row.Scan(&i.ID, &i.StartedAt)
+	return i, err
+}
+
+const upsertDuplicateGroup = `-- name: UpsertDuplicateGroup :one
+INSERT INTO duplicate_groups (member_key, tier, worst_bit_error_rate, last_seen_sweep_id)
+VALUES ($1, $2, $3, $4)
+ON CONFLICT (member_key) DO UPDATE
+   SET tier                 = EXCLUDED.tier,
+       worst_bit_error_rate = EXCLUDED.worst_bit_error_rate,
+       last_seen_sweep_id   = EXCLUDED.last_seen_sweep_id
+ WHERE duplicate_groups.status = 'pending'
+RETURNING id
+`
+
+type UpsertDuplicateGroupParams struct {
+	MemberKey         string
+	Tier              string
+	WorstBitErrorRate *float32
+	SweepID           pgtype.UUID
+}
+
+// Proposes a group, or refreshes one already pending. A group already dismissed
+// or merged is left exactly as it is: the WHERE on the update makes the conflict
+// a no-op, and the caller sees no row.
+func (q *Queries) UpsertDuplicateGroup(ctx context.Context, arg UpsertDuplicateGroupParams) (pgtype.UUID, error) {
+	row := q.db.QueryRow(ctx, upsertDuplicateGroup,
+		arg.MemberKey,
+		arg.Tier,
+		arg.WorstBitErrorRate,
+		arg.SweepID,
+	)
+	var id pgtype.UUID
+	err := row.Scan(&id)
+	return id, err
+}
@@ -0,0 +1,72 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: fingerprint_settings.sql
+
+package dbq
+
+import (
+	"context"
+)
+
+const getFingerprintSettings = `-- name: GetFingerprintSettings :one
+SELECT id, enabled, chromaprint_length_sec, acoustic_max_bit_error_rate, backfill_concurrency, sweep_interval_hours, updated_at FROM fingerprint_settings WHERE id = true
+`
+
+func (q *Queries) GetFingerprintSettings(ctx context.Context) (FingerprintSetting, error) {
+	row := q.db.QueryRow(ctx, getFingerprintSettings)
+	var i FingerprintSetting
+	err := row.Scan(
+		&i.ID,
+		&i.Enabled,
+		&i.ChromaprintLengthSec,
+		&i.AcousticMaxBitErrorRate,
+		&i.BackfillConcurrency,
+		&i.SweepIntervalHours,
+		&i.UpdatedAt,
+	)
+	return i, err
+}
+
+const updateFingerprintSettings = `-- name: UpdateFingerprintSettings :one
+UPDATE fingerprint_settings
+   SET enabled                     = $1,
+       chromaprint_length_sec      = $2,
+       acoustic_max_bit_error_rate = $3,
+       backfill_concurrency        = $4,
+       sweep_interval_hours        = $5,
+       updated_at                  = now()
+ WHERE id = true
+RETURNING id, enabled, chromaprint_length_sec, acoustic_max_bit_error_rate, backfill_concurrency, sweep_interval_hours, updated_at
+`
+
+type UpdateFingerprintSettingsParams struct {
+	Enabled                 bool
+	ChromaprintLengthSec    int32
+	AcousticMaxBitErrorRate float64
+	BackfillConcurrency     int32
+	SweepIntervalHours      int32
+}
+
+// Whole-row write from the admin card; migration 0061's CHECKs are the backstop
+// behind the service's own validation.
+func (q *Queries) UpdateFingerprintSettings(ctx context.Context, arg UpdateFingerprintSettingsParams) (FingerprintSetting, error) {
+	row := q.db.QueryRow(ctx, updateFingerprintSettings,
+		arg.Enabled,
+		arg.ChromaprintLengthSec,
+		arg.AcousticMaxBitErrorRate,
+		arg.BackfillConcurrency,
+		arg.SweepIntervalHours,
+	)
+	var i FingerprintSetting
+	err := row.Scan(
+		&i.ID,
+		&i.Enabled,
+		&i.ChromaprintLengthSec,
+		&i.AcousticMaxBitErrorRate,
+		&i.BackfillConcurrency,
+		&i.SweepIntervalHours,
+		&i.UpdatedAt,
+	)
+	return i, err
+}
@@ -0,0 +1,172 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: fingerprints.sql
+
+package dbq
+
+import (
+	"context"
+
+	"github.com/jackc/pgx/v5/pgtype"
+)
+
+const deleteTrackFingerprint = `-- name: DeleteTrackFingerprint :exec
+DELETE FROM track_fingerprints WHERE track_id = $1
+`
+
+// A file changed but could not be fingerprinted, for a reason unrelated to the
+// file. The stored row describes the OLD bytes, so it goes and the backfill
+// re-derives it — nothing may keep trusting a stale identity.
+func (q *Queries) DeleteTrackFingerprint(ctx context.Context, trackID pgtype.UUID) error {
+	_, err := q.db.Exec(ctx, deleteTrackFingerprint, trackID)
+	return err
+}
+
+const getFingerprintCoverage = `-- name: GetFingerprintCoverage :one
+SELECT count(*)::bigint AS total,
+       count(*) FILTER (
+           WHERE f.fingerprint_version >= $1
+             AND f.chromaprint_length_sec = $2
+             AND f.audio_stream_sha256 IS NOT NULL AND f.chromaprint IS NOT NULL
+       )::bigint AS fingerprinted,
+       count(*) FILTER (
+           WHERE f.fingerprint_version >= $1
+             AND f.chromaprint_length_sec = $2
+             AND (f.audio_stream_sha256 IS NULL OR f.chromaprint IS NULL)
+       )::bigint AS rejected,
+       count(*) FILTER (
+           WHERE f.track_id IS NULL
+              OR f.fingerprint_version < $1
+              OR f.chromaprint_length_sec <> $2
+       )::bigint AS pending
+  FROM tracks t
+  LEFT JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL
+`
+
+type GetFingerprintCoverageParams struct {
+	CurrentVersion       int16
+	ChromaprintLengthSec int32
+}
+
+type GetFingerprintCoverageRow struct {
+	Total         int64
+	Fingerprinted int64
+	Rejected      int64
+	Pending       int64
+}
+
+// The admin gauge for the backfill. fingerprinted + rejected + pending = total.
+// "Current" means derived by the current method AT the current length: a row at
+// another length is pending, because the backfill will re-derive it. rejected is
+// a current row with a NULL half: a tool ran and refused the file, which is
+// settled rather than waiting. Missing tracks are excluded, or the gauge could
+// never reach the end.
+func (q *Queries) GetFingerprintCoverage(ctx context.Context, arg GetFingerprintCoverageParams) (GetFingerprintCoverageRow, error) {
+	row := q.db.QueryRow(ctx, getFingerprintCoverage, arg.CurrentVersion, arg.ChromaprintLengthSec)
+	var i GetFingerprintCoverageRow
+	err := row.Scan(
+		&i.Total,
+		&i.Fingerprinted,
+		&i.Rejected,
+		&i.Pending,
+	)
+	return i, err
+}
+
+const listTracksNeedingFingerprint = `-- name: ListTracksNeedingFingerprint :many
+SELECT t.id, t.file_path
+  FROM tracks t
+  LEFT JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL
+   -- A row taken at another length is as stale as one from an older method:
+   -- chromaprints at two lengths cannot be compared (#3913).
+   AND (f.track_id IS NULL
+        OR f.fingerprint_version < $1
+        OR f.chromaprint_length_sec <> $2)
+   AND t.id > $3
+ ORDER BY t.id
+ LIMIT $4
+`
+
+type ListTracksNeedingFingerprintParams struct {
+	CurrentVersion       int16
+	ChromaprintLengthSec int32
+	AfterID              pgtype.UUID
+	BatchLimit           int32
+}
+
+type ListTracksNeedingFingerprintRow struct {
+	ID       pgtype.UUID
+	FilePath string
+}
+
+// The backfill's work queue (#3908): tracks with no fingerprint, or one derived
+// by an older method. Keyset-paged on id so a pass visits each track at most
+// once. That cursor is load-bearing: an inconclusive attempt writes no row, so
+// without it a file that keeps timing out would be listed again straight away
+// and retried in a tight loop. Missing tracks are skipped — there is no file to
+// read.
+func (q *Queries) ListTracksNeedingFingerprint(ctx context.Context, arg ListTracksNeedingFingerprintParams) ([]ListTracksNeedingFingerprintRow, error) {
+	rows, err := q.db.Query(ctx, listTracksNeedingFingerprint,
+		arg.CurrentVersion,
+		arg.ChromaprintLengthSec,
+		arg.AfterID,
+		arg.BatchLimit,
+	)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListTracksNeedingFingerprintRow
+	for rows.Next() {
+		var i ListTracksNeedingFingerprintRow
+		if err := rows.Scan(&i.ID, &i.FilePath); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const upsertTrackFingerprint = `-- name: UpsertTrackFingerprint :exec
+INSERT INTO track_fingerprints (
+    track_id, audio_stream_sha256, chromaprint, fingerprint_version, chromaprint_length_sec
+) VALUES (
+    $1, $2, $3,
+    $4, $5
+)
+ON CONFLICT (track_id) DO UPDATE SET
+    audio_stream_sha256    = EXCLUDED.audio_stream_sha256,
+    chromaprint            = EXCLUDED.chromaprint,
+    fingerprint_version    = EXCLUDED.fingerprint_version,
+    chromaprint_length_sec = EXCLUDED.chromaprint_length_sec,
+    computed_at            = now()
+`
+
+type UpsertTrackFingerprintParams struct {
+	TrackID              pgtype.UUID
+	AudioStreamSha256    []byte
+	Chromaprint          []int32
+	FingerprintVersion   int16
+	ChromaprintLengthSec int32
+}
+
+// Written whenever a track's fingerprint is derived: by the scan when a file is
+// new or its bytes changed, and by the backfill (#3908) for rows derived by an
+// older method. Replaces the row wholesale — a fingerprint of the old bytes has
+// no standing once the file has changed.
+func (q *Queries) UpsertTrackFingerprint(ctx context.Context, arg UpsertTrackFingerprintParams) error {
+	_, err := q.db.Exec(ctx, upsertTrackFingerprint,
+		arg.TrackID,
+		arg.AudioStreamSha256,
+		arg.Chromaprint,
+		arg.FingerprintVersion,
+		arg.ChromaprintLengthSec,
+	)
+	return err
+}
@@ -0,0 +1,339 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: merge.sql
+
+package dbq
+
+import (
+	"context"
+
+	"github.com/jackc/pgx/v5/pgtype"
+)
+
+const listDuplicateGroupMergeMembers = `-- name: ListDuplicateGroupMergeMembers :many
+SELECT t.id, t.file_path, t.file_format, t.file_size, t.added_at, t.album_id,
+       t.mbid, albums.mbid AS album_mbid
+  FROM duplicate_group_members m
+  JOIN tracks t  ON t.id = m.track_id
+  JOIN albums    ON albums.id = t.album_id
+ WHERE m.group_id = $1
+ ORDER BY t.id
+`
+
+type ListDuplicateGroupMergeMembersRow struct {
+	ID         pgtype.UUID
+	FilePath   string
+	FileFormat string
+	FileSize   int64
+	AddedAt    pgtype.Timestamptz
+	AlbumID    pgtype.UUID
+	Mbid       *string
+	AlbumMbid  *string
+}
+
+func (q *Queries) ListDuplicateGroupMergeMembers(ctx context.Context, groupID pgtype.UUID) ([]ListDuplicateGroupMergeMembersRow, error) {
+	rows, err := q.db.Query(ctx, listDuplicateGroupMergeMembers, groupID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []ListDuplicateGroupMergeMembersRow
+	for rows.Next() {
+		var i ListDuplicateGroupMergeMembersRow
+		if err := rows.Scan(
+			&i.ID,
+			&i.FilePath,
+			&i.FileFormat,
+			&i.FileSize,
+			&i.AddedAt,
+			&i.AlbumID,
+			&i.Mbid,
+			&i.AlbumMbid,
+		); err != nil {
+			return nil, err
+		}
+		items = append(items, i)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const lockDuplicateGroupForMerge = `-- name: LockDuplicateGroupForMerge :one
+
+SELECT id, tier, status
+  FROM duplicate_groups
+ WHERE id = $1
+   FOR UPDATE
+`
+
+type LockDuplicateGroupForMergeRow struct {
+	ID     pgtype.UUID
+	Tier   string
+	Status string
+}
+
+// Duplicate merge (Scribe #3911). Every statement here runs inside the one
+// transaction library.MergeDuplicateGroup opens, after the removed copy's file
+// is already gone. The loser's own track row is deleted last with DeleteTrack;
+// what these do is move everything it carries onto the survivor first, so that
+// delete's CASCADE finds nothing left to destroy.
+// Locks the group for the rest of the transaction, so two merges of one group
+// cannot run at once.
+func (q *Queries) LockDuplicateGroupForMerge(ctx context.Context, id pgtype.UUID) (LockDuplicateGroupForMergeRow, error) {
+	row := q.db.QueryRow(ctx, lockDuplicateGroupForMerge, id)
+	var i LockDuplicateGroupForMergeRow
+	err := row.Scan(&i.ID, &i.Tier, &i.Status)
+	return i, err
+}
+
+const markDuplicateGroupMerged = `-- name: MarkDuplicateGroupMerged :execrows
+UPDATE duplicate_groups
+   SET status = 'merged', resolved_at = now()
+ WHERE id = $1 AND status = 'pending'
+`
+
+func (q *Queries) MarkDuplicateGroupMerged(ctx context.Context, id pgtype.UUID) (int64, error) {
+	result, err := q.db.Exec(ctx, markDuplicateGroupMerged, id)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeCopyGeneralLikes = `-- name: MergeCopyGeneralLikes :many
+
+INSERT INTO general_likes (user_id, track_id, liked_at)
+SELECT user_id, $1::uuid, liked_at
+  FROM general_likes
+ WHERE track_id = $2::uuid
+ON CONFLICT (user_id, track_id) DO UPDATE
+   SET liked_at = LEAST(general_likes.liked_at, EXCLUDED.liked_at)
+RETURNING user_id
+`
+
+type MergeCopyGeneralLikesParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+// Collision-safe merges: a unique key includes track_id, so the survivor may
+// already hold a matching row. Copy what it lacks; DeleteTrack's CASCADE then
+// removes the loser's originals.
+// One like per user. A user who liked both copies keeps a single like, dated to
+// the earlier of the two.
+func (q *Queries) MergeCopyGeneralLikes(ctx context.Context, arg MergeCopyGeneralLikesParams) ([]pgtype.UUID, error) {
+	rows, err := q.db.Query(ctx, mergeCopyGeneralLikes, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []pgtype.UUID
+	for rows.Next() {
+		var user_id pgtype.UUID
+		if err := rows.Scan(&user_id); err != nil {
+			return nil, err
+		}
+		items = append(items, user_id)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const mergeCopyTrackSimilarity = `-- name: MergeCopyTrackSimilarity :execrows
+INSERT INTO track_similarity (track_a_id, track_b_id, score, source, fetched_at)
+SELECT CASE WHEN track_a_id = $1::uuid THEN $2::uuid ELSE track_a_id END,
+       CASE WHEN track_b_id = $1::uuid THEN $2::uuid ELSE track_b_id END,
+       score, source, fetched_at
+  FROM track_similarity
+ WHERE (track_a_id = $1::uuid OR track_b_id = $1::uuid)
+   AND (CASE WHEN track_a_id = $1::uuid THEN $2::uuid ELSE track_a_id END)
+    <> (CASE WHEN track_b_id = $1::uuid THEN $2::uuid ELSE track_b_id END)
+ON CONFLICT (track_a_id, track_b_id, source) DO NOTHING
+`
+
+type MergeCopyTrackSimilarityParams struct {
+	LoserID    pgtype.UUID
+	SurvivorID pgtype.UUID
+}
+
+// Rewrites the loser to the survivor on either side of an edge. An edge between
+// the two copies would become a track similar to itself — the table forbids
+// that, and it means nothing — so it is dropped. An edge the survivor already
+// has from the same source is kept as it is.
+func (q *Queries) MergeCopyTrackSimilarity(ctx context.Context, arg MergeCopyTrackSimilarityParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeCopyTrackSimilarity, arg.LoserID, arg.SurvivorID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeCopyTrackTags = `-- name: MergeCopyTrackTags :execrows
+INSERT INTO track_tags (track_id, tag, weight)
+SELECT $1::uuid, tag, weight
+  FROM track_tags
+ WHERE track_id = $2::uuid
+ON CONFLICT (track_id, tag) DO NOTHING
+`
+
+type MergeCopyTrackTagsParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+func (q *Queries) MergeCopyTrackTags(ctx context.Context, arg MergeCopyTrackTagsParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeCopyTrackTags, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeInheritTrackMbid = `-- name: MergeInheritTrackMbid :exec
+UPDATE tracks AS survivor
+   SET mbid = loser.mbid
+  FROM tracks AS loser
+ WHERE survivor.id = $1::uuid
+   AND loser.id = $2::uuid
+   AND survivor.mbid IS NULL
+   AND loser.mbid IS NOT NULL
+`
+
+type MergeInheritTrackMbidParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+// A recording MBID is what the similarity pipeline keys on. If only the removed
+// copy carried one, the survivor takes it rather than going dark to similarity.
+func (q *Queries) MergeInheritTrackMbid(ctx context.Context, arg MergeInheritTrackMbidParams) error {
+	_, err := q.db.Exec(ctx, mergeInheritTrackMbid, arg.SurvivorID, arg.LoserID)
+	return err
+}
+
+const mergeRepointContextualLikes = `-- name: MergeRepointContextualLikes :execrows
+UPDATE contextual_likes SET track_id = $1::uuid WHERE track_id = $2::uuid
+`
+
+type MergeRepointContextualLikesParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+func (q *Queries) MergeRepointContextualLikes(ctx context.Context, arg MergeRepointContextualLikesParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeRepointContextualLikes, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeRepointLidarrRequests = `-- name: MergeRepointLidarrRequests :execrows
+UPDATE lidarr_requests SET matched_track_id = $1::uuid
+ WHERE matched_track_id = $2::uuid
+`
+
+type MergeRepointLidarrRequestsParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+func (q *Queries) MergeRepointLidarrRequests(ctx context.Context, arg MergeRepointLidarrRequestsParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeRepointLidarrRequests, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeRepointPlayEvents = `-- name: MergeRepointPlayEvents :execrows
+
+UPDATE play_events SET track_id = $1::uuid WHERE track_id = $2::uuid
+`
+
+type MergeRepointPlayEventsParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+// Plain repoints: no unique key involves track_id, so moving rows cannot collide.
+func (q *Queries) MergeRepointPlayEvents(ctx context.Context, arg MergeRepointPlayEventsParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeRepointPlayEvents, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeRepointPlaybackErrors = `-- name: MergeRepointPlaybackErrors :execrows
+UPDATE playback_errors SET track_id = $1::uuid WHERE track_id = $2::uuid
+`
+
+type MergeRepointPlaybackErrorsParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+func (q *Queries) MergeRepointPlaybackErrors(ctx context.Context, arg MergeRepointPlaybackErrorsParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeRepointPlaybackErrors, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
+
+const mergeRepointPlaylistTracks = `-- name: MergeRepointPlaylistTracks :many
+UPDATE playlist_tracks SET track_id = $1::uuid
+ WHERE track_id = $2::uuid
+RETURNING playlist_id
+`
+
+type MergeRepointPlaylistTracksParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+// playlist_tracks is keyed by (playlist_id, position), so repointing keeps every
+// entry exactly where it was. A playlist that held both copies simply holds the
+// survivor twice — the user put two entries there, and both stay.
+func (q *Queries) MergeRepointPlaylistTracks(ctx context.Context, arg MergeRepointPlaylistTracksParams) ([]pgtype.UUID, error) {
+	rows, err := q.db.Query(ctx, mergeRepointPlaylistTracks, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var items []pgtype.UUID
+	for rows.Next() {
+		var playlist_id pgtype.UUID
+		if err := rows.Scan(&playlist_id); err != nil {
+			return nil, err
+		}
+		items = append(items, playlist_id)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return items, nil
+}
+
+const mergeRepointSkipEvents = `-- name: MergeRepointSkipEvents :execrows
+UPDATE skip_events SET track_id = $1::uuid WHERE track_id = $2::uuid
+`
+
+type MergeRepointSkipEventsParams struct {
+	SurvivorID pgtype.UUID
+	LoserID    pgtype.UUID
+}
+
+func (q *Queries) MergeRepointSkipEvents(ctx context.Context, arg MergeRepointSkipEventsParams) (int64, error) {
+	result, err := q.db.Exec(ctx, mergeRepointSkipEvents, arg.SurvivorID, arg.LoserID)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected(), nil
+}
@@ -297,6 +297,42 @@ type DiscoverTuning struct {
 	UpdatedAt        pgtype.Timestamptz
 }

+type DuplicateGroup struct {
+	ID                pgtype.UUID
+	MemberKey         string
+	Tier              string
+	WorstBitErrorRate *float32
+	Status            string
+	DetectedAt        pgtype.Timestamptz
+	LastSeenSweepID   pgtype.UUID
+	ResolvedAt        pgtype.Timestamptz
+}
+
+type DuplicateGroupMember struct {
+	GroupID pgtype.UUID
+	TrackID pgtype.UUID
+}
+
+type DuplicateSweep struct {
+	ID               pgtype.UUID
+	StartedAt        pgtype.Timestamptz
+	FinishedAt       pgtype.Timestamptz
+	Candidates       *int32
+	GroupsFound      *int32
+	OversizeClusters *int32
+	ErrorMessage     *string
+}
+
+type FingerprintSetting struct {
+	ID                      bool
+	Enabled                 bool
+	ChromaprintLengthSec    int32
+	AcousticMaxBitErrorRate float64
+	BackfillConcurrency     int32
+	SweepIntervalHours      int32
+	UpdatedAt               pgtype.Timestamptz
+}
+
 type GeneralLike struct {
 	UserID  pgtype.UUID
 	TrackID pgtype.UUID
@@ -667,6 +703,15 @@ type Track struct {
 	MissingSince      pgtype.Timestamptz
 }

+type TrackFingerprint struct {
+	TrackID              pgtype.UUID
+	AudioStreamSha256    []byte
+	Chromaprint          []int32
+	FingerprintVersion   int16
+	ComputedAt           pgtype.Timestamptz
+	ChromaprintLengthSec int32
+}
+
 type TrackSimilarity struct {
 	TrackAID  pgtype.UUID
 	TrackBID  pgtype.UUID
@@ -829,7 +829,7 @@ similar_artists AS (
    JOIN seed_artist sa ON asim.artist_a_id = sa.artist_id
    WHERE asim.source = 'listenbrainz'
      AND t.id NOT IN (SELECT id FROM excluded_ids)
-    ORDER BY asim.score DESC, random()
+    ORDER BY asim.score DESC, md5(t.id::text || $12::text)
    LIMIT $6
 ),
 tag_overlap AS (
@@ -857,7 +857,7 @@ likes_overlap AS (
          WHERE t.id = gl.track_id
            AND trim(g_overlap.g) IN (SELECT tag FROM seed_tags)
      )
-    ORDER BY random()
+    ORDER BY md5(gl.track_id::text || $12::text)
    LIMIT $8
 ),
 taste_overlap AS (
@@ -884,7 +884,7 @@ coplay_artists AS (
    WHERE asim.source = 'user_cooccurrence'
      AND t.id NOT IN (SELECT id FROM excluded_ids)
      AND t.id <> $2
-    ORDER BY asim.score DESC, random()
+    ORDER BY asim.score DESC, md5(t.id::text || $12::text)
    LIMIT $11
 ),
 random_fill AS (
@@ -900,7 +900,7 @@ random_fill AS (
          UNION SELECT track_id FROM taste_overlap
          UNION SELECT track_id FROM coplay_artists
      )
-    ORDER BY random()
+    ORDER BY md5(t.id::text || $12::text)
    LIMIT $9
 )
 SELECT
@@ -938,17 +938,18 @@ GROUP BY t.id, t.title, t.album_id, t.artist_id, t.duration_ms, t.file_path,
 `

 type LoadRadioCandidatesV2Params struct {
-	UserID  pgtype.UUID
-	ID      pgtype.UUID
-	Column3 interface{}
-	Column4 []pgtype.UUID
-	Limit   int32
-	Limit_2 int32
-	Limit_3 int32
-	Limit_4 int32
-	Limit_5 int32
-	Limit_6 int32
-	Limit_7 int32
+	UserID   pgtype.UUID
+	ID       pgtype.UUID
+	Column3  interface{}
+	Column4  []pgtype.UUID
+	Limit    int32
+	Limit_2  int32
+	Limit_3  int32
+	Limit_4  int32
+	Limit_5  int32
+	Limit_6  int32
+	Limit_7  int32
+	Column12 string
 }

 type LoadRadioCandidatesV2Row struct {
@@ -971,8 +972,22 @@ type LoadRadioCandidatesV2Row struct {
 //	enter the pool even when the similarity/random arms miss them; scored
 //	in Go via TasteMatch, so sim_score here is 0 pool-inclusion),
 //	$11 coplay_artists K (#1533 — tracks by artists co-played across the
-//	instance with the seed's artist; source='user_cooccurrence').
+//	instance with the seed's artist; source='user_cooccurrence'),
+//	$12 order_seed (text) — see below.
 //
+// $12 REPLACES `ORDER BY random()` IN FOUR ARMS (#3889). Those arms returned
+// a stable set only while their LIMIT exceeded the rows eligible for them: at
+// that point they returned all of them and the order stopped mattering,
+// because the caller sorts by track id before scoring. Below that threshold
+// they returned a random SUBSET, and two builds on the same day drew
+// different ones — so "daily determinism" held by accident, and only for
+// libraries smaller than the limits.
+//
+// md5(id || seed) keeps the intent — an arbitrary spread that changes when
+// the seed does — while making it reproducible for a given seed. The CALLER
+// decides what that means: system mixes pass a per-(user, day) string and get
+// the determinism they promise; radio passes a fresh value per request and
+// keeps varying, which is what a radio should do.
 // Returns same shape as LoadRadioCandidates plus similarity_score column.
 func (q *Queries) LoadRadioCandidatesV2(ctx context.Context, arg LoadRadioCandidatesV2Params) ([]LoadRadioCandidatesV2Row, error) {
 	rows, err := q.db.Query(ctx, loadRadioCandidatesV2,
@@ -987,6 +1002,7 @@ func (q *Queries) LoadRadioCandidatesV2(ctx context.Context, arg LoadRadioCandid
 		arg.Limit_5,
 		arg.Limit_6,
 		arg.Limit_7,
+		arg.Column12,
 	)
 	if err != nil {
 		return nil, err
@@ -148,39 +148,36 @@ func (q *Queries) DeleteTrack(ctx context.Context, id pgtype.UUID) (DeleteTrackR
 	return i, err
 }

-const findMissingTrackByFingerprint = `-- name: FindMissingTrackByFingerprint :many
-SELECT id, file_path FROM tracks
- WHERE missing_since IS NOT NULL
-   AND file_size = $1
-   AND duration_ms = $2
+const findMissingTrackByAudioHash = `-- name: FindMissingTrackByAudioHash :many
+SELECT t.id, t.file_path
+  FROM tracks t
+  JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NOT NULL
+   AND f.audio_stream_sha256 = $1
 LIMIT 2
 `

-type FindMissingTrackByFingerprintParams struct {
-	FileSize   int64
-	DurationMs int32
-}
-
-type FindMissingTrackByFingerprintRow struct {
+type FindMissingTrackByAudioHashRow struct {
 	ID       pgtype.UUID
 	FilePath string
 }

-// Move detection fallback for files with no MBID (#2528). Exact byte size AND
-// exact decoded duration is a strong pair: a plain move or rename preserves
-// both, while a re-encode changes at least one — and a re-encode genuinely is a
-// different file, so failing to match there is correct rather than a gap.
+// Move detection fallback for files with no MBID (#2528, #3914). The audio stream
+// hash identifies the encoded audio itself, so it survives a rename, a move and a
+// retag — anything short of a re-encode. It replaced (file_size, duration_ms),
+// which could pair two unrelated files that happened to share a byte count and a
+// duration, and missed a file retagged in place, whose size changes.
 //
 // Same missing-only constraint and same LIMIT 2 rationale as the MBID variant.
-func (q *Queries) FindMissingTrackByFingerprint(ctx context.Context, arg FindMissingTrackByFingerprintParams) ([]FindMissingTrackByFingerprintRow, error) {
-	rows, err := q.db.Query(ctx, findMissingTrackByFingerprint, arg.FileSize, arg.DurationMs)
+func (q *Queries) FindMissingTrackByAudioHash(ctx context.Context, audioStreamSha256 []byte) ([]FindMissingTrackByAudioHashRow, error) {
+	rows, err := q.db.Query(ctx, findMissingTrackByAudioHash, audioStreamSha256)
 	if err != nil {
 		return nil, err
 	}
 	defer rows.Close()
-	var items []FindMissingTrackByFingerprintRow
+	var items []FindMissingTrackByAudioHashRow
 	for rows.Next() {
-		var i FindMissingTrackByFingerprintRow
+		var i FindMissingTrackByAudioHashRow
 		if err := rows.Scan(&i.ID, &i.FilePath); err != nil {
 			return nil, err
 		}
@@ -0,0 +1,16 @@
+-- Drop the rows the narrower constraints are about to forbid, or re-adding
+-- them fails against existing data (the 0051 down-migration pattern).
+DELETE FROM recommendation_weight_profiles WHERE profile = 'songs_like';
+DELETE FROM recommendation_tuning_audit WHERE scope = 'songs_like';
+
+ALTER TABLE recommendation_tuning_audit
+    DROP CONSTRAINT recommendation_tuning_audit_scope_check;
+ALTER TABLE recommendation_tuning_audit
+    ADD CONSTRAINT recommendation_tuning_audit_scope_check
+    CHECK (scope IN ('radio', 'daily_mix', 'taste', 'discover'));
+
+ALTER TABLE recommendation_weight_profiles
+    DROP CONSTRAINT recommendation_weight_profiles_profile_check;
+ALTER TABLE recommendation_weight_profiles
+    ADD CONSTRAINT recommendation_weight_profiles_profile_check
+    CHECK (profile IN ('radio', 'daily_mix'));
@@ -0,0 +1,37 @@
+-- 0057_songs_like_tuning.up.sql — a THIRD weight profile, for Songs-like
+-- (Scribe #3881, milestone #398).
+--
+-- Songs-like shared the `daily_mix` profile with For-You, and that is the bug.
+-- The two surfaces want opposite things: For-You is a broad "what will they
+-- enjoy today", Songs-like answers "what sounds like THIS", and under one set
+-- of weights the broad answer wins. Operator, 2026-09-10: "I'm expecting to
+-- get a consistent sound and style from the experience... I was getting a
+-- seeming wide variety of music from each one when I was hoping to stay in a
+-- certain neighborhood."
+--
+-- Under the shared daily_mix weights, an UNRELATED track the user had liked
+-- and not played recently scored 1.0 + 2.0 + 1.0 = 4.0 before taste, while a
+-- PERFECT similarity match they had not liked scored 1.0 + 1.5 = 2.5. Liking
+-- something outranked sounding like the seed. Splitting the profile is what
+-- lets similarity dominate here without making For-You narrow.
+--
+-- Rows are seeded by the recsettings boot reconcile, not here, so shipped
+-- defaults live in exactly one place (Go) — same as 0040.
+
+-- Rule #36: a new value for a CHECK-gated column needs the constraint
+-- rewritten in the SAME change, or the first row written under the new
+-- profile fails at runtime rather than at migrate time.
+ALTER TABLE recommendation_weight_profiles
+    DROP CONSTRAINT recommendation_weight_profiles_profile_check;
+ALTER TABLE recommendation_weight_profiles
+    ADD CONSTRAINT recommendation_weight_profiles_profile_check
+    CHECK (profile IN ('radio', 'daily_mix', 'songs_like'));
+
+-- The audit table gates the same name on a separate constraint. Missing this
+-- one would let the profile be seeded and then fail on the first knob turn —
+-- green at boot, 500 on first use.
+ALTER TABLE recommendation_tuning_audit
+    DROP CONSTRAINT recommendation_tuning_audit_scope_check;
+ALTER TABLE recommendation_tuning_audit
+    ADD CONSTRAINT recommendation_tuning_audit_scope_check
+    CHECK (scope IN ('radio', 'daily_mix', 'taste', 'discover', 'songs_like'));
@@ -0,0 +1 @@
+DROP TABLE track_fingerprints;
@@ -0,0 +1,38 @@
+-- 0058_track_fingerprints.up.sql — an acoustic identity per track (Scribe
+-- milestone #400: #3905, #3906).
+--
+-- A table of its own rather than columns on tracks, for the hot path's sake:
+-- tracks is read with SELECT * by eight queries, among them ListTracksByAlbum,
+-- SearchTracks and GetTracksByIDs — album pages, search, the Subsonic surface.
+-- A ~4 KB chromaprint column on tracks would be de-TOASTed on every one of
+-- those reads to carry a value only the duplicate sweep ever looks at.
+--
+-- What a row means, which the backfill depends on:
+--   no row                          never fingerprinted
+--   fingerprint_version < current   derived by an older method; re-derive it
+--   fingerprint_version = current   attempted; a NULL value means that tool
+--                                   failed on this file, and it is not retried
+--                                   until the file changes
+-- A failure that says nothing about the file — a timeout, a cancelled scan, a
+-- missing binary — writes no row at all, so the backfill tries again.
+CREATE TABLE track_fingerprints (
+    -- CASCADE is right here, unlike for the likes and play history M400's
+    -- merge has to carry across: a fingerprint describes one file's bytes and
+    -- means nothing once that file's row is gone.
+    track_id            uuid        PRIMARY KEY REFERENCES tracks (id) ON DELETE CASCADE,
+    -- SHA-256 of the ENCODED audio packets (ffmpeg -c:a copy -f hash), not of
+    -- decoded samples. internal/library/fingerprint.go says why.
+    audio_stream_sha256 bytea
+        CHECK (audio_stream_sha256 IS NULL OR octet_length(audio_stream_sha256) = 32),
+    -- fpcalc -raw -signed: the same 32 bits per item, stored signed because
+    -- integer is.
+    chromaprint         integer[],
+    fingerprint_version smallint    NOT NULL,
+    computed_at         timestamptz NOT NULL DEFAULT now()
+);
+
+-- The exact duplicate tier is an equality match on this column. Partial
+-- because a NULL is never looked up — it only means the hash was not taken.
+CREATE INDEX track_fingerprints_audio_stream_sha256
+    ON track_fingerprints (audio_stream_sha256)
+    WHERE audio_stream_sha256 IS NOT NULL;
@@ -0,0 +1,4 @@
+DROP INDEX IF EXISTS tracks_duration_id_idx;
+DROP TABLE duplicate_group_members;
+DROP TABLE duplicate_groups;
+DROP TABLE duplicate_sweeps;
@@ -0,0 +1,54 @@
+-- 0059_duplicate_groups.up.sql — proposed duplicates and the sweeps that find
+-- them (Scribe milestone #400: #3910).
+--
+-- The sweep compares fingerprints (track_fingerprints, 0058) and proposes groups
+-- of tracks that hold one recording. Nothing here merges anything: a group is a
+-- proposal the operator reviews, and the merge (#3911) is a separate act.
+
+-- One row per sweep. Lets the report tell "the sweep has never run" apart from
+-- "it ran and found nothing", and gives the in-flight guard something to check,
+-- the same way scan_runs does for the library scan.
+CREATE TABLE duplicate_sweeps (
+    id                uuid        PRIMARY KEY DEFAULT gen_random_uuid(),
+    started_at        timestamptz NOT NULL DEFAULT now(),
+    finished_at       timestamptz,
+    candidates        integer,
+    groups_found      integer,
+    oversize_clusters integer,
+    error_message     text
+);
+CREATE INDEX duplicate_sweeps_started_at_idx ON duplicate_sweeps (started_at DESC);
+
+CREATE TABLE duplicate_groups (
+    id                   uuid        PRIMARY KEY DEFAULT gen_random_uuid(),
+    -- The group's identity: its member track ids, sorted and joined. A sweep
+    -- that finds the same tracks again updates this row rather than proposing
+    -- them twice, and a dismissal stays attached to the set it was made about.
+    member_key           text        NOT NULL UNIQUE,
+    -- Rule 36: a new value for either CHECK swaps the constraint in the same
+    -- migration.
+    tier                 text        NOT NULL CHECK (tier IN ('exact', 'acoustic')),
+    -- Largest disagreement between any two members; NULL for exact groups,
+    -- which have no score.
+    worst_bit_error_rate real,
+    status               text        NOT NULL DEFAULT 'pending'
+        CHECK (status IN ('pending', 'dismissed', 'merged')),
+    detected_at          timestamptz NOT NULL DEFAULT now(),
+    last_seen_sweep_id   uuid        REFERENCES duplicate_sweeps (id) ON DELETE SET NULL,
+    resolved_at          timestamptz
+);
+CREATE INDEX duplicate_groups_status_idx ON duplicate_groups (status);
+
+CREATE TABLE duplicate_group_members (
+    group_id uuid NOT NULL REFERENCES duplicate_groups (id) ON DELETE CASCADE,
+    -- CASCADE is right here: a track that genuinely leaves the library has no
+    -- place in a proposal about its duplicates.
+    track_id uuid NOT NULL REFERENCES tracks (id) ON DELETE CASCADE,
+    PRIMARY KEY (group_id, track_id)
+);
+CREATE INDEX duplicate_group_members_track_idx ON duplicate_group_members (track_id);
+
+-- The sweep streams candidates in (duration_ms, id) order, keyset-paged, so it
+-- only ever holds a few seconds' worth of durations in memory. Without this each
+-- page would sort the whole library again.
+CREATE INDEX tracks_duration_id_idx ON tracks (duration_ms, id);
@@ -0,0 +1 @@
+DROP INDEX IF EXISTS play_events_track_idx;
@@ -0,0 +1,8 @@
+-- 0060_play_events_track_index.up.sql — play_events by track (Scribe #3912, #3911).
+--
+-- play_events is indexed by (user_id, started_at) and (user_id, track_id), both
+-- led by user. Nothing reached it by track alone until the duplicates report,
+-- which shows each copy's play count — a scan of the whole table per copy — and
+-- the merge (#3911), which repoints a duplicate's play history onto the copy
+-- being kept. Both ask "every play of this track", whoever played it.
+CREATE INDEX play_events_track_idx ON play_events (track_id);
@@ -0,0 +1,2 @@
+ALTER TABLE track_fingerprints DROP COLUMN chromaprint_length_sec;
+DROP TABLE fingerprint_settings;
@@ -0,0 +1,50 @@
+-- 0061_fingerprint_settings.up.sql — fingerprinting's knobs, in admin Settings
+-- (Scribe #3913, milestone #400). Rule 25: anything an operator might tune is a
+-- database row, changed without a restart. Singleton in the style of
+-- reacquisition_settings (0056).
+CREATE TABLE fingerprint_settings (
+    id                          boolean PRIMARY KEY DEFAULT true,
+
+    -- Fingerprinting new files, the backfill, and the duplicate sweep. Off stops
+    -- the decode work entirely — the reason to turn it off is a slow NAS, and
+    -- that is the operator's call. On by default: a library that cannot tell its
+    -- duplicates apart is what milestone #400 exists to end.
+    enabled                     boolean NOT NULL DEFAULT true,
+
+    -- Seconds of audio fpcalc fingerprints. Chromaprints taken at different
+    -- lengths cannot be compared, which is why track_fingerprints records the
+    -- length each row was taken at (below): change this and every chromaprint is
+    -- re-derived, and until then only rows at the new length are compared.
+    chromaprint_length_sec      integer NOT NULL DEFAULT 120,
+
+    -- The most disagreement two aligned fingerprints may show and still be
+    -- proposed as one recording. Unrelated audio sits near 0.5, so the ceiling
+    -- stays well clear of it.
+    acoustic_max_bit_error_rate double precision NOT NULL DEFAULT 0.15,
+
+    -- Files the backfill decodes at once. Decoding competes with playback
+    -- transcoding for CPU and with streaming for the mount.
+    backfill_concurrency        integer NOT NULL DEFAULT 2,
+
+    -- The least time between duplicate sweeps. A sweep still runs only when
+    -- fingerprints have changed since the last one.
+    sweep_interval_hours        integer NOT NULL DEFAULT 1,
+
+    -- When the settings were last saved. A new threshold or length can change
+    -- what a sweep finds, so a save makes a sweep due.
+    updated_at                  timestamptz NOT NULL DEFAULT now(),
+
+    CONSTRAINT fingerprint_settings_singleton CHECK (id = true),
+    CONSTRAINT fingerprint_settings_length_range
+        CHECK (chromaprint_length_sec >= 30 AND chromaprint_length_sec <= 600),
+    CONSTRAINT fingerprint_settings_threshold_range
+        CHECK (acoustic_max_bit_error_rate >= 0.01 AND acoustic_max_bit_error_rate <= 0.35),
+    CONSTRAINT fingerprint_settings_concurrency_range
+        CHECK (backfill_concurrency >= 1 AND backfill_concurrency <= 8),
+    CONSTRAINT fingerprint_settings_sweep_interval_range
+        CHECK (sweep_interval_hours >= 1 AND sweep_interval_hours <= 168)
+);
+INSERT INTO fingerprint_settings (id) VALUES (true) ON CONFLICT (id) DO NOTHING;
+
+-- Every row written so far was taken at fpcalc's default length.
+ALTER TABLE track_fingerprints ADD COLUMN chromaprint_length_sec integer NOT NULL DEFAULT 120;
@@ -0,0 +1,156 @@
+-- name: StartDuplicateSweep :one
+INSERT INTO duplicate_sweeps DEFAULT VALUES RETURNING id, started_at;
+
+-- name: FinishDuplicateSweep :exec
+UPDATE duplicate_sweeps
+   SET finished_at       = now(),
+       candidates        = sqlc.arg(candidates),
+       groups_found      = sqlc.arg(groups_found),
+       oversize_clusters = sqlc.arg(oversize_clusters),
+       error_message     = NULLIF(sqlc.arg(error_message)::text, '')
+ WHERE id = sqlc.arg(id);
+
+-- name: GetInFlightDuplicateSweep :one
+-- The guard against two sweeps at once: "in flight" is finished_at IS NULL.
+SELECT id, started_at
+  FROM duplicate_sweeps
+ WHERE finished_at IS NULL
+ ORDER BY started_at DESC
+ LIMIT 1;
+
+-- name: GetLatestDuplicateSweep :one
+SELECT id, started_at, finished_at, candidates, groups_found, oversize_clusters, error_message
+  FROM duplicate_sweeps
+ ORDER BY started_at DESC
+ LIMIT 1;
+
+-- name: GetLatestFingerprintComputedAt :one
+-- Whether a sweep has anything new to look at: fingerprints written since the
+-- last sweep started.
+SELECT max(computed_at)::timestamptz AS latest FROM track_fingerprints;
+
+-- name: ListExactDuplicateHashes :many
+-- The exact tier, library-wide in one pass: identical encoded audio shared by
+-- more than one present track.
+SELECT f.audio_stream_sha256,
+       array_agg(t.id ORDER BY t.id)::uuid[] AS track_ids
+  FROM track_fingerprints f
+  JOIN tracks t ON t.id = f.track_id
+ WHERE t.missing_since IS NULL
+   AND f.fingerprint_version >= sqlc.arg(current_version)
+   AND f.audio_stream_sha256 IS NOT NULL
+ GROUP BY f.audio_stream_sha256
+HAVING count(*) > 1;
+
+-- name: ListDuplicateCandidates :many
+-- The acoustic tier's input, one page at a time in (duration_ms, id) order so the
+-- sweep holds only a sliding window of durations. Tracks without a chromaprint
+-- cannot be compared acoustically and are left out; any exact duplicates among
+-- them come from ListExactDuplicateHashes.
+SELECT t.id, t.duration_ms, f.audio_stream_sha256, f.chromaprint
+  FROM tracks t
+  JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL
+   AND f.fingerprint_version >= sqlc.arg(current_version)
+   AND f.chromaprint IS NOT NULL
+   -- Only chromaprints taken at the current length: prints at two lengths are not
+   -- comparable, and after a length change the backfill is still re-deriving the
+   -- rest (#3913).
+   AND f.chromaprint_length_sec = sqlc.arg(chromaprint_length_sec)
+   AND (t.duration_ms, t.id) > (sqlc.arg(after_duration_ms)::integer, sqlc.arg(after_id)::uuid)
+ ORDER BY t.duration_ms, t.id
+ LIMIT sqlc.arg(page_limit);
+
+-- name: ListDismissedDuplicateMemberSets :many
+-- What the operator has already said are not duplicates. A new proposal whose
+-- every member sat together in one of these is not proposed again.
+SELECT g.id, array_agg(m.track_id ORDER BY m.track_id)::uuid[] AS track_ids
+  FROM duplicate_groups g
+  JOIN duplicate_group_members m ON m.group_id = g.id
+ WHERE g.status = 'dismissed'
+ GROUP BY g.id;
+
+-- name: UpsertDuplicateGroup :one
+-- Proposes a group, or refreshes one already pending. A group already dismissed
+-- or merged is left exactly as it is: the WHERE on the update makes the conflict
+-- a no-op, and the caller sees no row.
+INSERT INTO duplicate_groups (member_key, tier, worst_bit_error_rate, last_seen_sweep_id)
+VALUES (sqlc.arg(member_key), sqlc.arg(tier), sqlc.narg(worst_bit_error_rate), sqlc.arg(sweep_id))
+ON CONFLICT (member_key) DO UPDATE
+   SET tier                 = EXCLUDED.tier,
+       worst_bit_error_rate = EXCLUDED.worst_bit_error_rate,
+       last_seen_sweep_id   = EXCLUDED.last_seen_sweep_id
+ WHERE duplicate_groups.status = 'pending'
+RETURNING id;
+
+-- name: AddDuplicateGroupMember :exec
+INSERT INTO duplicate_group_members (group_id, track_id)
+VALUES (sqlc.arg(group_id), sqlc.arg(track_id))
+ON CONFLICT DO NOTHING;
+
+-- name: DeleteStalePendingDuplicateGroups :execrows
+-- A pending proposal this sweep did not find again no longer describes the
+-- library: a member was re-fingerprinted, merged away or went missing. Dismissed
+-- groups are kept regardless — they are the memory of a decision.
+--
+-- Only proposals last confirmed by an EARLIER sweep go. Should two sweeps ever
+-- overlap (a manual trigger racing the worker), neither may delete what the other
+-- has just found.
+DELETE FROM duplicate_groups g
+ WHERE g.status = 'pending'
+   AND g.last_seen_sweep_id IS DISTINCT FROM sqlc.arg(sweep_id)
+   AND (g.last_seen_sweep_id IS NULL
+        OR (SELECT s.started_at FROM duplicate_sweeps s WHERE s.id = g.last_seen_sweep_id)
+           < (SELECT s.started_at FROM duplicate_sweeps s WHERE s.id = sqlc.arg(sweep_id)));
+
+-- name: CountPendingDuplicateGroups :one
+-- Proposals awaiting review. A group left with one member — its other tracks
+-- deleted since the sweep — is no proposal at all and is not counted; the next
+-- sweep retires it.
+SELECT count(*)::bigint
+  FROM duplicate_groups g
+ WHERE g.status = 'pending'
+   AND (SELECT count(*) FROM duplicate_group_members m WHERE m.group_id = g.id) >= 2;
+
+-- name: ListPendingDuplicateGroupMembers :many
+-- One page of proposals, newest first, flattened to one row per member so the
+-- handler folds them without a query per group. What each copy carries — likes
+-- and plays from every user — is here because it is what the operator weighs
+-- when deciding which copy to keep.
+WITH page AS (
+    SELECT g.id, g.tier, g.worst_bit_error_rate, g.detected_at
+      FROM duplicate_groups g
+     WHERE g.status = 'pending'
+       AND (SELECT count(*) FROM duplicate_group_members m WHERE m.group_id = g.id) >= 2
+     ORDER BY g.detected_at DESC, g.id
+     LIMIT sqlc.arg(page_limit) OFFSET sqlc.arg(page_offset)
+)
+SELECT p.id                  AS group_id,
+       p.tier,
+       p.worst_bit_error_rate,
+       p.detected_at,
+       t.id                  AS track_id,
+       t.title,
+       artists.name          AS artist_name,
+       albums.id             AS album_id,
+       albums.title          AS album_title,
+       t.file_path,
+       t.file_format,
+       t.file_size,
+       t.duration_ms,
+       t.added_at,
+       (SELECT count(*) FROM general_likes l WHERE l.track_id = t.id)::bigint AS like_count,
+       (SELECT count(*) FROM play_events e WHERE e.track_id = t.id)::bigint   AS play_count
+  FROM page p
+  JOIN duplicate_group_members m ON m.group_id = p.id
+  JOIN tracks t  ON t.id = m.track_id
+  JOIN albums    ON albums.id  = t.album_id
+  JOIN artists   ON artists.id = t.artist_id
+ ORDER BY p.detected_at DESC, p.id, t.id;
+
+-- name: DismissDuplicateGroup :execrows
+-- "These are not duplicates." Only a pending group can be dismissed; zero rows
+-- means it was already resolved or no longer exists.
+UPDATE duplicate_groups
+   SET status = 'dismissed', resolved_at = now()
+ WHERE id = sqlc.arg(id) AND status = 'pending';
@@ -0,0 +1,15 @@
+-- name: GetFingerprintSettings :one
+SELECT * FROM fingerprint_settings WHERE id = true;
+
+-- name: UpdateFingerprintSettings :one
+-- Whole-row write from the admin card; migration 0061's CHECKs are the backstop
+-- behind the service's own validation.
+UPDATE fingerprint_settings
+   SET enabled                     = sqlc.arg(enabled),
+       chromaprint_length_sec      = sqlc.arg(chromaprint_length_sec),
+       acoustic_max_bit_error_rate = sqlc.arg(acoustic_max_bit_error_rate),
+       backfill_concurrency        = sqlc.arg(backfill_concurrency),
+       sweep_interval_hours        = sqlc.arg(sweep_interval_hours),
+       updated_at                  = now()
+ WHERE id = true
+RETURNING *;
@@ -0,0 +1,70 @@
+-- name: UpsertTrackFingerprint :exec
+-- Written whenever a track's fingerprint is derived: by the scan when a file is
+-- new or its bytes changed, and by the backfill (#3908) for rows derived by an
+-- older method. Replaces the row wholesale — a fingerprint of the old bytes has
+-- no standing once the file has changed.
+INSERT INTO track_fingerprints (
+    track_id, audio_stream_sha256, chromaprint, fingerprint_version, chromaprint_length_sec
+) VALUES (
+    sqlc.arg(track_id), sqlc.narg(audio_stream_sha256), sqlc.narg(chromaprint),
+    sqlc.arg(fingerprint_version), sqlc.arg(chromaprint_length_sec)
+)
+ON CONFLICT (track_id) DO UPDATE SET
+    audio_stream_sha256    = EXCLUDED.audio_stream_sha256,
+    chromaprint            = EXCLUDED.chromaprint,
+    fingerprint_version    = EXCLUDED.fingerprint_version,
+    chromaprint_length_sec = EXCLUDED.chromaprint_length_sec,
+    computed_at            = now();
+
+-- name: DeleteTrackFingerprint :exec
+-- A file changed but could not be fingerprinted, for a reason unrelated to the
+-- file. The stored row describes the OLD bytes, so it goes and the backfill
+-- re-derives it — nothing may keep trusting a stale identity.
+DELETE FROM track_fingerprints WHERE track_id = $1;
+
+-- name: ListTracksNeedingFingerprint :many
+-- The backfill's work queue (#3908): tracks with no fingerprint, or one derived
+-- by an older method. Keyset-paged on id so a pass visits each track at most
+-- once. That cursor is load-bearing: an inconclusive attempt writes no row, so
+-- without it a file that keeps timing out would be listed again straight away
+-- and retried in a tight loop. Missing tracks are skipped — there is no file to
+-- read.
+SELECT t.id, t.file_path
+  FROM tracks t
+  LEFT JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL
+   -- A row taken at another length is as stale as one from an older method:
+   -- chromaprints at two lengths cannot be compared (#3913).
+   AND (f.track_id IS NULL
+        OR f.fingerprint_version < sqlc.arg(current_version)
+        OR f.chromaprint_length_sec <> sqlc.arg(chromaprint_length_sec))
+   AND t.id > sqlc.arg(after_id)
+ ORDER BY t.id
+ LIMIT sqlc.arg(batch_limit);
+
+-- name: GetFingerprintCoverage :one
+-- The admin gauge for the backfill. fingerprinted + rejected + pending = total.
+-- "Current" means derived by the current method AT the current length: a row at
+-- another length is pending, because the backfill will re-derive it. rejected is
+-- a current row with a NULL half: a tool ran and refused the file, which is
+-- settled rather than waiting. Missing tracks are excluded, or the gauge could
+-- never reach the end.
+SELECT count(*)::bigint AS total,
+       count(*) FILTER (
+           WHERE f.fingerprint_version >= sqlc.arg(current_version)
+             AND f.chromaprint_length_sec = sqlc.arg(chromaprint_length_sec)
+             AND f.audio_stream_sha256 IS NOT NULL AND f.chromaprint IS NOT NULL
+       )::bigint AS fingerprinted,
+       count(*) FILTER (
+           WHERE f.fingerprint_version >= sqlc.arg(current_version)
+             AND f.chromaprint_length_sec = sqlc.arg(chromaprint_length_sec)
+             AND (f.audio_stream_sha256 IS NULL OR f.chromaprint IS NULL)
+       )::bigint AS rejected,
+       count(*) FILTER (
+           WHERE f.track_id IS NULL
+              OR f.fingerprint_version < sqlc.arg(current_version)
+              OR f.chromaprint_length_sec <> sqlc.arg(chromaprint_length_sec)
+       )::bigint AS pending
+  FROM tracks t
+  LEFT JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NULL;
@@ -0,0 +1,101 @@
+-- Duplicate merge (Scribe #3911). Every statement here runs inside the one
+-- transaction library.MergeDuplicateGroup opens, after the removed copy's file
+-- is already gone. The loser's own track row is deleted last with DeleteTrack;
+-- what these do is move everything it carries onto the survivor first, so that
+-- delete's CASCADE finds nothing left to destroy.
+
+-- name: LockDuplicateGroupForMerge :one
+-- Locks the group for the rest of the transaction, so two merges of one group
+-- cannot run at once.
+SELECT id, tier, status
+  FROM duplicate_groups
+ WHERE id = sqlc.arg(id)
+   FOR UPDATE;
+
+-- name: ListDuplicateGroupMergeMembers :many
+SELECT t.id, t.file_path, t.file_format, t.file_size, t.added_at, t.album_id,
+       t.mbid, albums.mbid AS album_mbid
+  FROM duplicate_group_members m
+  JOIN tracks t  ON t.id = m.track_id
+  JOIN albums    ON albums.id = t.album_id
+ WHERE m.group_id = sqlc.arg(group_id)
+ ORDER BY t.id;
+
+-- Plain repoints: no unique key involves track_id, so moving rows cannot collide.
+
+-- name: MergeRepointPlayEvents :execrows
+UPDATE play_events SET track_id = sqlc.arg(survivor_id)::uuid WHERE track_id = sqlc.arg(loser_id)::uuid;
+
+-- name: MergeRepointSkipEvents :execrows
+UPDATE skip_events SET track_id = sqlc.arg(survivor_id)::uuid WHERE track_id = sqlc.arg(loser_id)::uuid;
+
+-- name: MergeRepointContextualLikes :execrows
+UPDATE contextual_likes SET track_id = sqlc.arg(survivor_id)::uuid WHERE track_id = sqlc.arg(loser_id)::uuid;
+
+-- name: MergeRepointPlaybackErrors :execrows
+UPDATE playback_errors SET track_id = sqlc.arg(survivor_id)::uuid WHERE track_id = sqlc.arg(loser_id)::uuid;
+
+-- name: MergeRepointLidarrRequests :execrows
+UPDATE lidarr_requests SET matched_track_id = sqlc.arg(survivor_id)::uuid
+ WHERE matched_track_id = sqlc.arg(loser_id)::uuid;
+
+-- name: MergeRepointPlaylistTracks :many
+-- playlist_tracks is keyed by (playlist_id, position), so repointing keeps every
+-- entry exactly where it was. A playlist that held both copies simply holds the
+-- survivor twice — the user put two entries there, and both stay.
+UPDATE playlist_tracks SET track_id = sqlc.arg(survivor_id)::uuid
+ WHERE track_id = sqlc.arg(loser_id)::uuid
+RETURNING playlist_id;
+
+-- Collision-safe merges: a unique key includes track_id, so the survivor may
+-- already hold a matching row. Copy what it lacks; DeleteTrack's CASCADE then
+-- removes the loser's originals.
+
+-- name: MergeCopyGeneralLikes :many
+-- One like per user. A user who liked both copies keeps a single like, dated to
+-- the earlier of the two.
+INSERT INTO general_likes (user_id, track_id, liked_at)
+SELECT user_id, sqlc.arg(survivor_id)::uuid, liked_at
+  FROM general_likes
+ WHERE track_id = sqlc.arg(loser_id)::uuid
+ON CONFLICT (user_id, track_id) DO UPDATE
+   SET liked_at = LEAST(general_likes.liked_at, EXCLUDED.liked_at)
+RETURNING user_id;
+
+-- name: MergeCopyTrackTags :execrows
+INSERT INTO track_tags (track_id, tag, weight)
+SELECT sqlc.arg(survivor_id)::uuid, tag, weight
+  FROM track_tags
+ WHERE track_id = sqlc.arg(loser_id)::uuid
+ON CONFLICT (track_id, tag) DO NOTHING;
+
+-- name: MergeCopyTrackSimilarity :execrows
+-- Rewrites the loser to the survivor on either side of an edge. An edge between
+-- the two copies would become a track similar to itself — the table forbids
+-- that, and it means nothing — so it is dropped. An edge the survivor already
+-- has from the same source is kept as it is.
+INSERT INTO track_similarity (track_a_id, track_b_id, score, source, fetched_at)
+SELECT CASE WHEN track_a_id = sqlc.arg(loser_id)::uuid THEN sqlc.arg(survivor_id)::uuid ELSE track_a_id END,
+       CASE WHEN track_b_id = sqlc.arg(loser_id)::uuid THEN sqlc.arg(survivor_id)::uuid ELSE track_b_id END,
+       score, source, fetched_at
+  FROM track_similarity
+ WHERE (track_a_id = sqlc.arg(loser_id)::uuid OR track_b_id = sqlc.arg(loser_id)::uuid)
+   AND (CASE WHEN track_a_id = sqlc.arg(loser_id)::uuid THEN sqlc.arg(survivor_id)::uuid ELSE track_a_id END)
+    <> (CASE WHEN track_b_id = sqlc.arg(loser_id)::uuid THEN sqlc.arg(survivor_id)::uuid ELSE track_b_id END)
+ON CONFLICT (track_a_id, track_b_id, source) DO NOTHING;
+
+-- name: MergeInheritTrackMbid :exec
+-- A recording MBID is what the similarity pipeline keys on. If only the removed
+-- copy carried one, the survivor takes it rather than going dark to similarity.
+UPDATE tracks AS survivor
+   SET mbid = loser.mbid
+  FROM tracks AS loser
+ WHERE survivor.id = sqlc.arg(survivor_id)::uuid
+   AND loser.id = sqlc.arg(loser_id)::uuid
+   AND survivor.mbid IS NULL
+   AND loser.mbid IS NOT NULL;
+
+-- name: MarkDuplicateGroupMerged :execrows
+UPDATE duplicate_groups
+   SET status = 'merged', resolved_at = now()
+ WHERE id = sqlc.arg(id) AND status = 'pending';
@@ -45,7 +45,22 @@ WHERE t.id <> $2
 --   enter the pool even when the similarity/random arms miss them; scored
 --   in Go via TasteMatch, so sim_score here is 0 pool-inclusion),
 --   $11 coplay_artists K (#1533 — tracks by artists co-played across the
--   instance with the seed's artist; source='user_cooccurrence').
+--   instance with the seed's artist; source='user_cooccurrence'),
+--   $12 order_seed (text) — see below.
+--
+-- $12 REPLACES `ORDER BY random()` IN FOUR ARMS (#3889). Those arms returned
+-- a stable set only while their LIMIT exceeded the rows eligible for them: at
+-- that point they returned all of them and the order stopped mattering,
+-- because the caller sorts by track id before scoring. Below that threshold
+-- they returned a random SUBSET, and two builds on the same day drew
+-- different ones — so "daily determinism" held by accident, and only for
+-- libraries smaller than the limits.
+--
+-- md5(id || seed) keeps the intent — an arbitrary spread that changes when
+-- the seed does — while making it reproducible for a given seed. The CALLER
+-- decides what that means: system mixes pass a per-(user, day) string and get
+-- the determinism they promise; radio passes a fresh value per request and
+-- keeps varying, which is what a radio should do.
 -- Returns same shape as LoadRadioCandidates plus similarity_score column.

 WITH
@@ -87,7 +102,7 @@ similar_artists AS (
    JOIN seed_artist sa ON asim.artist_a_id = sa.artist_id
    WHERE asim.source = 'listenbrainz'
      AND t.id NOT IN (SELECT id FROM excluded_ids)
-    ORDER BY asim.score DESC, random()
+    ORDER BY asim.score DESC, md5(t.id::text || $12::text)
    LIMIT $6
 ),
 tag_overlap AS (
@@ -115,7 +130,7 @@ likes_overlap AS (
          WHERE t.id = gl.track_id
            AND trim(g_overlap.g) IN (SELECT tag FROM seed_tags)
      )
-    ORDER BY random()
+    ORDER BY md5(gl.track_id::text || $12::text)
    LIMIT $8
 ),
 taste_overlap AS (
@@ -142,7 +157,7 @@ coplay_artists AS (
    WHERE asim.source = 'user_cooccurrence'
      AND t.id NOT IN (SELECT id FROM excluded_ids)
      AND t.id <> $2
-    ORDER BY asim.score DESC, random()
+    ORDER BY asim.score DESC, md5(t.id::text || $12::text)
    LIMIT $11
 ),
 random_fill AS (
@@ -158,7 +173,7 @@ random_fill AS (
          UNION SELECT track_id FROM taste_overlap
          UNION SELECT track_id FROM coplay_artists
      )
-    ORDER BY random()
+    ORDER BY md5(t.id::text || $12::text)
    LIMIT $9
 )
 SELECT
@@ -155,17 +155,19 @@ SELECT id, file_path FROM tracks
   AND mbid = sqlc.arg(mbid)::text
 LIMIT 2;

-- name: FindMissingTrackByFingerprint :many
-- Move detection fallback for files with no MBID (#2528). Exact byte size AND
-- exact decoded duration is a strong pair: a plain move or rename preserves
-- both, while a re-encode changes at least one — and a re-encode genuinely is a
-- different file, so failing to match there is correct rather than a gap.
+-- name: FindMissingTrackByAudioHash :many
+-- Move detection fallback for files with no MBID (#2528, #3914). The audio stream
+-- hash identifies the encoded audio itself, so it survives a rename, a move and a
+-- retag — anything short of a re-encode. It replaced (file_size, duration_ms),
+-- which could pair two unrelated files that happened to share a byte count and a
+-- duration, and missed a file retagged in place, whose size changes.
 --
 -- Same missing-only constraint and same LIMIT 2 rationale as the MBID variant.
-SELECT id, file_path FROM tracks
- WHERE missing_since IS NOT NULL
-   AND file_size = sqlc.arg(file_size)
-   AND duration_ms = sqlc.arg(duration_ms)
+SELECT t.id, t.file_path
+  FROM tracks t
+  JOIN track_fingerprints f ON f.track_id = t.id
+ WHERE t.missing_since IS NOT NULL
+   AND f.audio_stream_sha256 = sqlc.arg(audio_stream_sha256)
 LIMIT 2;

 -- name: AdoptTrackPath :execrows
@@ -87,6 +87,10 @@ var dataTables = []string{
 	// pristine Discover knobs rather than whatever a previous test tuned.
 	"discover_tuning",
 	"recommendation_tuning_audit",
+	"duplicate_group_members", // M400
+	"duplicate_groups",
+	"duplicate_sweeps",
+	"track_fingerprints", // M400
 	"tracks",
 	"albums",
 	"artists",
@@ -126,4 +130,15 @@ func ResetDB(t *testing.T, pool *pgxpool.Pool) {
 	); err != nil {
 		t.Fatalf("dbtest.ResetDB reset tag-sources version: %v", err)
 	}
+	// Fingerprinting settings (M400 #3913), a singleton like the counters above.
+	// Every column goes back to its migration default rather than to literals
+	// written here, so a test can pin the Go defaults to the migration's.
+	if _, err := pool.Exec(ctx, `
+		UPDATE fingerprint_settings
+		   SET enabled = DEFAULT, chromaprint_length_sec = DEFAULT,
+		       acoustic_max_bit_error_rate = DEFAULT, backfill_concurrency = DEFAULT,
+		       sweep_interval_hours = DEFAULT, updated_at = DEFAULT`,
+	); err != nil {
+		t.Fatalf("dbtest.ResetDB reset fingerprint settings: %v", err)
+	}
 }
@@ -5,12 +5,16 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"log/slog"
 	"os"
+	"path/filepath"
+	"syscall"

 	"github.com/jackc/pgx/v5"
 	"github.com/jackc/pgx/v5/pgtype"
 	"github.com/jackc/pgx/v5/pgxpool"

+	"git.fabledsword.com/bvandeusen/minstrel/internal/coverart"
 	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
 	syncpkg "git.fabledsword.com/bvandeusen/minstrel/internal/sync"
 )
@@ -19,54 +23,179 @@ import (
 // that has no row in tracks.
 var ErrTrackNotFound = errors.New("library: track not found")

-// DeleteTrackFile removes a track file from disk and its row from the
-// tracks table. Album and artist rows are left untouched.
+// removeFile is os.Remove behind a variable so a test can make removal fail the
+// way a read-only mount or a wrongly-owned directory does. A chmod-based test
+// cannot stand in for that: root ignores permission bits, so in a CI container
+// running as root it would pass without ever exercising the failure.
+var removeFile = os.Remove
+
+// FileRemoveError reports that a track's file exists but could not be removed.
+// When DeleteTrackFile returns one, NOTHING was deleted: the row, its likes, its
+// play history and its playlist memberships are all intact.
+type FileRemoveError struct {
+	Path string
+	// UID and GID are the identity the server process runs as — the half of a
+	// permission problem the operator cannot see from the host side.
+	UID, GID int
+	Err      error
+}
+
+func (e *FileRemoveError) Error() string { return fmt.Sprintf("remove track file: %v", e.Err) }
+
+func (e *FileRemoveError) Unwrap() error { return e.Err }
+
+// Dir is the directory removal needs write access to. Unlinking a file writes to
+// its PARENT, so a world-writable file inside a read-only directory still cannot
+// be removed — naming the file's own permissions would send the operator to the
+// wrong place.
+func (e *FileRemoveError) Dir() string { return filepath.Dir(e.Path) }
+
+// NotWritable reports whether the library is unwritable for this process — a
+// read-only mount or a permission denial — rather than an I/O fault. It is the
+// case the operator can fix, so callers answer it differently.
+func (e *FileRemoveError) NotWritable() bool {
+	return errors.Is(e.Err, fs.ErrPermission) || errors.Is(e.Err, syscall.EROFS)
+}
+
+// Reason is the underlying cause without the path os.Remove already wrapped
+// around it, for messages that name the directory themselves.
+func (e *FileRemoveError) Reason() string {
+	var pathErr *fs.PathError
+	if errors.As(e.Err, &pathErr) {
+		return pathErr.Err.Error()
+	}
+	return e.Err.Error()
+}
+
+// DeletedTrack reports what a delete tidied away beyond the track itself.
+type DeletedTrack struct {
+	// AlbumID is set when the track was its album's last, so the album went too.
+	AlbumID *pgtype.UUID
+	// ArtistID is set when that album was its artist's last, so the artist went too.
+	ArtistID *pgtype.UUID
+}
+
+// DeleteTrackFile removes a track's file from disk and then its row, tidying
+// away an album or artist the delete leaves empty. It is the ONLY path that
+// deletes a track file: the admin remove-track endpoint and quarantine's Delete
+// file both come through here (#3918).
 //
-// Steps:
-//  1. Look up the track to get its file_path.
-//  2. Remove the file from disk. fs.ErrNotExist is OK — already gone.
-//  3. Delete the tracks row.
+// Order is the whole contract. The file goes first, and if it cannot go — a
+// read-only mount, a permission denial, an I/O error — nothing else happens and
+// a *FileRemoveError comes back. Proceeding past that failure is how #3918 lost
+// history: tracks CASCADEs to play_events, general_likes, contextual_likes,
+// playlist_tracks, track_tags and playback_errors, so the row and everything
+// hanging off it were destroyed while the file survived, and the next scan
+// re-imported it as a brand-new track with none of it.
 //
-// Order matters: file first, then DB. If the file delete fails (permission,
-// I/O error), we leave the DB row alone so the admin can retry.
+// A file that is already gone (fs.ErrNotExist) is not a failure; the row is
+// removed as asked.
 //
-// The reverse failure mode — file gone, DB row still present — IS reconciled
-// now, and not by this function: the scan's reconcile pass stamps
-// tracks.missing_since (#2523), every selection path filters on it, and a file
-// that returns is un-marked or adopted at its new path (#2528). That is the
-// normal life of a vanished file and it is deliberately non-destructive: the
-// row, its play history and its likes survive, because a missing file is a
-// track Minstrel still knows about (#2527).
+// This is NOT the missing-file path. That lifecycle is deliberately
+// non-destructive: reconcile stamps missing_since (#2523), selection paths
+// filter on it, and a returning file is un-marked or adopted (#2528). This is the
+// explicit, irreversible "remove this recording", never the way to tidy up a row
+// whose file merely went away.
 //
-// So this function is NOT the missing-file path. It is the explicit admin
-// action "remove this recording from disk and from the library", and it is
-// irreversible: tracks CASCADEs to play_events, general_likes_tracks,
-// contextual_likes, track_tags and playback_errors. Reach for it when the
-// operator means to destroy the record, never to tidy up a row whose file
-// merely went away.
-func DeleteTrackFile(ctx context.Context, pool *pgxpool.Pool, trackID pgtype.UUID) error {
+// dataDir, when set, also clears the cached art of an artist the delete removed.
+// logger may be nil.
+func DeleteTrackFile(
+	ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger, dataDir string, trackID pgtype.UUID,
+) (DeletedTrack, error) {
+	if logger == nil {
+		logger = slog.Default()
+	}
 	q := dbq.New(pool)
 	track, err := q.GetTrackByID(ctx, trackID)
 	if err != nil {
 		if errors.Is(err, pgx.ErrNoRows) {
-			return ErrTrackNotFound
+			return DeletedTrack{}, ErrTrackNotFound
 		}
-		return fmt.Errorf("get track: %w", err)
+		return DeletedTrack{}, fmt.Errorf("get track: %w", err)
 	}

-	if err := os.Remove(track.FilePath); err != nil && !errors.Is(err, fs.ErrNotExist) {
-		return fmt.Errorf("remove file: %w", err)
+	if err := removeTrackFileOnDisk(track.FilePath); err != nil {
+		return DeletedTrack{}, err
 	}

-	if _, err := pool.Exec(ctx, "DELETE FROM tracks WHERE id = $1", trackID); err != nil {
-		return fmt.Errorf("delete row: %w", err)
+	// The row and any album or artist it empties go together, so a failure
+	// partway cannot leave a deleted track with a ghost album behind it.
+	tx, err := pool.Begin(ctx)
+	if err != nil {
+		return DeletedTrack{}, fmt.Errorf("begin tx: %w", err)
 	}
-	// Log the change after the delete succeeds. Best-effort: a Warn-level
-	// failure here would leave the cache index orphaned on offline clients
-	// until the next scan touches the surrounding album.
+	defer func() { _ = tx.Rollback(ctx) }()
+	tq := dbq.New(tx)
+
+	deleted, err := tq.DeleteTrack(ctx, trackID)
+	if err != nil {
+		if errors.Is(err, pgx.ErrNoRows) {
+			// Removed by someone else between the lookup and here.
+			return DeletedTrack{}, ErrTrackNotFound
+		}
+		return DeletedTrack{}, fmt.Errorf("delete track: %w", err)
+	}
+
+	out, err := tidyEmptiedAlbum(ctx, tq, deleted.AlbumID)
+	if err != nil {
+		return DeletedTrack{}, err
+	}
+
+	if err := tx.Commit(ctx); err != nil {
+		return DeletedTrack{}, fmt.Errorf("commit: %w", err)
+	}
+
+	// Both of these run after the delete has committed, so neither may fail
+	// it: the recording is gone either way. An unlogged change leaves the track
+	// in offline clients' caches until the next scan touches its album; a
+	// leftover art directory is only disk.
 	if err := syncpkg.LogChange(ctx, pool, syncpkg.EntityTrack,
 		syncpkg.FormatUUID(trackID), syncpkg.OpDelete); err != nil {
-		return fmt.Errorf("log change: %w", err)
+		logger.Warn("track delete: LogChange failed", "track_id", syncpkg.FormatUUID(trackID), "err", err)
+	}
+	if out.ArtistID != nil && dataDir != "" {
+		if err := coverart.CleanupArtistArt(dataDir, *out.ArtistID); err != nil {
+			logger.Warn("track delete: artist-art cleanup failed",
+				"artist_id", syncpkg.FormatUUID(*out.ArtistID), "err", err)
+		}
+	}
+	return out, nil
+}
+
+// removeTrackFileOnDisk is the one rule for removing a track's file, shared by
+// DeleteTrackFile and the duplicate merge. A file already gone is fine; anything
+// else comes back as a *FileRemoveError, and the caller must then change nothing
+// in the database (#3918).
+func removeTrackFileOnDisk(path string) error {
+	if err := removeFile(path); err != nil && !errors.Is(err, fs.ErrNotExist) {
+		return &FileRemoveError{Path: path, UID: os.Getuid(), GID: os.Getgid(), Err: err}
 	}
 	return nil
 }
+
+// tidyEmptiedAlbum removes an album a track delete left with no tracks, and its
+// artist if that album was the artist's last. It runs on the caller's
+// transaction, so the tidy-up commits or rolls back with the delete itself.
+func tidyEmptiedAlbum(ctx context.Context, tq *dbq.Queries, albumID pgtype.UUID) (DeletedTrack, error) {
+	var out DeletedTrack
+	album, err := tq.DeleteAlbumIfEmpty(ctx, albumID)
+	switch {
+	case err == nil:
+		id := album.ID
+		out.AlbumID = &id
+		artistID, aerr := tq.DeleteArtistIfEmpty(ctx, album.ArtistID)
+		switch {
+		case aerr == nil:
+			out.ArtistID = &artistID
+		case errors.Is(aerr, pgx.ErrNoRows):
+			// The artist still has other albums or stray tracks.
+		default:
+			return DeletedTrack{}, fmt.Errorf("delete artist if empty: %w", aerr)
+		}
+	case errors.Is(err, pgx.ErrNoRows):
+		// The album still has other tracks.
+	default:
+		return DeletedTrack{}, fmt.Errorf("delete album if empty: %w", err)
+	}
+	return out, nil
+}
@@ -4,9 +4,11 @@ import (
 	"context"
 	"errors"
 	"io"
+	"io/fs"
 	"log/slog"
 	"os"
 	"path/filepath"
+	"syscall"
 	"testing"

 	"github.com/jackc/pgx/v5/pgtype"
@@ -64,6 +66,15 @@ func seedTrack(t *testing.T, pool *pgxpool.Pool, filePath string) (dbq.Track, db
 	return track, album, artist
 }

+// stubRemoveFile makes file removal fail (or succeed) on demand for one test.
+// See removeFile for why this is a seam rather than a chmod.
+func stubRemoveFile(t *testing.T, fn func(string) error) {
+	t.Helper()
+	orig := removeFile
+	removeFile = fn
+	t.Cleanup(func() { removeFile = orig })
+}
+
 func TestDeleteTrackFile_HappyPath(t *testing.T) {
 	pool := newPool(t)
 	q := dbq.New(pool)
@@ -73,9 +84,18 @@ func TestDeleteTrackFile_HappyPath(t *testing.T) {
 	if err := os.WriteFile(path, []byte("payload"), 0o644); err != nil {
 		t.Fatalf("write file: %v", err)
 	}
-	track, album, _ := seedTrack(t, pool, path)
+	track, album, artist := seedTrack(t, pool, path)
+	// A sibling keeps the album non-empty, so this case pins that the tidy-up
+	// only removes an album the delete actually emptied.
+	if _, err := q.UpsertTrack(context.Background(), dbq.UpsertTrackParams{
+		Title: "Sibling", AlbumID: album.ID, ArtistID: artist.ID,
+		DurationMs: 1000, FilePath: filepath.Join(dir, "sibling.mp3"), FileSize: 100, FileFormat: "mp3",
+	}); err != nil {
+		t.Fatalf("sibling: %v", err)
+	}

-	if err := DeleteTrackFile(context.Background(), pool, track.ID); err != nil {
+	got, err := DeleteTrackFile(context.Background(), pool, nil, "", track.ID)
+	if err != nil {
 		t.Fatalf("DeleteTrackFile: %v", err)
 	}

@@ -85,9 +105,99 @@ func TestDeleteTrackFile_HappyPath(t *testing.T) {
 	if _, err := q.GetTrackByID(context.Background(), track.ID); err == nil {
 		t.Errorf("track row still exists")
 	}
-	// Album row preserved (other tracks may reference it).
 	if _, err := q.GetAlbumByID(context.Background(), album.ID); err != nil {
-		t.Errorf("album row vanished: %v", err)
+		t.Errorf("album with a remaining track vanished: %v", err)
+	}
+	if got.AlbumID != nil || got.ArtistID != nil {
+		t.Errorf("reported tidy-up %+v for an album that still has a track", got)
+	}
+}
+
+func TestDeleteTrackFile_EmptiedAlbumAndArtistGoToo(t *testing.T) {
+	pool := newPool(t)
+	q := dbq.New(pool)
+
+	path := filepath.Join(t.TempDir(), "lone.mp3")
+	if err := os.WriteFile(path, []byte("payload"), 0o644); err != nil {
+		t.Fatalf("write file: %v", err)
+	}
+	track, album, artist := seedTrack(t, pool, path)
+
+	got, err := DeleteTrackFile(context.Background(), pool, nil, "", track.ID)
+	if err != nil {
+		t.Fatalf("DeleteTrackFile: %v", err)
+	}
+	if got.AlbumID == nil || *got.AlbumID != album.ID {
+		t.Errorf("AlbumID = %v, want %v", got.AlbumID, album.ID)
+	}
+	if got.ArtistID == nil || *got.ArtistID != artist.ID {
+		t.Errorf("ArtistID = %v, want %v", got.ArtistID, artist.ID)
+	}
+	if _, err := q.GetAlbumByID(context.Background(), album.ID); err == nil {
+		t.Errorf("emptied album row still exists")
+	}
+	if _, err := q.GetArtistByID(context.Background(), artist.ID); err == nil {
+		t.Errorf("emptied artist row still exists")
+	}
+}
+
+// The #3918 proof. A file that cannot be removed must leave EVERYTHING in place:
+// the row is what carries likes, plays and playlist memberships, and the file
+// surviving means the next scan would re-import it as a stranger.
+func TestDeleteTrackFile_UnremovableFileDeletesNothing(t *testing.T) {
+	cases := []struct {
+		name        string
+		errno       syscall.Errno
+		notWritable bool
+	}{
+		{"read-only mount", syscall.EROFS, true},
+		{"permission denied", syscall.EACCES, true},
+		{"operation not permitted", syscall.EPERM, true},
+		{"i/o error", syscall.EIO, false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			pool := newPool(t)
+			q := dbq.New(pool)
+
+			dir := t.TempDir()
+			path := filepath.Join(dir, "track.mp3")
+			if err := os.WriteFile(path, []byte("payload"), 0o644); err != nil {
+				t.Fatalf("write file: %v", err)
+			}
+			track, album, _ := seedTrack(t, pool, path)
+			stubRemoveFile(t, func(name string) error {
+				return &fs.PathError{Op: "remove", Path: name, Err: tc.errno}
+			})
+
+			_, err := DeleteTrackFile(context.Background(), pool, nil, "", track.ID)
+
+			var fre *FileRemoveError
+			if !errors.As(err, &fre) {
+				t.Fatalf("err = %v, want a *FileRemoveError", err)
+			}
+			if fre.NotWritable() != tc.notWritable {
+				t.Errorf("NotWritable = %v, want %v", fre.NotWritable(), tc.notWritable)
+			}
+			if fre.Dir() != dir {
+				t.Errorf("Dir = %q, want the parent directory %q", fre.Dir(), dir)
+			}
+			if fre.Reason() != tc.errno.Error() {
+				t.Errorf("Reason = %q, want %q", fre.Reason(), tc.errno.Error())
+			}
+			if fre.UID != os.Getuid() || fre.GID != os.Getgid() {
+				t.Errorf("identity = %d:%d, want this process's %d:%d", fre.UID, fre.GID, os.Getuid(), os.Getgid())
+			}
+			if _, err := q.GetTrackByID(context.Background(), track.ID); err != nil {
+				t.Errorf("track row was deleted although its file was not: %v", err)
+			}
+			if _, err := q.GetAlbumByID(context.Background(), album.ID); err != nil {
+				t.Errorf("album row was deleted although the track's file was not: %v", err)
+			}
+			if _, err := os.Stat(path); err != nil {
+				t.Errorf("file gone although removal was refused: %v", err)
+			}
+		})
 	}
 }

@@ -97,7 +207,7 @@ func TestDeleteTrackFile_FileAlreadyGoneSucceeds(t *testing.T) {

 	track, _, _ := seedTrack(t, pool, "/no/such/file/anywhere.mp3")

-	if err := DeleteTrackFile(context.Background(), pool, track.ID); err != nil {
+	if _, err := DeleteTrackFile(context.Background(), pool, nil, "", track.ID); err != nil {
 		t.Fatalf("DeleteTrackFile with missing file: %v", err)
 	}
 	if _, err := q.GetTrackByID(context.Background(), track.ID); err == nil {
@@ -112,7 +222,7 @@ func TestDeleteTrackFile_NotFoundReturnsErr(t *testing.T) {
 	bogus.Bytes = [16]byte{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}
 	bogus.Valid = true

-	err := DeleteTrackFile(context.Background(), pool, bogus)
+	_, err := DeleteTrackFile(context.Background(), pool, nil, "", bogus)
 	if !errors.Is(err, ErrTrackNotFound) {
 		t.Errorf("err = %v, want ErrTrackNotFound", err)
 	}
@@ -0,0 +1,430 @@
+package library
+
+import (
+	"math"
+	"math/bits"
+	"sort"
+)
+
+// Duplicate matching (M400 #3909).
+//
+// Pure functions over fingerprints: no database, no files. This is the part that
+// decides whether two tracks in the operator's library are proposed as one
+// recording, so every rule in it has to be falsifiable in a unit test.
+//
+// Two tiers, answering different questions:
+//
+//	exact     equal audio_stream_sha256 — the same encoded audio bytes. No score,
+//	          no threshold, no false positives (the #3885 pair).
+//	acoustic  chromaprint fingerprints that agree closely once aligned — the
+//	          same recording at another bitrate or in another codec.
+//
+// The acoustic comparison follows the approach of AcoustID's pg_acoustid
+// (acoustid_compare.c): vote on the relative offset between two fingerprints
+// using items that agree in their high bits, then measure disagreement at the
+// winning offset. Reimplemented from that description; no code was copied.
+// The alignment window and match-bit width below are taken from it.
+
+// maxAlignOffsetItems bounds how far apart two fingerprints may be shifted and
+// still be compared: ±120 items, about 15 seconds at chromaprint's ~8 items per
+// second. Covers a leading silence trimmed differently or a short intro; the
+// same bound pg_acoustid uses (ACOUSTID_MAX_ALIGN_OFFSET).
+const maxAlignOffsetItems = 120
+
+// alignMatchBits is how many high bits two items must share to vote for an
+// offset. Matching whole 32-bit items would miss the same recording at another
+// bitrate, whose low bits are noisier; 14 is pg_acoustid's MATCH_BITS.
+const alignMatchBits = 14
+
+// minOverlapItems is the least overlap worth a verdict: about 10 seconds. A few
+// items agreeing perfectly is not evidence that two recordings are one.
+const minOverlapItems = 80
+
+// minDistinctFraction rejects low-information fingerprints before they can
+// match. Near-silence, a sustained tone or a click track produces the same few
+// items over and over, and two such tracks agree closely without being the
+// same recording. Real music is overwhelmingly distinct item to item, so this
+// floor only catches the pathological case. A judgment value, not a measured
+// one — revisit if the sweep reports real tracks refused for it.
+const minDistinctFraction = 0.3
+
+// defaultAcousticMaxBitErrorRate is the most disagreement two aligned
+// fingerprints may show and still be proposed as one recording. Unrelated audio
+// sits near 0.5; the same recording re-encoded lands well under 0.1.
+//
+// Deliberately conservative. The operator's stated worry is the opposite of a
+// missed duplicate: "the same song can appear in different albums, usually it's
+// a different recording", and an instrumental shares its vocal version's
+// harmony, which chroma features capture. A false merge is the failure that
+// matters, and the report is reviewed anyway. This is an unmeasured default:
+// calibrate it against real pairs once the backfill (#3908) has populated the
+// library, then expose it in Settings (#3913).
+const defaultAcousticMaxBitErrorRate = 0.15
+
+// durationToleranceMs is how far apart two tracks' durations may be and still be
+// compared. Encoders pad and trim a little; different edits differ by more.
+const durationToleranceMs = 3000
+
+// maxAcousticGroupSize caps an acoustic group. A cluster bigger than this is far
+// more likely a shared jingle, a skit or a low-information pattern than eight
+// copies of one recording, and proposing it would bury the real duplicates.
+// Exact-tier groups are not capped: identical bytes are identical however many.
+const maxAcousticGroupSize = 8
+
+// acousticScore is the result of comparing two fingerprints.
+type acousticScore struct {
+	// Offset is how many items b is shifted against a: b[i+Offset] aligns with
+	// a[i].
+	Offset int
+	// Overlap is how many aligned items were compared.
+	Overlap int
+	// BitErrorRate is the fraction of differing bits over the overlap, 0..1.
+	BitErrorRate float64
+}
+
+// preparedPrint is a fingerprint with the parts every comparison needs worked
+// out once. The sweep compares each track with every other track within a few
+// seconds of its duration, so rebuilding the alignment index for each pair would
+// dominate its cost.
+type preparedPrint struct {
+	items       []int32
+	index       map[uint32][]int
+	informative bool
+}
+
+// preparePrint indexes a fingerprint's items by their high bits and records
+// whether it varies enough to be compared at all.
+func preparePrint(fp []int32) *preparedPrint {
+	// Each bucket keeps only a few positions: a value repeating many times is
+	// uninformative, and letting it vote once per repeat would make every
+	// pairing O(n²).
+	const keepPerBucket = 4
+	p := &preparedPrint{items: fp, index: make(map[uint32][]int, len(fp))}
+	seen := make(map[int32]struct{}, len(fp))
+	for i, v := range fp {
+		seen[v] = struct{}{}
+		key := alignKey(v)
+		if pos := p.index[key]; len(pos) < keepPerBucket {
+			p.index[key] = append(pos, i)
+		}
+	}
+	p.informative = len(fp) > 0 && float64(len(seen)) >= minDistinctFraction*float64(len(fp))
+	return p
+}
+
+func alignKey(v int32) uint32 { return uint32(v) >> (32 - alignMatchBits) }
+
+// compareChromaprint aligns two raw fingerprints and measures how much they
+// disagree. ok is false when no verdict is possible: no offset gathered any
+// votes, the overlap at the best offset is too short, or either side carries
+// too little information to mean anything.
+func compareChromaprint(a, b []int32) (acousticScore, bool) {
+	return comparePrepared(preparePrint(a), preparePrint(b))
+}
+
+// comparePrepared is compareChromaprint over fingerprints already prepared.
+func comparePrepared(a, b *preparedPrint) (acousticScore, bool) {
+	if len(a.items) < minOverlapItems || len(b.items) < minOverlapItems {
+		return acousticScore{}, false
+	}
+	if !a.informative || !b.informative {
+		return acousticScore{}, false
+	}
+
+	offset, ok := bestOffset(a, b)
+	if !ok {
+		return acousticScore{}, false
+	}
+
+	// a[i] aligns with b[i+offset]; walk the indices valid on both sides.
+	start := max(0, -offset)
+	end := min(len(a.items), len(b.items)-offset)
+	overlap := end - start
+	if overlap < minOverlapItems {
+		return acousticScore{}, false
+	}
+	errBits := 0
+	for i := start; i < end; i++ {
+		errBits += bits.OnesCount32(uint32(a.items[i]) ^ uint32(b.items[i+offset]))
+	}
+	return acousticScore{
+		Offset:       offset,
+		Overlap:      overlap,
+		BitErrorRate: float64(errBits) / float64(32*overlap),
+	}, true
+}
+
+// bestOffset returns the relative shift most items agree on.
+func bestOffset(a, b *preparedPrint) (int, bool) {
+	votes := make([]int, 2*maxAlignOffsetItems+1)
+	for j, v := range b.items {
+		for _, i := range a.index[alignKey(v)] {
+			off := j - i
+			if off >= -maxAlignOffsetItems && off <= maxAlignOffsetItems {
+				votes[off+maxAlignOffsetItems]++
+			}
+		}
+	}
+
+	best, bestVotes := 0, 0
+	for k, n := range votes {
+		// Strictly greater keeps the smallest shift on a tie, which is the more
+		// likely truth and keeps the result deterministic.
+		if n > bestVotes || (n == bestVotes && n > 0 && abs(k-maxAlignOffsetItems) < abs(best)) {
+			best, bestVotes = k-maxAlignOffsetItems, n
+		}
+	}
+	return best, bestVotes > 0
+}
+
+// fingerprintCandidate is one track as the grouping sees it.
+type fingerprintCandidate struct {
+	ID           string
+	DurationMs   int32
+	StreamSHA256 []byte
+	Chromaprint  []int32
+}
+
+// duplicateTier names what a group's evidence is.
+type duplicateTier string
+
+const (
+	tierExact    duplicateTier = "exact"
+	tierAcoustic duplicateTier = "acoustic"
+)
+
+// duplicateGroup is a set of tracks proposed as one recording. Members are
+// sorted by ID.
+type duplicateGroup struct {
+	Tier    duplicateTier
+	Members []string
+	// WorstBitErrorRate is the largest disagreement between any two members of
+	// an acoustic group — the weakest evidence the group rests on. Zero for
+	// exact groups.
+	WorstBitErrorRate float64
+}
+
+// groupingResult is what one grouping pass found.
+type groupingResult struct {
+	Groups []duplicateGroup
+	// OversizeClusters counts acoustic clusters discarded for exceeding
+	// maxAcousticGroupSize. Reported rather than silent: a sudden rise means the
+	// cap or the information floor needs attention.
+	OversizeClusters int
+}
+
+// groupUnit is one thing the acoustic pass compares: a single track, or an exact
+// group standing in for all its byte-identical copies.
+type groupUnit struct {
+	ids        []string // every member, sorted
+	durationMs int32
+	sortKey    string // the representative's id: ties on duration break on it
+	print      *preparedPrint
+	exact      bool // more than one member with identical audio
+	assigned   bool
+}
+
+// streamGrouper is the acoustic pass over units arriving in (durationMs, sortKey)
+// order. It holds only the units within durationToleranceMs of the oldest one
+// not yet settled, so memory is bounded by the densest few seconds of the
+// library rather than by its size — the whole library's fingerprints would be
+// hundreds of megabytes.
+//
+// A seed can be settled as soon as a unit arrives beyond its window: everything
+// it could group with has already arrived, and no later seed can reach back to
+// it because seeds only look forward. That is what makes the streamed result
+// identical to running the same pass over the whole sorted list.
+//
+// Grouping is COMPLETE-LINKAGE: a unit joins a group only if it matches every
+// unit already in it, within the duration tolerance and the bit-error limit.
+// Single-linkage would let a chain of near-misses — A close to B, B close to C —
+// drag A and C, which are not close, into one proposed merge. Complete linkage
+// also means any member can be chosen as the survivor (#3911).
+//
+// When an acoustic group absorbs an exact group, the result is tier acoustic: a
+// group is only as certain as its weakest link.
+type streamGrouper struct {
+	maxBitErrorRate float64
+	window          []*groupUnit
+	res             groupingResult
+}
+
+func newStreamGrouper(maxBitErrorRate float64) *streamGrouper {
+	return &streamGrouper{maxBitErrorRate: maxBitErrorRate}
+}
+
+// push adds the next unit. Units must arrive in non-decreasing
+// (durationMs, sortKey) order.
+func (g *streamGrouper) push(u *groupUnit) {
+	g.window = append(g.window, u)
+	for len(g.window) > 1 && u.durationMs-g.window[0].durationMs > durationToleranceMs {
+		g.settleOldest()
+	}
+}
+
+// finish settles every unit still waiting and returns what was found. Groups
+// are in no particular order; callers sort with sortGroups.
+func (g *streamGrouper) finish() groupingResult {
+	for len(g.window) > 0 {
+		g.settleOldest()
+	}
+	return g.res
+}
+
+func (g *streamGrouper) settleOldest() {
+	seed := g.window[0]
+	g.window[0] = nil // release it: the window's backing array outlives the slide
+	g.window = g.window[1:]
+	if seed.assigned {
+		return
+	}
+
+	group := []*groupUnit{seed}
+	worst := 0.0
+	for _, cand := range g.window {
+		if cand.durationMs-seed.durationMs > durationToleranceMs {
+			break
+		}
+		if cand.assigned {
+			continue
+		}
+		joined, worstWithCand := true, worst
+		for _, member := range group {
+			if abs32(cand.durationMs-member.durationMs) > durationToleranceMs {
+				joined = false
+				break
+			}
+			score, ok := comparePrepared(member.print, cand.print)
+			if !ok || score.BitErrorRate > g.maxBitErrorRate {
+				joined = false
+				break
+			}
+			worstWithCand = math.Max(worstWithCand, score.BitErrorRate)
+		}
+		if joined {
+			group = append(group, cand)
+			worst = worstWithCand
+		}
+	}
+
+	if len(group) == 1 {
+		if seed.exact {
+			g.res.Groups = append(g.res.Groups, duplicateGroup{Tier: tierExact, Members: seed.ids})
+		}
+		return
+	}
+	// Count units, not tracks: an absorbed exact group is one piece of acoustic
+	// evidence however many identical files it holds.
+	if len(group) > maxAcousticGroupSize {
+		g.res.OversizeClusters++
+		for _, member := range group {
+			member.assigned = true
+			// The acoustic evidence is untrustworthy; identical bytes are not.
+			// An exact group caught inside an oversize cluster is still proposed.
+			if member.exact {
+				g.res.Groups = append(g.res.Groups, duplicateGroup{Tier: tierExact, Members: member.ids})
+			}
+		}
+		return
+	}
+	var members []string
+	for _, member := range group {
+		member.assigned = true
+		members = append(members, member.ids...)
+	}
+	sort.Strings(members)
+	g.res.Groups = append(g.res.Groups, duplicateGroup{
+		Tier: tierAcoustic, Members: members, WorstBitErrorRate: worst,
+	})
+}
+
+// groupDuplicates proposes duplicate groups among candidates held in memory. It
+// runs the same streamGrouper the sweep uses, so there is one grouping rule.
+//
+// Exact groups come first: tracks sharing an audio stream hash. Each becomes a
+// single unit for the acoustic pass, represented by its member with the lowest
+// (duration, id) that has a chromaprint. That is the member the sweep's
+// duration-ordered stream meets first, which keeps the two identical. An exact
+// group with no chromaprint at all cannot be compared acoustically and stands
+// on its own.
+//
+// The output does not depend on input order.
+func groupDuplicates(cands []fingerprintCandidate, maxBitErrorRate float64) groupingResult {
+	byHash := map[string][]fingerprintCandidate{}
+	var units []*groupUnit
+	var printless []duplicateGroup
+	for _, c := range cands {
+		if len(c.StreamSHA256) > 0 {
+			byHash[string(c.StreamSHA256)] = append(byHash[string(c.StreamSHA256)], c)
+			continue
+		}
+		if len(c.Chromaprint) > 0 {
+			units = append(units, &groupUnit{
+				ids: []string{c.ID}, durationMs: c.DurationMs, sortKey: c.ID, print: preparePrint(c.Chromaprint),
+			})
+		}
+	}
+	for _, group := range byHash {
+		ids := make([]string, len(group))
+		for i, m := range group {
+			ids[i] = m.ID
+		}
+		sort.Strings(ids)
+
+		var rep *fingerprintCandidate
+		for i := range group {
+			m := &group[i]
+			if len(m.Chromaprint) == 0 {
+				continue
+			}
+			if rep == nil || m.DurationMs < rep.DurationMs || (m.DurationMs == rep.DurationMs && m.ID < rep.ID) {
+				rep = m
+			}
+		}
+		if rep == nil {
+			if len(group) > 1 {
+				printless = append(printless, duplicateGroup{Tier: tierExact, Members: ids})
+			}
+			continue
+		}
+		units = append(units, &groupUnit{
+			ids: ids, durationMs: rep.DurationMs, sortKey: rep.ID,
+			print: preparePrint(rep.Chromaprint), exact: len(group) > 1,
+		})
+	}
+
+	sort.Slice(units, func(i, j int) bool {
+		if units[i].durationMs != units[j].durationMs {
+			return units[i].durationMs < units[j].durationMs
+		}
+		return units[i].sortKey < units[j].sortKey
+	})
+	g := newStreamGrouper(maxBitErrorRate)
+	for _, u := range units {
+		g.push(u)
+	}
+	res := g.finish()
+	res.Groups = append(res.Groups, printless...)
+	sortGroups(res.Groups)
+	return res
+}
+
+// sortGroups orders groups by their first member. Groups are disjoint, so that
+// is a total order.
+func sortGroups(groups []duplicateGroup) {
+	sort.Slice(groups, func(i, j int) bool { return groups[i].Members[0] < groups[j].Members[0] })
+}
+
+func abs(n int) int {
+	if n < 0 {
+		return -n
+	}
+	return n
+}
+
+func abs32(n int32) int32 {
+	if n < 0 {
+		return -n
+	}
+	return n
+}
@@ -0,0 +1,241 @@
+package library
+
+import (
+	"math"
+	"math/rand/v2"
+	"reflect"
+	"testing"
+)
+
+// printLen is a realistic fingerprint length: fpcalc's 120s at ~8 items/second.
+const printLen = 960
+
+// randomPrint is a deterministic stand-in for one recording's fingerprint.
+func randomPrint(seed uint64, n int) []int32 {
+	r := rand.New(rand.NewPCG(seed, seed^0x9e3779b97f4a7c15))
+	fp := make([]int32, n)
+	for i := range fp {
+		fp[i] = int32(r.Uint32())
+	}
+	return fp
+}
+
+// withBitNoise flips exactly round(fraction × all bits) distinct bits — the
+// same recording through a different encoder, at a known bit-error rate.
+func withBitNoise(fp []int32, fraction float64, seed uint64) []int32 {
+	out := append([]int32(nil), fp...)
+	r := rand.New(rand.NewPCG(seed, seed^0x243f6a8885a308d3))
+	total := 32 * len(fp)
+	for _, pos := range r.Perm(total)[:int(math.Round(fraction*float64(total)))] {
+		out[pos/32] ^= int32(uint32(1) << (pos % 32))
+	}
+	return out
+}
+
+func constantPrint(v int32, n int) []int32 {
+	fp := make([]int32, n)
+	for i := range fp {
+		fp[i] = v
+	}
+	return fp
+}
+
+func TestCompareChromaprint(t *testing.T) {
+	base := randomPrint(1, printLen)
+
+	t.Run("identical", func(t *testing.T) {
+		got, ok := compareChromaprint(base, base)
+		if !ok || got.BitErrorRate != 0 || got.Offset != 0 || got.Overlap != printLen {
+			t.Fatalf("got %+v ok=%v, want an exact alignment", got, ok)
+		}
+	})
+
+	t.Run("re-encoded: known bit noise is measured exactly", func(t *testing.T) {
+		got, ok := compareChromaprint(base, withBitNoise(base, 0.03, 2))
+		if !ok {
+			t.Fatal("a re-encode was not comparable")
+		}
+		if want := math.Round(0.03*32*printLen) / (32 * printLen); got.BitErrorRate != want {
+			t.Fatalf("BitErrorRate = %v, want %v", got.BitErrorRate, want)
+		}
+	})
+
+	// b starts 40 items later in the same audio: b[j] = a[j+40], so a[i] aligns
+	// with b[i-40].
+	t.Run("offset inside the window is recovered", func(t *testing.T) {
+		got, ok := compareChromaprint(base, base[40:])
+		if !ok || got.Offset != -40 || got.BitErrorRate != 0 || got.Overlap != printLen-40 {
+			t.Fatalf("got %+v ok=%v, want offset -40 with no error", got, ok)
+		}
+	})
+
+	t.Run("offset beyond the window never matches", func(t *testing.T) {
+		got, ok := compareChromaprint(base, base[200:])
+		if ok && got.BitErrorRate <= defaultAcousticMaxBitErrorRate {
+			t.Fatalf("a 200-item shift matched: %+v", got)
+		}
+	})
+
+	t.Run("unrelated recordings sit near 0.5", func(t *testing.T) {
+		got, ok := compareChromaprint(base, randomPrint(99, printLen))
+		if ok && got.BitErrorRate < 0.4 {
+			t.Fatalf("unrelated fingerprints scored %v", got.BitErrorRate)
+		}
+	})
+
+	t.Run("too short an overlap gives no verdict", func(t *testing.T) {
+		if got, ok := compareChromaprint(base, base[:minOverlapItems-1]); ok {
+			t.Fatalf("a %d-item fingerprint was compared: %+v", minOverlapItems-1, got)
+		}
+	})
+
+	// Two near-silent tracks agree perfectly without being one recording. The
+	// information floor is the only thing standing between them and a merge.
+	t.Run("low-information fingerprints give no verdict", func(t *testing.T) {
+		silence := constantPrint(0x1234, printLen)
+		if got, ok := compareChromaprint(silence, silence); ok {
+			t.Fatalf("silence compared as a match: %+v", got)
+		}
+	})
+
+	t.Run("the threshold separates close from not close", func(t *testing.T) {
+		near, _ := compareChromaprint(base, withBitNoise(base, 0.10, 3))
+		far, _ := compareChromaprint(base, withBitNoise(base, 0.20, 4))
+		if near.BitErrorRate > defaultAcousticMaxBitErrorRate {
+			t.Errorf("10%% noise (%v) is over the threshold", near.BitErrorRate)
+		}
+		if far.BitErrorRate <= defaultAcousticMaxBitErrorRate {
+			t.Errorf("20%% noise (%v) is under the threshold", far.BitErrorRate)
+		}
+	})
+}
+
+func TestGroupDuplicates_ExactTier(t *testing.T) {
+	hash := []byte("sha256-of-www-instrumental-bytes")
+	res := groupDuplicates([]fingerprintCandidate{
+		{ID: "www-01", DurationMs: 215000, StreamSHA256: hash},
+		{ID: "www-02", DurationMs: 215000, StreamSHA256: hash},
+		{ID: "lovesick", DurationMs: 198000, StreamSHA256: []byte("another")},
+	}, defaultAcousticMaxBitErrorRate)
+	want := []duplicateGroup{{Tier: tierExact, Members: []string{"www-01", "www-02"}}}
+	if !reflect.DeepEqual(res.Groups, want) {
+		t.Fatalf("groups = %+v, want %+v", res.Groups, want)
+	}
+}
+
+func TestGroupDuplicates_AcousticPair(t *testing.T) {
+	p := randomPrint(10, printLen)
+	res := groupDuplicates([]fingerprintCandidate{
+		{ID: "album", DurationMs: 240000, Chromaprint: p},
+		{ID: "compilation", DurationMs: 241000, Chromaprint: withBitNoise(p, 0.05, 11)},
+	}, defaultAcousticMaxBitErrorRate)
+	if len(res.Groups) != 1 || res.Groups[0].Tier != tierAcoustic ||
+		!reflect.DeepEqual(res.Groups[0].Members, []string{"album", "compilation"}) {
+		t.Fatalf("groups = %+v, want one acoustic pair", res.Groups)
+	}
+	if got := res.Groups[0].WorstBitErrorRate; math.Abs(got-0.05) > 0.001 {
+		t.Fatalf("WorstBitErrorRate = %v, want about 0.05", got)
+	}
+}
+
+// A is close to B and B is close to C, but A and C are not close. Under
+// single linkage all three would be proposed as one recording; complete linkage
+// must keep C out.
+func TestGroupDuplicates_NoChaining(t *testing.T) {
+	a := randomPrint(20, printLen)
+	b := withBitNoise(a, 0.10, 21)
+	c := withBitNoise(b, 0.10, 22)
+	if s, _ := compareChromaprint(a, c); s.BitErrorRate <= defaultAcousticMaxBitErrorRate {
+		t.Fatalf("fixture broken: A and C are close (%v), so this cannot test chaining", s.BitErrorRate)
+	}
+	res := groupDuplicates([]fingerprintCandidate{
+		{ID: "a", DurationMs: 200000, Chromaprint: a},
+		{ID: "b", DurationMs: 200000, Chromaprint: b},
+		{ID: "c", DurationMs: 200000, Chromaprint: c},
+	}, defaultAcousticMaxBitErrorRate)
+	if len(res.Groups) != 1 || !reflect.DeepEqual(res.Groups[0].Members, []string{"a", "b"}) {
+		t.Fatalf("groups = %+v, want only {a, b}", res.Groups)
+	}
+}
+
+func TestGroupDuplicates_DurationTolerance(t *testing.T) {
+	p := randomPrint(30, printLen)
+	res := groupDuplicates([]fingerprintCandidate{
+		{ID: "edit", DurationMs: 200000, Chromaprint: p},
+		{ID: "extended", DurationMs: 200000 + durationToleranceMs + 1, Chromaprint: p},
+	}, defaultAcousticMaxBitErrorRate)
+	if len(res.Groups) != 0 {
+		t.Fatalf("tracks %dms apart were grouped: %+v", durationToleranceMs+1, res.Groups)
+	}
+}
+
+// Nine tracks that all match are far likelier a shared jingle than nine copies
+// of one recording. The cluster must be reported, not proposed.
+func TestGroupDuplicates_OversizeClusterIsDiscarded(t *testing.T) {
+	p := randomPrint(40, printLen)
+	var cands []fingerprintCandidate
+	for i := range maxAcousticGroupSize + 1 {
+		cands = append(cands, fingerprintCandidate{
+			ID: string(rune('a' + i)), DurationMs: 30000, Chromaprint: withBitNoise(p, 0.01, uint64(100+i)),
+		})
+	}
+	res := groupDuplicates(cands, defaultAcousticMaxBitErrorRate)
+	if len(res.Groups) != 0 || res.OversizeClusters != 1 {
+		t.Fatalf("groups = %+v, oversize = %d; want none proposed and 1 oversize", res.Groups, res.OversizeClusters)
+	}
+}
+
+// Two byte-identical copies plus a re-encode of the same recording are one
+// group, and it is only as certain as its weakest link.
+func TestGroupDuplicates_ExactGroupAbsorbedIntoAcoustic(t *testing.T) {
+	p := randomPrint(50, printLen)
+	hash := []byte("same-bytes")
+	res := groupDuplicates([]fingerprintCandidate{
+		{ID: "x1", DurationMs: 180000, StreamSHA256: hash, Chromaprint: p},
+		{ID: "x2", DurationMs: 180000, StreamSHA256: hash, Chromaprint: p},
+		{ID: "y", DurationMs: 180000, StreamSHA256: []byte("other-bytes"), Chromaprint: withBitNoise(p, 0.03, 51)},
+	}, defaultAcousticMaxBitErrorRate)
+	want := []string{"x1", "x2", "y"}
+	if len(res.Groups) != 1 || res.Groups[0].Tier != tierAcoustic || !reflect.DeepEqual(res.Groups[0].Members, want) {
+		t.Fatalf("groups = %+v, want one acoustic group %v", res.Groups, want)
+	}
+}
+
+func TestGroupDuplicates_UnrelatedTracksNeverGroup(t *testing.T) {
+	var cands []fingerprintCandidate
+	for i := range 6 {
+		cands = append(cands, fingerprintCandidate{
+			ID: string(rune('a' + i)), DurationMs: 210000, Chromaprint: randomPrint(uint64(60+i), printLen),
+		})
+	}
+	if res := groupDuplicates(cands, defaultAcousticMaxBitErrorRate); len(res.Groups) != 0 {
+		t.Fatalf("unrelated recordings were grouped: %+v", res.Groups)
+	}
+}
+
+func TestGroupDuplicates_OrderIndependent(t *testing.T) {
+	p := randomPrint(70, printLen)
+	q := randomPrint(71, printLen)
+	hash := []byte("identical")
+	cands := []fingerprintCandidate{
+		{ID: "p1", DurationMs: 200000, Chromaprint: p},
+		{ID: "p2", DurationMs: 201000, Chromaprint: withBitNoise(p, 0.04, 72)},
+		{ID: "q1", DurationMs: 150000, Chromaprint: q},
+		{ID: "q2", DurationMs: 150500, Chromaprint: withBitNoise(q, 0.02, 73)},
+		{ID: "h1", DurationMs: 90000, StreamSHA256: hash},
+		{ID: "h2", DurationMs: 90000, StreamSHA256: hash},
+		{ID: "lone", DurationMs: 200000, Chromaprint: randomPrint(74, printLen)},
+	}
+	want := groupDuplicates(cands, defaultAcousticMaxBitErrorRate)
+	if len(want.Groups) != 3 {
+		t.Fatalf("fixture broken: %d groups, want 3 (p, q, h)", len(want.Groups))
+	}
+	r := rand.New(rand.NewPCG(75, 76))
+	for range 20 {
+		shuffled := append([]fingerprintCandidate(nil), cands...)
+		r.Shuffle(len(shuffled), func(i, j int) { shuffled[i], shuffled[j] = shuffled[j], shuffled[i] })
+		if got := groupDuplicates(shuffled, defaultAcousticMaxBitErrorRate); !reflect.DeepEqual(got, want) {
+			t.Fatalf("input order changed the result:\n got  %+v\n want %+v", got, want)
+		}
+	}
+}
@@ -0,0 +1,309 @@
+package library
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sort"
+
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/pgtype"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/coverart"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	syncpkg "git.fabledsword.com/bvandeusen/minstrel/internal/sync"
+)
+
+// Duplicate merge (M400 #3911).
+
+// ErrDuplicateGroupNotPending means the group was already merged or dismissed,
+// no longer exists, or no longer has two members to merge.
+var ErrDuplicateGroupNotPending = errors.New("library: duplicate group is not pending")
+
+// ErrSurvivorNotInGroup means the copy chosen to keep is not a member of the group.
+var ErrSurvivorNotInGroup = errors.New("library: survivor is not a member of the group")
+
+// MergedCopy is one copy a merge kept or removed.
+type MergedCopy struct {
+	TrackID   pgtype.UUID
+	FilePath  string
+	TrackMbid *string
+	AlbumMbid *string
+}
+
+// MergeResult says what a merge did.
+type MergeResult struct {
+	Tier     string
+	Survivor MergedCopy
+	Removed  []MergedCopy
+
+	// What moved onto the survivor — reported so the operator, and the audit
+	// log, can see that the history was kept rather than take it on trust.
+	PlayEvents      int64
+	SkipEvents      int64
+	Likes           int // users whose like now sits on the survivor
+	PlaylistEntries int
+
+	DeletedAlbumIDs  []pgtype.UUID
+	DeletedArtistIDs []pgtype.UUID
+}
+
+// MergeDuplicateGroup keeps one copy of a duplicate group and removes the rest,
+// carrying everything the removed copies held onto the one kept.
+//
+// survivorID chooses the copy to keep; an invalid (zero) id takes the proposal
+// from ProposeSurvivor.
+//
+// The danger this is built around: every table referencing tracks does so ON
+// DELETE CASCADE, so deleting a duplicate's row outright silently destroys its
+// likes, plays, playlist entries and tags. The merge moves all of that onto the
+// survivor first, and only then deletes the now-empty row.
+//
+// It deletes the removed copies' FILES too, and first, before any row changes
+// (#3918, note #3926). A merge that left the file behind would be undone by the
+// next scan, which re-imports it as a new track with no history. If a file cannot
+// be removed, the *FileRemoveError comes back and nothing in the database changes.
+// With several copies to remove, one file may already be gone when a later one
+// fails; that copy's row keeps all its history and is marked missing by the next
+// scan, and retrying the merge picks up where it stopped.
+//
+// Everything else happens in one transaction, which holds a lock on the group so
+// two merges of it cannot run at once. Sync changes for clients' caches are logged
+// inside it, the way the playlists service logs its own.
+func MergeDuplicateGroup(
+	ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger, dataDir string,
+	groupID, survivorID pgtype.UUID,
+) (MergeResult, error) {
+	if logger == nil {
+		logger = slog.Default()
+	}
+	tx, err := pool.Begin(ctx)
+	if err != nil {
+		return MergeResult{}, fmt.Errorf("begin merge: %w", err)
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+	tq := dbq.New(tx)
+
+	group, err := tq.LockDuplicateGroupForMerge(ctx, groupID)
+	if errors.Is(err, pgx.ErrNoRows) {
+		return MergeResult{}, ErrDuplicateGroupNotPending
+	}
+	if err != nil {
+		return MergeResult{}, fmt.Errorf("lock duplicate group: %w", err)
+	}
+	if group.Status != "pending" {
+		return MergeResult{}, ErrDuplicateGroupNotPending
+	}
+
+	members, err := tq.ListDuplicateGroupMergeMembers(ctx, groupID)
+	if err != nil {
+		return MergeResult{}, fmt.Errorf("list group members: %w", err)
+	}
+	if len(members) < 2 {
+		return MergeResult{}, ErrDuplicateGroupNotPending
+	}
+	survivor, losers, err := splitSurvivor(members, survivorID)
+	if err != nil {
+		return MergeResult{}, err
+	}
+
+	for _, l := range losers {
+		if err := removeTrackFileOnDisk(l.FilePath); err != nil {
+			return MergeResult{}, err
+		}
+	}
+
+	res := MergeResult{Tier: group.Tier, Survivor: mergedCopyOf(survivor)}
+	likers := map[string]struct{}{}
+	changes := mergeChanges{}
+	survivorKey := syncpkg.FormatUUID(survivor.ID)
+
+	for _, l := range losers {
+		ids := struct{ survivor, loser pgtype.UUID }{survivor.ID, l.ID}
+		loserKey := syncpkg.FormatUUID(l.ID)
+
+		n, err := tq.MergeRepointPlayEvents(ctx, dbq.MergeRepointPlayEventsParams{SurvivorID: ids.survivor, LoserID: ids.loser})
+		if err != nil {
+			return MergeResult{}, fmt.Errorf("move play events: %w", err)
+		}
+		res.PlayEvents += n
+		n, err = tq.MergeRepointSkipEvents(ctx, dbq.MergeRepointSkipEventsParams{SurvivorID: ids.survivor, LoserID: ids.loser})
+		if err != nil {
+			return MergeResult{}, fmt.Errorf("move skip events: %w", err)
+		}
+		res.SkipEvents += n
+		if _, err := tq.MergeRepointContextualLikes(ctx, dbq.MergeRepointContextualLikesParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("move contextual likes: %w", err)
+		}
+		if _, err := tq.MergeRepointPlaybackErrors(ctx, dbq.MergeRepointPlaybackErrorsParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("move playback errors: %w", err)
+		}
+		if _, err := tq.MergeRepointLidarrRequests(ctx, dbq.MergeRepointLidarrRequestsParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("move lidarr request matches: %w", err)
+		}
+
+		playlists, err := tq.MergeRepointPlaylistTracks(ctx, dbq.MergeRepointPlaylistTracksParams{SurvivorID: ids.survivor, LoserID: ids.loser})
+		if err != nil {
+			return MergeResult{}, fmt.Errorf("move playlist entries: %w", err)
+		}
+		res.PlaylistEntries += len(playlists)
+		for _, pl := range playlists {
+			plKey := syncpkg.FormatUUID(pl)
+			changes.playlistDelete(syncpkg.EncodePlaylistTrackID(plKey, loserKey))
+			changes.playlistUpsert(syncpkg.EncodePlaylistTrackID(plKey, survivorKey))
+		}
+
+		users, err := tq.MergeCopyGeneralLikes(ctx, dbq.MergeCopyGeneralLikesParams{SurvivorID: ids.survivor, LoserID: ids.loser})
+		if err != nil {
+			return MergeResult{}, fmt.Errorf("move likes: %w", err)
+		}
+		for _, u := range users {
+			userKey := syncpkg.FormatUUID(u)
+			likers[userKey] = struct{}{}
+			changes.likeDelete(syncpkg.EncodeLikeID(userKey, loserKey))
+			changes.likeUpsert(syncpkg.EncodeLikeID(userKey, survivorKey))
+		}
+
+		if _, err := tq.MergeCopyTrackTags(ctx, dbq.MergeCopyTrackTagsParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("merge tags: %w", err)
+		}
+		if _, err := tq.MergeCopyTrackSimilarity(ctx, dbq.MergeCopyTrackSimilarityParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("merge similarity: %w", err)
+		}
+		if err := tq.MergeInheritTrackMbid(ctx, dbq.MergeInheritTrackMbidParams{SurvivorID: ids.survivor, LoserID: ids.loser}); err != nil {
+			return MergeResult{}, fmt.Errorf("inherit recording mbid: %w", err)
+		}
+
+		// Everything the loser carried now sits on the survivor, so the CASCADE
+		// this delete sets off has nothing left to destroy.
+		deleted, err := tq.DeleteTrack(ctx, l.ID)
+		if err != nil {
+			return MergeResult{}, fmt.Errorf("delete merged copy: %w", err)
+		}
+		tidied, err := tidyEmptiedAlbum(ctx, tq, deleted.AlbumID)
+		if err != nil {
+			return MergeResult{}, err
+		}
+		if tidied.AlbumID != nil {
+			res.DeletedAlbumIDs = append(res.DeletedAlbumIDs, *tidied.AlbumID)
+		}
+		if tidied.ArtistID != nil {
+			res.DeletedArtistIDs = append(res.DeletedArtistIDs, *tidied.ArtistID)
+		}
+		res.Removed = append(res.Removed, mergedCopyOf(l))
+		changes.trackDelete(loserKey)
+	}
+	res.Likes = len(likers)
+
+	marked, err := tq.MarkDuplicateGroupMerged(ctx, groupID)
+	if err != nil {
+		return MergeResult{}, fmt.Errorf("mark group merged: %w", err)
+	}
+	if marked != 1 {
+		return MergeResult{}, ErrDuplicateGroupNotPending
+	}
+	if err := changes.log(ctx, tx); err != nil {
+		return MergeResult{}, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return MergeResult{}, fmt.Errorf("commit merge: %w", err)
+	}
+
+	// After commit, like DeleteTrackFile: a leftover art directory is only disk.
+	if dataDir != "" {
+		for _, artistID := range res.DeletedArtistIDs {
+			if err := coverart.CleanupArtistArt(dataDir, artistID); err != nil {
+				logger.Warn("duplicate merge: artist-art cleanup failed",
+					"artist_id", syncpkg.FormatUUID(artistID), "err", err)
+			}
+		}
+	}
+	return res, nil
+}
+
+// splitSurvivor separates the copy to keep from the copies to remove. An
+// invalid survivorID takes ProposeSurvivor's choice.
+func splitSurvivor(
+	members []dbq.ListDuplicateGroupMergeMembersRow, survivorID pgtype.UUID,
+) (dbq.ListDuplicateGroupMergeMembersRow, []dbq.ListDuplicateGroupMergeMembersRow, error) {
+	want := ""
+	if survivorID.Valid {
+		want = syncpkg.FormatUUID(survivorID)
+	} else {
+		cands := make([]SurvivorCandidate, len(members))
+		for i, m := range members {
+			cands[i] = SurvivorCandidate{
+				TrackID: syncpkg.FormatUUID(m.ID), FileFormat: m.FileFormat, FileSize: m.FileSize, AddedAt: m.AddedAt.Time,
+			}
+		}
+		want, _ = ProposeSurvivor(cands)
+	}
+
+	var survivor dbq.ListDuplicateGroupMergeMembersRow
+	found := false
+	var losers []dbq.ListDuplicateGroupMergeMembersRow
+	for _, m := range members {
+		if syncpkg.FormatUUID(m.ID) == want {
+			survivor, found = m, true
+			continue
+		}
+		losers = append(losers, m)
+	}
+	if !found {
+		return dbq.ListDuplicateGroupMergeMembersRow{}, nil, ErrSurvivorNotInGroup
+	}
+	return survivor, losers, nil
+}
+
+func mergedCopyOf(m dbq.ListDuplicateGroupMergeMembersRow) MergedCopy {
+	return MergedCopy{TrackID: m.ID, FilePath: m.FilePath, TrackMbid: m.Mbid, AlbumMbid: m.AlbumMbid}
+}
+
+// mergeChanges collects the sync-log entries a merge owes clients' caches, each
+// once: a user who liked two removed copies still gets one upsert for the
+// survivor.
+type mergeChanges struct {
+	tracks, likeDeletes, likeUpserts, playlistDeletes, playlistUpserts map[string]struct{}
+}
+
+func addTo(set *map[string]struct{}, id string) {
+	if *set == nil {
+		*set = map[string]struct{}{}
+	}
+	(*set)[id] = struct{}{}
+}
+
+func (c *mergeChanges) trackDelete(id string)    { addTo(&c.tracks, id) }
+func (c *mergeChanges) likeDelete(id string)     { addTo(&c.likeDeletes, id) }
+func (c *mergeChanges) likeUpsert(id string)     { addTo(&c.likeUpserts, id) }
+func (c *mergeChanges) playlistDelete(id string) { addTo(&c.playlistDeletes, id) }
+func (c *mergeChanges) playlistUpsert(id string) { addTo(&c.playlistUpserts, id) }
+
+func (c *mergeChanges) log(ctx context.Context, tx pgx.Tx) error {
+	for _, entry := range []struct {
+		kind syncpkg.EntityType
+		ids  map[string]struct{}
+		op   syncpkg.Op
+	}{
+		{syncpkg.EntityTrack, c.tracks, syncpkg.OpDelete},
+		{syncpkg.EntityLikeTrack, c.likeDeletes, syncpkg.OpDelete},
+		{syncpkg.EntityLikeTrack, c.likeUpserts, syncpkg.OpUpsert},
+		{syncpkg.EntityPlaylistTrack, c.playlistDeletes, syncpkg.OpDelete},
+		{syncpkg.EntityPlaylistTrack, c.playlistUpserts, syncpkg.OpUpsert},
+	} {
+		if len(entry.ids) == 0 {
+			continue
+		}
+		ids := make([]string, 0, len(entry.ids))
+		for id := range entry.ids {
+			ids = append(ids, id)
+		}
+		sort.Strings(ids)
+		if err := syncpkg.LogChanges(ctx, tx, entry.kind, ids, entry.op); err != nil {
+			return fmt.Errorf("log merge changes: %w", err)
+		}
+	}
+	return nil
+}
@@ -0,0 +1,281 @@
+package library
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"syscall"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/dbtest"
+)
+
+// mergeFixture is a library with one duplicate pair carrying history on both
+// copies, and a neighbour track for similarity edges.
+type mergeFixture struct {
+	pool                  *pgxpool.Pool
+	keep, remove, other   dbq.Track
+	keepPath, removePath  string
+	groupID               pgtype.UUID
+	alice, bob            dbq.User
+	playlistID            pgtype.UUID
+	removePlaylistPos     int32
+	aliceEarlierLikeOnRem time.Time
+}
+
+func newMergeFixture(t *testing.T) mergeFixture {
+	t.Helper()
+	pool := newPool(t)
+	ctx := context.Background()
+	q := dbq.New(pool)
+	dir := t.TempDir()
+	f := mergeFixture{pool: pool}
+
+	f.keepPath = filepath.Join(dir, "keep.flac")
+	f.removePath = filepath.Join(dir, "remove.mp3")
+	for _, p := range []string{f.keepPath, f.removePath} {
+		if err := os.WriteFile(p, []byte("audio"), 0o644); err != nil {
+			t.Fatalf("write %s: %v", p, err)
+		}
+	}
+	var album dbq.Album
+	var artist dbq.Artist
+	f.keep, album, artist = seedTrack(t, pool, f.keepPath)
+	upsert := func(title, path string) dbq.Track {
+		t.Helper()
+		tr, err := q.UpsertTrack(ctx, dbq.UpsertTrackParams{
+			Title: title, AlbumID: album.ID, ArtistID: artist.ID,
+			DurationMs: 215000, FilePath: path, FileSize: 100, FileFormat: "mp3",
+		})
+		if err != nil {
+			t.Fatalf("track %s: %v", title, err)
+		}
+		return tr
+	}
+	f.remove = upsert("WWW (copy)", f.removePath)
+	f.other = upsert("Neighbour", filepath.Join(dir, "other.mp3"))
+
+	mustExec := func(sql string, args ...any) {
+		t.Helper()
+		if _, err := pool.Exec(ctx, sql, args...); err != nil {
+			t.Fatalf("exec %q: %v", sql, err)
+		}
+	}
+	// Only the copy being removed carries a recording MBID.
+	mustExec(`UPDATE tracks SET mbid = 'rec-www' WHERE id = $1`, f.remove.ID)
+
+	user := func(name string) dbq.User {
+		t.Helper()
+		u, err := q.CreateUser(ctx, dbq.CreateUserParams{
+			Username: dbtest.TestUserPrefix + name, PasswordHash: "x", ApiToken: name + "-merge-token",
+		})
+		if err != nil {
+			t.Fatalf("user %s: %v", name, err)
+		}
+		return u
+	}
+	f.alice, f.bob = user("merge-alice"), user("merge-bob")
+
+	// Alice liked both copies, the removed one first; Bob liked only the removed one.
+	f.aliceEarlierLikeOnRem = time.Now().Add(-72 * time.Hour).UTC().Truncate(time.Microsecond)
+	mustExec(`INSERT INTO general_likes (user_id, track_id, liked_at) VALUES ($1, $2, $3), ($1, $4, now()), ($5, $2, now())`,
+		f.alice.ID, f.remove.ID, f.aliceEarlierLikeOnRem, f.keep.ID, f.bob.ID)
+
+	now := pgtype.Timestamptz{Time: time.Now(), Valid: true}
+	session, err := q.InsertPlaySession(ctx, dbq.InsertPlaySessionParams{UserID: f.alice.ID, StartedAt: now})
+	if err != nil {
+		t.Fatalf("session: %v", err)
+	}
+	for _, track := range []dbq.Track{f.remove, f.remove, f.keep} {
+		if _, err := q.InsertPlayEvent(ctx, dbq.InsertPlayEventParams{
+			UserID: f.alice.ID, TrackID: track.ID, SessionID: session.ID, StartedAt: now,
+		}); err != nil {
+			t.Fatalf("play event: %v", err)
+		}
+	}
+	if _, err := q.InsertSkipEvent(ctx, dbq.InsertSkipEventParams{
+		UserID: f.alice.ID, TrackID: f.remove.ID, SessionID: session.ID, SkippedAt: now, PositionMs: 1000,
+	}); err != nil {
+		t.Fatalf("skip event: %v", err)
+	}
+
+	pl, err := q.CreatePlaylist(ctx, dbq.CreatePlaylistParams{UserID: f.alice.ID, Name: "merge-mix"})
+	if err != nil {
+		t.Fatalf("playlist: %v", err)
+	}
+	f.playlistID = pl.ID
+	entry, err := q.AppendPlaylistTrack(ctx, dbq.AppendPlaylistTrackParams{PlaylistID: pl.ID, TrackID: f.remove.ID})
+	if err != nil {
+		t.Fatalf("playlist entry: %v", err)
+	}
+	f.removePlaylistPos = entry.Position
+
+	mustExec(`INSERT INTO track_tags (track_id, tag, weight) VALUES ($1, 'j-pop', 1), ($1, 'house', 0.5), ($2, 'house', 0.9)`,
+		f.remove.ID, f.keep.ID)
+	mustExec(`INSERT INTO track_similarity (track_a_id, track_b_id, score, source) VALUES
+		($1, $3, 0.8, 'listenbrainz'),
+		($2, $3, 0.7, 'listenbrainz'),
+		($1, $2, 0.99, 'listenbrainz'),
+		($3, $1, 0.6, 'musicbrainz_tag')`, f.remove.ID, f.keep.ID, f.other.ID)
+
+	if err := pool.QueryRow(ctx,
+		`INSERT INTO duplicate_groups (member_key, tier) VALUES ('merge-fixture', 'exact') RETURNING id`,
+	).Scan(&f.groupID); err != nil {
+		t.Fatalf("group: %v", err)
+	}
+	mustExec(`INSERT INTO duplicate_group_members (group_id, track_id) VALUES ($1, $2), ($1, $3)`,
+		f.groupID, f.keep.ID, f.remove.ID)
+	return f
+}
+
+func (f mergeFixture) count(t *testing.T, sql string, args ...any) int {
+	t.Helper()
+	var n int
+	if err := f.pool.QueryRow(context.Background(), sql, args...).Scan(&n); err != nil {
+		t.Fatalf("count %q: %v", sql, err)
+	}
+	return n
+}
+
+// The #3911 proof: after a merge, every piece of history the removed copy held
+// is on the copy kept, nothing is doubled, and the removed copy — row and file —
+// is gone.
+func TestMergeDuplicateGroup_Integration(t *testing.T) {
+	f := newMergeFixture(t)
+	ctx := context.Background()
+
+	res, err := MergeDuplicateGroup(ctx, f.pool, nil, "", f.groupID, f.keep.ID)
+	if err != nil {
+		t.Fatalf("merge: %v", err)
+	}
+	if len(res.Removed) != 1 || res.Removed[0].FilePath != f.removePath || res.Survivor.TrackID != f.keep.ID {
+		t.Fatalf("result = %+v, want the removed copy reported and the kept one as survivor", res)
+	}
+	if res.PlayEvents != 2 || res.SkipEvents != 1 || res.Likes != 2 || res.PlaylistEntries != 1 {
+		t.Errorf("moved = plays %d skips %d likes %d playlist %d, want 2, 1, 2, 1",
+			res.PlayEvents, res.SkipEvents, res.Likes, res.PlaylistEntries)
+	}
+
+	if _, err := os.Stat(f.removePath); !errors.Is(err, os.ErrNotExist) {
+		t.Errorf("removed copy's file still on disk: %v", err)
+	}
+	if _, err := os.Stat(f.keepPath); err != nil {
+		t.Errorf("kept copy's file is gone: %v", err)
+	}
+	if n := f.count(t, `SELECT count(*) FROM tracks WHERE id = $1`, f.remove.ID); n != 0 {
+		t.Errorf("removed copy's row still exists")
+	}
+
+	// Likes: one per user, Alice's dated to her earlier like.
+	if n := f.count(t, `SELECT count(*) FROM general_likes WHERE track_id = $1`, f.keep.ID); n != 2 {
+		t.Errorf("likes on the kept copy = %d, want 2 (Alice once, Bob)", n)
+	}
+	var aliceLiked time.Time
+	if err := f.pool.QueryRow(ctx, `SELECT liked_at FROM general_likes WHERE user_id = $1 AND track_id = $2`,
+		f.alice.ID, f.keep.ID).Scan(&aliceLiked); err != nil {
+		t.Fatalf("alice's like: %v", err)
+	}
+	if !aliceLiked.Equal(f.aliceEarlierLikeOnRem) {
+		t.Errorf("alice's like dated %v, want her earlier like %v", aliceLiked, f.aliceEarlierLikeOnRem)
+	}
+
+	// Plays and skips move exactly: none lost, none invented.
+	if n := f.count(t, `SELECT count(*) FROM play_events WHERE track_id = $1`, f.keep.ID); n != 3 {
+		t.Errorf("plays on the kept copy = %d, want 3", n)
+	}
+	if n := f.count(t, `SELECT count(*) FROM skip_events WHERE track_id = $1`, f.keep.ID); n != 1 {
+		t.Errorf("skips on the kept copy = %d, want 1", n)
+	}
+
+	// The playlist entry stays where it was and now plays the kept copy.
+	if n := f.count(t, `SELECT count(*) FROM playlist_tracks WHERE playlist_id = $1 AND position = $2 AND track_id = $3`,
+		f.playlistID, f.removePlaylistPos, f.keep.ID); n != 1 {
+		t.Errorf("playlist entry at position %d does not point at the kept copy", f.removePlaylistPos)
+	}
+
+	// Tags are a union; the kept copy's own weight wins where both had the tag.
+	if n := f.count(t, `SELECT count(*) FROM track_tags WHERE track_id = $1`, f.keep.ID); n != 2 {
+		t.Errorf("tags on the kept copy = %d, want 2 (house, j-pop)", n)
+	}
+	if n := f.count(t, `SELECT count(*) FROM track_tags WHERE track_id = $1 AND tag = 'house' AND weight = 0.9`, f.keep.ID); n != 1 {
+		t.Errorf("the kept copy's own house weight was overwritten")
+	}
+
+	// Similarity: rewritten onto the kept copy, no duplicate edge, no self-edge,
+	// nothing left pointing at the removed copy.
+	if n := f.count(t, `SELECT count(*) FROM track_similarity WHERE track_a_id = $1 AND track_b_id = $2 AND source = 'listenbrainz'`,
+		f.keep.ID, f.other.ID); n != 1 {
+		t.Errorf("listenbrainz edge keep→other = %d rows, want exactly 1", n)
+	}
+	if n := f.count(t, `SELECT count(*) FROM track_similarity WHERE track_a_id = $1 AND track_b_id = $2 AND source = 'musicbrainz_tag'`,
+		f.other.ID, f.keep.ID); n != 1 {
+		t.Errorf("musicbrainz_tag edge other→keep was not carried over")
+	}
+	if n := f.count(t, `SELECT count(*) FROM track_similarity WHERE track_a_id = track_b_id`); n != 0 {
+		t.Errorf("a self-edge was written")
+	}
+
+	// The removed copy's recording MBID is inherited; the group is closed.
+	if n := f.count(t, `SELECT count(*) FROM tracks WHERE id = $1 AND mbid = 'rec-www'`, f.keep.ID); n != 1 {
+		t.Errorf("the kept copy did not inherit the recording MBID")
+	}
+	if n := f.count(t, `SELECT count(*) FROM duplicate_groups WHERE id = $1 AND status = 'merged' AND resolved_at IS NOT NULL`, f.groupID); n != 1 {
+		t.Errorf("group was not marked merged")
+	}
+
+	// A second merge of the same group is refused rather than repeated.
+	if _, err := MergeDuplicateGroup(ctx, f.pool, nil, "", f.groupID, f.keep.ID); !errors.Is(err, ErrDuplicateGroupNotPending) {
+		t.Errorf("second merge err = %v, want ErrDuplicateGroupNotPending", err)
+	}
+}
+
+// When the removed copy's file cannot go, nothing may change: its likes, plays
+// and row stay exactly where they were, and the group stays pending.
+func TestMergeDuplicateGroup_UnremovableFileChangesNothing(t *testing.T) {
+	f := newMergeFixture(t)
+	stubRemoveFile(t, func(name string) error {
+		return &fs.PathError{Op: "remove", Path: name, Err: syscall.EROFS}
+	})
+
+	_, err := MergeDuplicateGroup(context.Background(), f.pool, nil, "", f.groupID, f.keep.ID)
+	var fre *FileRemoveError
+	if !errors.As(err, &fre) || !fre.NotWritable() {
+		t.Fatalf("err = %v, want a not-writable *FileRemoveError", err)
+	}
+	if n := f.count(t, `SELECT count(*) FROM tracks WHERE id = $1`, f.remove.ID); n != 1 {
+		t.Errorf("the copy's row was deleted although its file was not")
+	}
+	if n := f.count(t, `SELECT count(*) FROM general_likes WHERE track_id = $1`, f.remove.ID); n != 2 {
+		t.Errorf("likes on the copy = %d, want both still there", n)
+	}
+	if n := f.count(t, `SELECT count(*) FROM play_events WHERE track_id = $1`, f.remove.ID); n != 2 {
+		t.Errorf("plays on the copy = %d, want both still there", n)
+	}
+	if n := f.count(t, `SELECT count(*) FROM duplicate_groups WHERE id = $1 AND status = 'pending'`, f.groupID); n != 1 {
+		t.Errorf("group left pending = false, want it still pending")
+	}
+}
+
+func TestMergeDuplicateGroup_SurvivorMustBeAMember(t *testing.T) {
+	f := newMergeFixture(t)
+	var stranger pgtype.UUID
+	stranger.Bytes[15], stranger.Valid = 0xEE, true
+
+	_, err := MergeDuplicateGroup(context.Background(), f.pool, nil, "", f.groupID, stranger)
+	if !errors.Is(err, ErrSurvivorNotInGroup) {
+		t.Fatalf("err = %v, want ErrSurvivorNotInGroup", err)
+	}
+	if _, err := os.Stat(f.removePath); err != nil {
+		t.Errorf("a refused merge removed a file: %v", err)
+	}
+	if n := f.count(t, `SELECT count(*) FROM duplicate_groups WHERE id = $1 AND status = 'pending'`, f.groupID); n != 1 {
+		t.Errorf("a refused merge changed the group")
+	}
+}
@@ -0,0 +1,77 @@
+package library
+
+import (
+	"sort"
+	"strings"
+	"time"
+)
+
+// SurvivorCandidate is what choosing which copy to keep needs to know about one
+// member of a duplicate group.
+type SurvivorCandidate struct {
+	TrackID    string
+	FileFormat string
+	FileSize   int64
+	AddedAt    time.Time
+}
+
+// losslessFormats are the scanned extensions that are lossless by definition.
+// m4a is left out on purpose: it holds either ALAC or AAC, and the scanner
+// records only the extension, so calling it lossless would sometimes prefer an
+// AAC copy over a FLAC one.
+var losslessFormats = map[string]bool{"flac": true, "wav": true}
+
+// ProposeSurvivor picks which copy of a duplicate group to keep, and gives the
+// reason in words the operator reads beside it. It is a default, not a verdict:
+// the report shows it and the merge (#3911) lets the operator choose another.
+//
+// In order:
+//  1. lossless over lossy — the one difference no later step can recover
+//  2. the larger file — for one recording at one duration that is the higher
+//     bitrate. The scanner does not record bitrate (tracks.bitrate is never
+//     filled), so file size is the signal that actually exists
+//  3. the copy in the library longest — the one most likely to carry the play
+//     history and likes, so the merge moves the least
+//  4. the lowest track id, so the choice is stable between page loads
+func ProposeSurvivor(cands []SurvivorCandidate) (trackID, reason string) {
+	if len(cands) == 0 {
+		return "", ""
+	}
+	ranked := append([]SurvivorCandidate(nil), cands...)
+	sort.SliceStable(ranked, func(i, j int) bool { return survivorBefore(ranked[i], ranked[j]) })
+	best := ranked[0]
+	if len(ranked) == 1 {
+		return best.TrackID, "the only copy"
+	}
+
+	// The reason names the first rule that separated the best copy from the
+	// runner-up — the rule that actually decided, not every rule it passed.
+	next := ranked[1]
+	switch {
+	case isLossless(best) != isLossless(next):
+		return best.TrackID, "lossless (" + strings.ToLower(best.FileFormat) + ")"
+	case best.FileSize != next.FileSize:
+		return best.TrackID, "largest file"
+	case !best.AddedAt.Equal(next.AddedAt):
+		return best.TrackID, "in the library longest"
+	default:
+		return best.TrackID, "copies are otherwise identical"
+	}
+}
+
+func survivorBefore(a, b SurvivorCandidate) bool {
+	if isLossless(a) != isLossless(b) {
+		return isLossless(a)
+	}
+	if a.FileSize != b.FileSize {
+		return a.FileSize > b.FileSize
+	}
+	if !a.AddedAt.Equal(b.AddedAt) {
+		return a.AddedAt.Before(b.AddedAt)
+	}
+	return a.TrackID < b.TrackID
+}
+
+func isLossless(c SurvivorCandidate) bool {
+	return losslessFormats[strings.ToLower(c.FileFormat)]
+}
@@ -0,0 +1,91 @@
+package library
+
+import (
+	"testing"
+	"time"
+)
+
+func TestProposeSurvivor(t *testing.T) {
+	older := time.Date(2025, 1, 1, 0, 0, 0, 0, time.UTC)
+	newer := older.Add(24 * time.Hour)
+	cases := []struct {
+		name       string
+		cands      []SurvivorCandidate
+		wantID     string
+		wantReason string
+	}{
+		{
+			// Lossless wins even against a much larger lossy file, and even
+			// when the lossy copy has been in the library longer.
+			name: "lossless beats larger and older",
+			cands: []SurvivorCandidate{
+				{TrackID: "mp3", FileFormat: "mp3", FileSize: 90_000_000, AddedAt: older},
+				{TrackID: "flac", FileFormat: "FLAC", FileSize: 30_000_000, AddedAt: newer},
+			},
+			wantID: "flac", wantReason: "lossless (flac)",
+		},
+		{
+			// m4a may be AAC; it must not outrank an mp3 just for being m4a.
+			name: "m4a is not treated as lossless",
+			cands: []SurvivorCandidate{
+				{TrackID: "m4a", FileFormat: "m4a", FileSize: 5_000_000, AddedAt: older},
+				{TrackID: "mp3", FileFormat: "mp3", FileSize: 9_000_000, AddedAt: newer},
+			},
+			wantID: "mp3", wantReason: "largest file",
+		},
+		{
+			name: "larger file wins among lossy copies",
+			cands: []SurvivorCandidate{
+				{TrackID: "128k", FileFormat: "mp3", FileSize: 3_400_000, AddedAt: older},
+				{TrackID: "320k", FileFormat: "mp3", FileSize: 8_600_000, AddedAt: newer},
+			},
+			wantID: "320k", wantReason: "largest file",
+		},
+		{
+			// The #3885 pair: identical audio, sizes equal but for the tags.
+			name: "the longest-standing copy wins when size ties",
+			cands: []SurvivorCandidate{
+				{TrackID: "www-02", FileFormat: "mp3", FileSize: 6_900_000, AddedAt: newer},
+				{TrackID: "www-01", FileFormat: "mp3", FileSize: 6_900_000, AddedAt: older},
+			},
+			wantID: "www-01", wantReason: "in the library longest",
+		},
+		{
+			name: "a full tie falls back to the lowest id, stably",
+			cands: []SurvivorCandidate{
+				{TrackID: "b", FileFormat: "mp3", FileSize: 1, AddedAt: older},
+				{TrackID: "a", FileFormat: "mp3", FileSize: 1, AddedAt: older},
+			},
+			wantID: "a", wantReason: "copies are otherwise identical",
+		},
+		{
+			name:   "one copy",
+			cands:  []SurvivorCandidate{{TrackID: "only", FileFormat: "mp3", FileSize: 1, AddedAt: older}},
+			wantID: "only", wantReason: "the only copy",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			id, reason := ProposeSurvivor(tc.cands)
+			if id != tc.wantID || reason != tc.wantReason {
+				t.Fatalf("ProposeSurvivor = (%q, %q), want (%q, %q)", id, reason, tc.wantID, tc.wantReason)
+			}
+		})
+	}
+}
+
+// The reason must name the rule that decided. Across three copies that is the
+// comparison between first and second place, not the first rule any pair
+// differs on: here the lossy copy differs from the others by format, but the
+// two FLACs are separated by size.
+func TestProposeSurvivor_ReasonIsTheDecidingRule(t *testing.T) {
+	at := time.Date(2025, 1, 1, 0, 0, 0, 0, time.UTC)
+	id, reason := ProposeSurvivor([]SurvivorCandidate{
+		{TrackID: "mp3", FileFormat: "mp3", FileSize: 99_000_000, AddedAt: at},
+		{TrackID: "flac-small", FileFormat: "flac", FileSize: 20_000_000, AddedAt: at},
+		{TrackID: "flac-big", FileFormat: "flac", FileSize: 40_000_000, AddedAt: at},
+	})
+	if id != "flac-big" || reason != "largest file" {
+		t.Fatalf("got (%q, %q), want (flac-big, largest file)", id, reason)
+	}
+}
@@ -0,0 +1,398 @@
+package library
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/pgtype"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	syncpkg "git.fabledsword.com/bvandeusen/minstrel/internal/sync"
+)
+
+// Duplicate sweep (M400 #3910).
+//
+// Reads fingerprints, runs them through the matcher and records what it proposes
+// in duplicate_groups. It never merges or deletes anything the operator has not
+// asked for: a group is a proposal, reviewed in the admin report (#3912).
+//
+// Blocked on duration, deliberately not on title: the #3885 pair are titled
+// "WWW" and "WWW (instrumental)", so a title block would have missed the case
+// that started the milestone. Candidates stream in (duration_ms, id) order and
+// the grouper holds only a few seconds of durations at a time.
+
+// duplicateCandidatePage is how many candidates one query returns. Each row
+// carries a ~4 KB fingerprint, so a page is about 2 MB.
+const duplicateCandidatePage = 500
+
+// duplicateSweepTick is how often the worker checks whether a sweep is due. The
+// operator's sweep interval (#3913) is the least time between sweeps; the tick
+// only bounds how late past it one starts. With nothing due, a tick is two cheap
+// aggregate queries.
+const duplicateSweepTick = time.Hour
+
+// sweepIntervalSlack absorbs the moment between a tick and the sweep it starts
+// stamping started_at. Without it a one-hour interval checked on a one-hour tick
+// would find the last sweep a moment under an hour old, and skip every other tick.
+const sweepIntervalSlack = 5 * time.Minute
+
+// staleDuplicateSweepThreshold is the age past which an in-flight sweep is
+// assumed dead — a crash mid-sweep leaves finished_at NULL for ever — and another
+// may start. Twice the library scan's threshold, because a sweep compares
+// fingerprints across the whole library and can legitimately run long on a big
+// one.
+const staleDuplicateSweepThreshold = 2 * time.Hour
+
+// duplicateSweepFinishTimeout bounds recording that a sweep ended. It runs on a
+// context detached from the sweep's own, so a sweep cancelled at shutdown still
+// closes its row rather than leaving it in flight until the reaper.
+const duplicateSweepFinishTimeout = 10 * time.Second
+
+// DuplicateSweepResult tallies one sweep.
+type DuplicateSweepResult struct {
+	Candidates int // tracks with a chromaprint that were streamed
+	Groups     int // groups the matcher found
+	Proposed   int // written as pending, new or refreshed
+	Suppressed int // not proposed: already dismissed or resolved by the operator
+	Retired    int // pending proposals this sweep did not find again, removed
+	Oversize   int // acoustic clusters too large to propose
+}
+
+// RunDuplicateSweep runs one sweep and records it in duplicate_sweeps. cfg is a
+// snapshot: one sweep applies one threshold and one length throughout.
+func RunDuplicateSweep(
+	ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger, cfg FingerprintSettings,
+) (DuplicateSweepResult, error) {
+	return runDuplicateSweep(ctx, pool, logger, cfg, duplicateCandidatePage)
+}
+
+func runDuplicateSweep(
+	ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger, cfg FingerprintSettings, pageSize int32,
+) (DuplicateSweepResult, error) {
+	q := dbq.New(pool)
+	sweep, err := q.StartDuplicateSweep(ctx)
+	if err != nil {
+		return DuplicateSweepResult{}, fmt.Errorf("start duplicate sweep: %w", err)
+	}
+
+	res, runErr := sweepDuplicates(ctx, q, sweep.ID, cfg, pageSize)
+
+	finishCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), duplicateSweepFinishTimeout)
+	defer cancel()
+	errMsg := ""
+	if runErr != nil {
+		errMsg = runErr.Error()
+	}
+	candidates, groups, oversize := int32(res.Candidates), int32(res.Groups), int32(res.Oversize)
+	if ferr := q.FinishDuplicateSweep(finishCtx, dbq.FinishDuplicateSweepParams{
+		ID: sweep.ID, Candidates: &candidates, GroupsFound: &groups, OversizeClusters: &oversize,
+		ErrorMessage: errMsg,
+	}); ferr != nil {
+		logger.Error("duplicate sweep: recording the end of the sweep failed", "err", ferr)
+		if runErr == nil {
+			runErr = fmt.Errorf("finish duplicate sweep: %w", ferr)
+		}
+	}
+
+	logger.Info("duplicate sweep complete",
+		"candidates", res.Candidates, "groups", res.Groups, "proposed", res.Proposed,
+		"suppressed", res.Suppressed, "retired", res.Retired, "oversize", res.Oversize, "err", runErr)
+	return res, runErr
+}
+
+func sweepDuplicates(
+	ctx context.Context, q *dbq.Queries, sweepID pgtype.UUID, cfg FingerprintSettings, pageSize int32,
+) (DuplicateSweepResult, error) {
+	var res DuplicateSweepResult
+
+	// Exact tier, library-wide, in one query.
+	exactRows, err := q.ListExactDuplicateHashes(ctx, fingerprintVersion)
+	if err != nil {
+		return res, fmt.Errorf("list exact duplicates: %w", err)
+	}
+	exactMembers := make([][]string, len(exactRows))
+	exactOf := map[string]int{}
+	for i, row := range exactRows {
+		exactMembers[i] = formatUUIDs(row.TrackIds)
+		for _, id := range exactMembers[i] {
+			exactOf[id] = i
+		}
+	}
+	exactSeen := make([]bool, len(exactRows))
+
+	// Acoustic tier, streamed in duration order. The first member of an exact
+	// group the stream meets stands in for the whole group; the rest are skipped.
+	// Only prints at the current length are streamed: a print at another length
+	// cannot be compared, and is waiting on the backfill to be re-derived.
+	grouper := newStreamGrouper(cfg.AcousticMaxBitErrorRate)
+	params := dbq.ListDuplicateCandidatesParams{
+		CurrentVersion:       fingerprintVersion,
+		ChromaprintLengthSec: cfg.ChromaprintLengthSec,
+		// Durations are never negative, and the all-zero uuid sorts first: every
+		// row is after this cursor. Valid must be true, or "> NULL" matches nothing.
+		AfterDurationMs: -1,
+		AfterID:         pgtype.UUID{Valid: true},
+		PageLimit:       pageSize,
+	}
+	for {
+		if err := ctx.Err(); err != nil {
+			return res, err
+		}
+		rows, err := q.ListDuplicateCandidates(ctx, params)
+		if err != nil {
+			return res, fmt.Errorf("list duplicate candidates: %w", err)
+		}
+		for _, row := range rows {
+			res.Candidates++
+			id := syncpkg.FormatUUID(row.ID)
+			unit := &groupUnit{ids: []string{id}, durationMs: row.DurationMs, sortKey: id}
+			if gi, ok := exactOf[id]; ok {
+				if exactSeen[gi] {
+					continue
+				}
+				exactSeen[gi] = true
+				unit.ids, unit.exact = exactMembers[gi], true
+			}
+			unit.print = preparePrint(row.Chromaprint)
+			grouper.push(unit)
+		}
+		if int32(len(rows)) < pageSize {
+			break
+		}
+		last := rows[len(rows)-1]
+		params.AfterDurationMs, params.AfterID = last.DurationMs, last.ID
+	}
+
+	found := grouper.finish()
+	// Exact groups none of whose members has a chromaprint never reached the
+	// stream. Identical bytes need no acoustic evidence.
+	for gi, seen := range exactSeen {
+		if !seen {
+			found.Groups = append(found.Groups, duplicateGroup{Tier: tierExact, Members: exactMembers[gi]})
+		}
+	}
+	sortGroups(found.Groups)
+	res.Groups, res.Oversize = len(found.Groups), found.OversizeClusters
+
+	dismissed, err := q.ListDismissedDuplicateMemberSets(ctx)
+	if err != nil {
+		return res, fmt.Errorf("list dismissed duplicate groups: %w", err)
+	}
+	dismissedSets := make([]map[string]struct{}, 0, len(dismissed))
+	for _, d := range dismissed {
+		set := map[string]struct{}{}
+		for _, id := range formatUUIDs(d.TrackIds) {
+			set[id] = struct{}{}
+		}
+		dismissedSets = append(dismissedSets, set)
+	}
+
+	for _, group := range found.Groups {
+		if coveredByDismissal(group.Members, dismissedSets) {
+			res.Suppressed++
+			continue
+		}
+		up := dbq.UpsertDuplicateGroupParams{
+			MemberKey: strings.Join(group.Members, ","),
+			Tier:      string(group.Tier),
+			SweepID:   sweepID,
+		}
+		if group.Tier == tierAcoustic {
+			worst := float32(group.WorstBitErrorRate)
+			up.WorstBitErrorRate = &worst
+		}
+		groupID, err := q.UpsertDuplicateGroup(ctx, up)
+		if errors.Is(err, pgx.ErrNoRows) {
+			// This exact member set was already dismissed or merged.
+			res.Suppressed++
+			continue
+		}
+		if err != nil {
+			return res, fmt.Errorf("upsert duplicate group: %w", err)
+		}
+		for _, id := range group.Members {
+			var trackID pgtype.UUID
+			if err := trackID.Scan(id); err != nil {
+				return res, fmt.Errorf("parse track id %q: %w", id, err)
+			}
+			if err := q.AddDuplicateGroupMember(ctx, dbq.AddDuplicateGroupMemberParams{
+				GroupID: groupID, TrackID: trackID,
+			}); err != nil {
+				return res, fmt.Errorf("add duplicate group member: %w", err)
+			}
+		}
+		res.Proposed++
+	}
+
+	// Only after a complete sweep: a sweep that failed partway has no basis for
+	// concluding that anything it did not reach has gone.
+	retired, err := q.DeleteStalePendingDuplicateGroups(ctx, sweepID)
+	if err != nil {
+		return res, fmt.Errorf("retire stale duplicate groups: %w", err)
+	}
+	res.Retired = int(retired)
+	return res, nil
+}
+
+// coveredByDismissal reports whether every member of a proposal sat together in
+// one group the operator dismissed. A subset counts: dismissing {A, B, C} said
+// none of them are copies of each other, so proposing {A, B} again would be
+// asking the same question twice. A superset does not count: a new copy joining
+// is new evidence, and worth asking about.
+func coveredByDismissal(members []string, dismissed []map[string]struct{}) bool {
+	for _, set := range dismissed {
+		covered := true
+		for _, id := range members {
+			if _, ok := set[id]; !ok {
+				covered = false
+				break
+			}
+		}
+		if covered {
+			return true
+		}
+	}
+	return false
+}
+
+func formatUUIDs(ids []pgtype.UUID) []string {
+	out := make([]string, len(ids))
+	for i, id := range ids {
+		out[i] = syncpkg.FormatUUID(id)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// TryStartDuplicateSweep starts a sweep in the background unless one is already
+// running, reaping a sweep that has been in flight past
+// staleDuplicateSweepThreshold. Mirrors TryStartScan. The sweep runs on ctx, so
+// a caller answering an HTTP request must pass a context that outlives it.
+func TryStartDuplicateSweep(
+	ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger, cfg FingerprintSettings,
+) (bool, error) {
+	q := dbq.New(pool)
+	row, err := q.GetInFlightDuplicateSweep(ctx)
+	switch {
+	case err == nil:
+		age := time.Since(row.StartedAt.Time)
+		if age <= staleDuplicateSweepThreshold {
+			return false, nil
+		}
+		logger.Warn("reaping stale duplicate sweep", "id", syncpkg.FormatUUID(row.ID), "age", age)
+		if ferr := q.FinishDuplicateSweep(ctx, dbq.FinishDuplicateSweepParams{
+			ID: row.ID, ErrorMessage: "reaped (stale)",
+		}); ferr != nil {
+			return false, fmt.Errorf("reap stale duplicate sweep: %w", ferr)
+		}
+	case !errors.Is(err, pgx.ErrNoRows):
+		return false, fmt.Errorf("duplicate sweep in-flight check: %w", err)
+	}
+
+	go func() {
+		if _, err := RunDuplicateSweep(ctx, pool, logger, cfg); err != nil {
+			logger.Warn("duplicate sweep failed", "err", err)
+		}
+	}()
+	return true, nil
+}
+
+// DuplicateSweepWorker sweeps whenever its input has changed, at most once per
+// the operator's sweep interval.
+type DuplicateSweepWorker struct {
+	pool     *pgxpool.Pool
+	logger   *slog.Logger
+	settings *FingerprintSettingsService
+	tick     time.Duration
+}
+
+// NewDuplicateSweepWorker builds a worker with the production cadence. settings
+// is shared with the admin API; nil runs on defaults.
+func NewDuplicateSweepWorker(
+	pool *pgxpool.Pool, logger *slog.Logger, settings *FingerprintSettingsService,
+) *DuplicateSweepWorker {
+	return &DuplicateSweepWorker{pool: pool, logger: logger, settings: settings, tick: duplicateSweepTick}
+}
+
+// Run blocks until ctx is cancelled, checking once at start and then each tick.
+func (w *DuplicateSweepWorker) Run(ctx context.Context) {
+	w.tickOnce(ctx)
+	t := time.NewTicker(w.tick)
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			w.tickOnce(ctx)
+		}
+	}
+}
+
+// tickOnce contains one check so nothing it does can stop the next tick (rule 157).
+func (w *DuplicateSweepWorker) tickOnce(ctx context.Context) {
+	defer func() {
+		if r := recover(); r != nil {
+			w.logger.Error("duplicate sweep: tick panicked", "panic", r)
+		}
+	}()
+	cfg := w.settings.Get()
+	due, err := duplicateSweepDue(ctx, dbq.New(w.pool), cfg, time.Now())
+	if err != nil {
+		if ctx.Err() == nil {
+			w.logger.Warn("duplicate sweep: due check failed", "err", err)
+		}
+		return
+	}
+	if !due {
+		return
+	}
+	if _, err := TryStartDuplicateSweep(ctx, w.pool, w.logger, cfg); err != nil {
+		w.logger.Warn("duplicate sweep: start failed", "err", err)
+	}
+}
+
+// duplicateSweepDue reads what sweepIsDue decides on.
+func duplicateSweepDue(ctx context.Context, q *dbq.Queries, cfg FingerprintSettings, now time.Time) (bool, error) {
+	latest, err := q.GetLatestFingerprintComputedAt(ctx)
+	if err != nil {
+		return false, fmt.Errorf("latest fingerprint: %w", err)
+	}
+	var lastStart pgtype.Timestamptz
+	last, err := q.GetLatestDuplicateSweep(ctx)
+	switch {
+	case err == nil:
+		lastStart = last.StartedAt
+	case !errors.Is(err, pgx.ErrNoRows):
+		return false, fmt.Errorf("latest duplicate sweep: %w", err)
+	}
+	return sweepIsDue(latest, lastStart, cfg, now), nil
+}
+
+// sweepIsDue reports whether a sweep should start: something it reads has
+// changed since the last sweep started, and the operator's interval has passed.
+//
+// Two things can change its answer. Fingerprints are its input, so any written
+// after the last sweep began count; while the backfill runs that is true every
+// tick, which is what the interval is for. And a settings save counts, because a
+// new threshold or length changes what the same fingerprints group into.
+func sweepIsDue(latestFingerprint, lastSweepStart pgtype.Timestamptz, cfg FingerprintSettings, now time.Time) bool {
+	if !latestFingerprint.Valid {
+		return false // nothing fingerprinted yet
+	}
+	if !lastSweepStart.Valid {
+		return true // never swept
+	}
+	interval := time.Duration(cfg.SweepIntervalHours) * time.Hour
+	if now.Sub(lastSweepStart.Time) < interval-sweepIntervalSlack {
+		return false
+	}
+	return latestFingerprint.Time.After(lastSweepStart.Time) || cfg.UpdatedAt.After(lastSweepStart.Time)
+}
@@ -0,0 +1,250 @@
+package library
+
+import (
+	"bytes"
+	"context"
+	"io"
+	"log/slog"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	syncpkg "git.fabledsword.com/bvandeusen/minstrel/internal/sync"
+)
+
+func TestCoveredByDismissal(t *testing.T) {
+	dismissed := []map[string]struct{}{{"a": {}, "b": {}, "c": {}}}
+	for _, tc := range []struct {
+		name    string
+		members []string
+		want    bool
+	}{
+		{"the same set", []string{"a", "b", "c"}, true},
+		{"a subset of it", []string{"a", "b"}, true},
+		// A new copy joining is new evidence: ask again.
+		{"a superset of it", []string{"a", "b", "c", "d"}, false},
+		{"overlapping only in part", []string{"a", "d"}, false},
+		{"unrelated", []string{"x", "y"}, false},
+	} {
+		if got := coveredByDismissal(tc.members, dismissed); got != tc.want {
+			t.Errorf("%s: coveredByDismissal = %v, want %v", tc.name, got, tc.want)
+		}
+	}
+}
+
+// TestDuplicateSweep_Integration pins what the sweep proposes, what it leaves out,
+// and how re-sweeping treats a dismissal and a proposal that no longer holds.
+func TestDuplicateSweep_Integration(t *testing.T) {
+	pool := newPool(t)
+	ctx := context.Background()
+	q := dbq.New(pool)
+	dir := t.TempDir()
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+
+	// seedTrack's own track has no fingerprint row: it must be absent from the
+	// report, not grouped with every other track lacking one.
+	_, album, artist := seedTrack(t, pool, filepath.Join(dir, "unfingerprinted.mp3"))
+	hash := func(b byte) []byte { return bytes.Repeat([]byte{b}, 32) }
+	add := func(name string, durationMs int32, sum []byte, print []int32) string {
+		t.Helper()
+		tr, err := q.UpsertTrack(ctx, dbq.UpsertTrackParams{
+			Title: name, AlbumID: album.ID, ArtistID: artist.ID, DurationMs: durationMs,
+			FilePath: filepath.Join(dir, name+".mp3"), FileSize: 100, FileFormat: "mp3",
+		})
+		if err != nil {
+			t.Fatalf("track %s: %v", name, err)
+		}
+		if err := q.UpsertTrackFingerprint(ctx, dbq.UpsertTrackFingerprintParams{
+			TrackID: tr.ID, AudioStreamSha256: sum, Chromaprint: print, FingerprintVersion: fingerprintVersion,
+			ChromaprintLengthSec: defaultChromaprintLengthSec,
+		}); err != nil {
+			t.Fatalf("fingerprint %s: %v", name, err)
+		}
+		return syncpkg.FormatUUID(tr.ID)
+	}
+	key := func(ids ...string) string {
+		sorted := append([]string(nil), ids...)
+		sort.Strings(sorted)
+		return strings.Join(sorted, ",")
+	}
+
+	recording := randomPrint(200, printLen)
+	onAlbum := add("recording-album", 240000, hash(1), recording)
+	onCompilation := add("recording-compilation", 241000, hash(2), withBitNoise(recording, 0.05, 201))
+	www1 := add("www-01", 215000, hash(9), randomPrint(210, printLen))
+	www2 := add("www-02", 215000, hash(9), randomPrint(210, printLen))
+	// Near-identical duration to the recording, different audio.
+	add("different-song", 240500, hash(3), randomPrint(220, printLen))
+	// Identical to the album copy, but its file is gone: nothing to compare.
+	missing := add("missing-copy", 240000, hash(4), recording)
+	if _, err := pool.Exec(ctx, "UPDATE tracks SET missing_since = now() WHERE file_path LIKE '%missing-copy.mp3'"); err != nil {
+		t.Fatalf("mark missing: %v", err)
+	}
+
+	type stored struct {
+		tier, status string
+	}
+	groups := func() map[string]stored {
+		t.Helper()
+		rows, err := pool.Query(ctx, `SELECT member_key, tier, status FROM duplicate_groups`)
+		if err != nil {
+			t.Fatalf("read groups: %v", err)
+		}
+		defer rows.Close()
+		out := map[string]stored{}
+		for rows.Next() {
+			var k string
+			var s stored
+			if err := rows.Scan(&k, &s.tier, &s.status); err != nil {
+				t.Fatalf("scan group: %v", err)
+			}
+			out[k] = s
+		}
+		return out
+	}
+
+	// 1. A page size of one forces the keyset cursor across every candidate.
+	res, err := runDuplicateSweep(ctx, pool, logger, DefaultFingerprintSettings, 1)
+	if err != nil {
+		t.Fatalf("first sweep: %v", err)
+	}
+	// Five tracks carry a chromaprint and a present file.
+	if res.Candidates != 5 || res.Groups != 2 || res.Proposed != 2 {
+		t.Fatalf("first sweep = %+v, want 5 candidates, 2 groups, 2 proposed", res)
+	}
+	acousticKey, exactKey := key(onAlbum, onCompilation), key(www1, www2)
+	got := groups()
+	want := map[string]stored{
+		acousticKey: {"acoustic", "pending"},
+		exactKey:    {"exact", "pending"},
+	}
+	if len(got) != len(want) || got[acousticKey] != want[acousticKey] || got[exactKey] != want[exactKey] {
+		t.Fatalf("groups = %+v, want %+v", got, want)
+	}
+	for k := range got {
+		if strings.Contains(k, missing) {
+			t.Fatalf("a missing track was proposed: %s", k)
+		}
+	}
+
+	// 2. A dismissed group is not proposed again, and the pending one is
+	// refreshed in place rather than duplicated.
+	if _, err := pool.Exec(ctx, "UPDATE duplicate_groups SET status = 'dismissed' WHERE member_key = $1", acousticKey); err != nil {
+		t.Fatalf("dismiss: %v", err)
+	}
+	res, err = runDuplicateSweep(ctx, pool, logger, DefaultFingerprintSettings, duplicateCandidatePage)
+	if err != nil {
+		t.Fatalf("second sweep: %v", err)
+	}
+	if res.Proposed != 1 || res.Suppressed != 1 {
+		t.Fatalf("second sweep = %+v, want 1 proposed, 1 suppressed", res)
+	}
+	got = groups()
+	if len(got) != 2 || got[acousticKey].status != "dismissed" || got[exactKey].status != "pending" {
+		t.Fatalf("after dismissal groups = %+v, want the dismissal kept and one pending group", got)
+	}
+
+	// 3. A proposal that no longer holds is retired; the dismissal survives it.
+	if _, err := pool.Exec(ctx,
+		"DELETE FROM track_fingerprints f USING tracks t WHERE f.track_id = t.id AND t.file_path LIKE '%www-02.mp3'"); err != nil {
+		t.Fatalf("drop fingerprint: %v", err)
+	}
+	res, err = runDuplicateSweep(ctx, pool, logger, DefaultFingerprintSettings, duplicateCandidatePage)
+	if err != nil {
+		t.Fatalf("third sweep: %v", err)
+	}
+	if res.Retired != 1 {
+		t.Fatalf("third sweep = %+v, want 1 retired", res)
+	}
+	got = groups()
+	if len(got) != 1 || got[acousticKey].status != "dismissed" {
+		t.Fatalf("after retiring groups = %+v, want only the dismissal", got)
+	}
+
+	// 4. The sweep record reflects the last run.
+	last, err := q.GetLatestDuplicateSweep(ctx)
+	if err != nil {
+		t.Fatalf("latest sweep: %v", err)
+	}
+	if !last.FinishedAt.Valid || last.ErrorMessage != nil {
+		t.Fatalf("latest sweep = %+v, want finished without error", last)
+	}
+
+	// 5. Prints taken at another length are never compared (#3913). Every print
+	// here was taken at the default length, so a sweep at 60s has nothing to read,
+	// rather than scoring 120s prints against each other as if they were 60s ones.
+	atOtherLength := DefaultFingerprintSettings
+	atOtherLength.ChromaprintLengthSec = 60
+	res, err = runDuplicateSweep(ctx, pool, logger, atOtherLength, duplicateCandidatePage)
+	if err != nil {
+		t.Fatalf("sweep at another length: %v", err)
+	}
+	if res.Candidates != 0 || res.Groups != 0 {
+		t.Fatalf("sweep at another length = %+v, want no candidates and no groups", res)
+	}
+
+	// 6. The sweep applies the threshold it is given. The recording's two copies
+	// disagree on about 5% of their bits: grouped at the default, not at 1%.
+	if _, err := pool.Exec(ctx, "DELETE FROM duplicate_groups"); err != nil {
+		t.Fatalf("clear groups: %v", err)
+	}
+	strict := DefaultFingerprintSettings
+	strict.AcousticMaxBitErrorRate = 0.01
+	res, err = runDuplicateSweep(ctx, pool, logger, strict, duplicateCandidatePage)
+	if err != nil {
+		t.Fatalf("strict sweep: %v", err)
+	}
+	if res.Groups != 0 {
+		t.Fatalf("sweep at a 1%% threshold = %+v, want the copies 5%% apart left ungrouped", res)
+	}
+	res, err = runDuplicateSweep(ctx, pool, logger, DefaultFingerprintSettings, duplicateCandidatePage)
+	if err != nil {
+		t.Fatalf("default sweep: %v", err)
+	}
+	if got := groups(); res.Groups != 1 || got[acousticKey] != (stored{"acoustic", "pending"}) {
+		t.Fatalf("sweep at the default threshold = %+v, groups %+v; want the recording's copies proposed", res, got)
+	}
+}
+
+func TestSweepIsDue(t *testing.T) {
+	now := time.Date(2026, 9, 11, 12, 0, 0, 0, time.UTC)
+	at := func(ago time.Duration) pgtype.Timestamptz {
+		return pgtype.Timestamptz{Time: now.Add(-ago), Valid: true}
+	}
+	never := pgtype.Timestamptz{}
+	hourly := DefaultFingerprintSettings
+	daily := DefaultFingerprintSettings
+	daily.SweepIntervalHours = 24
+	savedAgo := func(ago time.Duration) FingerprintSettings {
+		s := DefaultFingerprintSettings
+		s.UpdatedAt = now.Add(-ago)
+		return s
+	}
+	for _, tc := range []struct {
+		name                   string
+		latestPrint, lastSweep pgtype.Timestamptz
+		cfg                    FingerprintSettings
+		want                   bool
+	}{
+		{"nothing fingerprinted", never, never, hourly, false},
+		{"never swept", at(time.Minute), never, hourly, true},
+		{"new fingerprints since the last sweep", at(10 * time.Minute), at(2 * time.Hour), hourly, true},
+		{"nothing new since the last sweep", at(3 * time.Hour), at(2 * time.Hour), hourly, false},
+		// The sweep started a moment after the previous tick, so one tick later
+		// it is a moment under an hour old. Without the slack this is false.
+		{"one tick after an hourly sweep", at(time.Minute), at(time.Hour - 2*time.Second), hourly, true},
+		{"new fingerprints inside the interval", at(time.Minute), at(3 * time.Hour), daily, false},
+		{"new fingerprints past the interval", at(time.Minute), at(25 * time.Hour), daily, true},
+		{"settings saved since the last sweep", at(3 * time.Hour), at(2 * time.Hour), savedAgo(time.Hour), true},
+		{"settings saved before the last sweep", at(3 * time.Hour), at(2 * time.Hour), savedAgo(4 * time.Hour), false},
+	} {
+		if got := sweepIsDue(tc.latestPrint, tc.lastSweep, tc.cfg, now); got != tc.want {
+			t.Errorf("%s: sweepIsDue = %v, want %v", tc.name, got, tc.want)
+		}
+	}
+}
@@ -0,0 +1,339 @@
+package library
+
+import (
+	"bytes"
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"log/slog"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+)
+
+// Acoustic identity (M400).
+//
+// Two values per track, because they answer different questions:
+//
+//	audio_stream_sha256  a SHA-256 of the ENCODED audio packets. Equal means the
+//	                     same audio bytes, whatever the tags or container around
+//	                     them say. No threshold and no false positives — this is
+//	                     what catches two copies of one MP3 that differ only in
+//	                     their ID3 (#3885).
+//
+//	chromaprint          fpcalc's raw fingerprint. Close means the same
+//	                     recording, even at another bitrate or in another codec
+//	                     — the case an exact hash cannot see.
+//
+// Both shell out, in the shape probeDurationMs already set: a deadline on every
+// call, and a failure that leaves the value unset rather than failing the file.
+// A track with no fingerprint is never a duplicate candidate; it is still a
+// track.
+
+// fingerprintTimeout bounds one ffmpeg hash or fpcalc call. Longer than
+// probeTimeout because both read the audio rather than a header: the hash reads
+// every packet and fpcalc decodes up to its -length. 60s leaves room for a large
+// lossless file on a slow network mount; a call needing more is a stall, not a
+// big file.
+const fingerprintTimeout = 60 * time.Second
+
+// fingerprintWaitDelay bounds how long Output may keep waiting on the tool's
+// pipes after the deadline has killed it. Without it, a child that left a
+// descendant holding stdout open would block the scan past its own timeout.
+const fingerprintWaitDelay = 5 * time.Second
+
+// fingerprintVersion stamps how a track_fingerprints row was derived. Bump it
+// whenever the derivation changes — the hash arguments or fpcalc's flags — and
+// the backfill re-derives every row below it. Fingerprints taken by two methods
+// are not comparable, and nothing else would reveal that the library held a mix.
+//
+// The length is deliberately not part of it: it is an operator setting (#3913),
+// so each row records the length it was taken at and readers compare only rows
+// at the current one. See fingerprint_settings.go.
+const fingerprintVersion int16 = 1
+
+// errFingerprintTimeout marks a tool that ran out of time. Distinct from a
+// failed exit because a stall is a fact about the mount, not about the file.
+var errFingerprintTimeout = errors.New("fingerprint tool timed out")
+
+// defaultChromaprintLengthSec is the shipped value of the length setting (#3913):
+// how many seconds of audio fpcalc fingerprints. 120 is fpcalc's own default.
+const defaultChromaprintLengthSec = 120
+
+// errChromaprintSkipped marks a chromaprint not taken because fingerprinting is
+// switched off. Inconclusive rather than a verdict: nothing was learned about
+// the file.
+var errChromaprintSkipped = errors.New("chromaprint skipped: fingerprinting is off")
+
+// fpcalcStderrTail caps how much of a failing tool's stderr reaches the log.
+const fpcalcStderrTail = 512
+
+// streamHashArgs hashes the encoded audio packets, never decoded samples.
+//
+// -c:a copy is the point, not an optimisation. A decoded hash of a lossy file
+// depends on the decoder's float maths and sample conversion, which can move
+// between ffmpeg releases — so an image upgrade could silently change every
+// stored hash, and yesterday's duplicate would stop matching today's copy.
+// Packet bytes do not move. It is also far cheaper: demux only, no decode.
+//
+// -map 0:a keeps embedded cover art (an attached-picture video stream) out of
+// the hash, so two copies of one recording carrying different art still match.
+func streamHashArgs(path string) []string {
+	return []string{
+		"-v", "error",
+		"-i", path,
+		"-map", "0:a",
+		"-c:a", "copy",
+		"-f", "hash", "-hash", "sha256",
+		"-",
+	}
+}
+
+// fpcalcArgs asks for the raw fingerprint as SIGNED integers.
+//
+// -raw because the matcher compares items bit by bit, which the compressed form
+// cannot do without being unpacked first. -signed because the column is Postgres
+// integer[], which is signed: fpcalc's default prints uint32, and half of those
+// values do not fit. Signed output is the same 32 bits with no reinterpretation
+// step left to get wrong.
+func fpcalcArgs(path string, lengthSec int) []string {
+	return []string{
+		"-raw", "-signed",
+		"-length", strconv.Itoa(lengthSec),
+		path,
+	}
+}
+
+// fingerprintResult is one attempt at both halves of a track's identity. They
+// fail independently: a file ffmpeg can demux may still defeat fpcalc.
+type fingerprintResult struct {
+	streamSHA256 []byte
+	chromaprint  []int32
+	hashErr      error
+	printErr     error
+}
+
+// fingerprintOptions is what the settings decide for one attempt. Captured once
+// per file, so the length a chromaprint was taken at is the length stored with it
+// even if the setting changes mid-attempt.
+type fingerprintOptions struct {
+	lengthSec int32
+	// chromaprint false takes the stream hash alone: a demux, no decode.
+	chromaprint bool
+}
+
+// computeFingerprint derives the halves opts asks for, for the file at path.
+func computeFingerprint(ctx context.Context, path string, opts fingerprintOptions) fingerprintResult {
+	var r fingerprintResult
+	r.streamSHA256, r.hashErr = computeAudioStreamSHA256(ctx, path)
+	if !opts.chromaprint {
+		r.printErr = errChromaprintSkipped
+		return r
+	}
+	r.chromaprint, r.printErr = computeChromaprint(ctx, path, opts.lengthSec)
+	return r
+}
+
+// inconclusive reports whether either half failed for a reason that says
+// nothing about the file. Such a result must never be stored: stamped at the
+// current version it would read as "tried, and this file cannot be
+// fingerprinted", and the backfill would never try it again.
+func (r fingerprintResult) inconclusive() bool {
+	return isInconclusive(r.hashErr) || isInconclusive(r.printErr)
+}
+
+// isInconclusive names the failures that are not a verdict on the file: a
+// stall, a cancelled scan, a tool that is not installed, and a chromaprint
+// skipped because fingerprinting is off. A missing tool matters outside the
+// image — a dev binary run without fpcalc on PATH must not stamp every track in
+// the library as unfingerprintable.
+func isInconclusive(err error) bool {
+	return errors.Is(err, errFingerprintTimeout) ||
+		errors.Is(err, errChromaprintSkipped) ||
+		errors.Is(err, context.Canceled) ||
+		errors.Is(err, context.DeadlineExceeded) ||
+		errors.Is(err, exec.ErrNotFound)
+}
+
+// fingerprintFile runs the scanner's fingerprinter. A Scanner built without New
+// gets the real tools rather than a nil-func panic halfway through a scan.
+func (s *Scanner) fingerprintFile(ctx context.Context, path string, opts fingerprintOptions) fingerprintResult {
+	if s.fingerprint == nil {
+		return computeFingerprint(ctx, path, opts)
+	}
+	return s.fingerprint(ctx, path, opts)
+}
+
+// fingerprintOutcome is what storeFingerprint did with one attempt.
+type fingerprintOutcome int
+
+const (
+	outcomeFingerprinted fingerprintOutcome = iota // both halves stored
+	outcomeRejected                                // stored with a NULL half: a verdict
+	outcomeInconclusive                            // nothing stored; worth trying again
+	outcomeStoreFailed                             // the write itself failed
+)
+
+// storeFingerprint records one attempt, for the scan (new or changed bytes) and
+// the backfill (#3908) alike, so there is one rule for what gets written. It
+// never fails its caller: a missing fingerprint only keeps a track out of
+// duplicate detection, which is not worth dropping a scan or a pass over.
+//
+// lengthSec is the length fp's chromaprint was taken at, stored with it (#3913).
+func storeFingerprint(
+	ctx context.Context, q *dbq.Queries, logger *slog.Logger,
+	trackID pgtype.UUID, path string, fp fingerprintResult, lengthSec int32,
+) fingerprintOutcome {
+	if fp.hashErr != nil {
+		logger.Warn("fingerprint: audio stream hash failed", "path", path, "err", fp.hashErr)
+	}
+	if fp.printErr != nil {
+		logger.Warn("fingerprint: chromaprint failed", "path", path, "err", fp.printErr)
+	}
+	if fp.inconclusive() {
+		// Any row this track holds describes bytes we could not confirm — the
+		// previous bytes for the scan, an older derivation for the backfill.
+		// Drop it rather than stamp a failure that says nothing about the file.
+		if err := q.DeleteTrackFingerprint(ctx, trackID); err != nil {
+			logger.Warn("fingerprint: clearing stale fingerprint failed", "path", path, "err", err)
+		}
+		return outcomeInconclusive
+	}
+	// A NULL half here is a verdict — the tool ran and rejected this file — and
+	// is stamped at the current version so the backfill does not retry it on
+	// every pass. It is retried when the file changes.
+	if err := q.UpsertTrackFingerprint(ctx, dbq.UpsertTrackFingerprintParams{
+		TrackID:              trackID,
+		AudioStreamSha256:    fp.streamSHA256,
+		Chromaprint:          fp.chromaprint,
+		FingerprintVersion:   fingerprintVersion,
+		ChromaprintLengthSec: lengthSec,
+	}); err != nil {
+		logger.Warn("fingerprint: storing fingerprint failed", "path", path, "err", err)
+		return outcomeStoreFailed
+	}
+	if fp.hashErr != nil || fp.printErr != nil {
+		return outcomeRejected
+	}
+	return outcomeFingerprinted
+}
+
+// computeAudioStreamSHA256 returns the SHA-256 of the file's encoded audio.
+func computeAudioStreamSHA256(ctx context.Context, path string) ([]byte, error) {
+	out, err := runFingerprintTool(ctx, "ffmpeg", streamHashArgs(path))
+	if err != nil {
+		return nil, err
+	}
+	return parseStreamHash(out)
+}
+
+// computeChromaprint returns the raw acoustic fingerprint of the first
+// lengthSec seconds of the file.
+func computeChromaprint(ctx context.Context, path string, lengthSec int32) ([]int32, error) {
+	out, err := runFingerprintTool(ctx, "fpcalc", fpcalcArgs(path, int(lengthSec)))
+	if err != nil {
+		return nil, err
+	}
+	return parseFpcalcRaw(out)
+}
+
+// runFingerprintTool runs one tool under fingerprintTimeout.
+//
+// Any non-zero exit is an error, and that deliberately includes fpcalc's exit 3:
+// "reading failed, but here is a fingerprint of what I got". A partial
+// fingerprint of a damaged file is not that file's identity. Stored, it would
+// score against a healthy copy over whatever prefix survived, and could group
+// or fail to group either way. Absent is better than wrong.
+func runFingerprintTool(ctx context.Context, name string, args []string) ([]byte, error) {
+	runCtx, cancel := context.WithTimeout(ctx, fingerprintTimeout)
+	defer cancel()
+
+	cmd := exec.CommandContext(runCtx, name, args...)
+	cmd.WaitDelay = fingerprintWaitDelay
+	out, err := cmd.Output()
+	if err == nil {
+		return out, nil
+	}
+	// The caller gave up (a cancelled scan). Report that rather than the
+	// signal-killed exit it caused, so it is never mistaken for a verdict on
+	// the file.
+	if ctx.Err() != nil {
+		return nil, fmt.Errorf("%s: %w", name, ctx.Err())
+	}
+	// Named separately so a stall reads as a stall, not as a crash.
+	if errors.Is(runCtx.Err(), context.DeadlineExceeded) {
+		return nil, fmt.Errorf("%s: no result within %s: %w", name, fingerprintTimeout, errFingerprintTimeout)
+	}
+	var exitErr *exec.ExitError
+	if errors.As(err, &exitErr) {
+		return nil, fmt.Errorf("%s exited %d: %s", name, exitErr.ExitCode(), stderrTail(exitErr.Stderr))
+	}
+	return nil, fmt.Errorf("%s: %w", name, err)
+}
+
+// stderrTail keeps the END of a failing tool's stderr. ffmpeg and fpcalc print
+// the actual reason last, after any banner or per-frame warnings, so a cap that
+// kept the head would log the noise and drop the cause.
+func stderrTail(stderr []byte) []byte {
+	stderr = bytes.TrimSpace(stderr)
+	if len(stderr) > fpcalcStderrTail {
+		stderr = stderr[len(stderr)-fpcalcStderrTail:]
+	}
+	return stderr
+}
+
+// parseStreamHash reads the ffmpeg hash muxer's "SHA256=<hex>" line.
+func parseStreamHash(out []byte) ([]byte, error) {
+	for _, line := range strings.Split(string(out), "\n") {
+		hexed, ok := strings.CutPrefix(strings.TrimSpace(line), "SHA256=")
+		if !ok {
+			continue
+		}
+		sum, err := hex.DecodeString(hexed)
+		if err != nil {
+			return nil, fmt.Errorf("stream hash %q: %w", hexed, err)
+		}
+		if len(sum) != sha256.Size {
+			return nil, fmt.Errorf("stream hash is %d bytes, want %d", len(sum), sha256.Size)
+		}
+		return sum, nil
+	}
+	return nil, errors.New("ffmpeg printed no SHA256= line")
+}
+
+// parseFpcalcRaw reads fpcalc's text output:
+//
+//	DURATION=<seconds>
+//	FINGERPRINT=<int32>,<int32>,...
+func parseFpcalcRaw(out []byte) ([]int32, error) {
+	for _, line := range strings.Split(string(out), "\n") {
+		list, ok := strings.CutPrefix(strings.TrimSpace(line), "FINGERPRINT=")
+		if !ok {
+			continue
+		}
+		if list == "" {
+			return nil, errors.New("fpcalc returned an empty fingerprint")
+		}
+		items := strings.Split(list, ",")
+		fp := make([]int32, len(items))
+		for i, item := range items {
+			// ParseInt at 32 bits, not ParseUint: a value past int32 means the
+			// output was unsigned — -signed went missing from the invocation —
+			// and nothing downstream would reinterpret it. Refuse it here.
+			v, err := strconv.ParseInt(item, 10, 32)
+			if err != nil {
+				return nil, fmt.Errorf("fingerprint item %d %q: %w", i, item, err)
+			}
+			fp[i] = int32(v)
+		}
+		return fp, nil
+	}
+	return nil, errors.New("fpcalc printed no FINGERPRINT= line")
+}
@@ -0,0 +1,224 @@
+package library
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"sync"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+)
+
+// Fingerprint backfill (M400 #3908).
+//
+// The scan fingerprints only bytes it has not seen (see scanFile), so every track
+// imported before fingerprinting existed — and every row derived by an older
+// fingerprintVersion — needs a pass of its own. That pass is this worker.
+//
+// Its own worker rather than a stage in RunScan, for two reasons, both about
+// time:
+//   - RunScan runs at boot and then every safetyNetScanInterval (12h), and an
+//     in-flight scan older than StuckScanThreshold (1h) is reaped and a second
+//     one started beside it. A stage would have to stop well inside the hour — a
+//     few hundred decodes — so a 50k-track library would take about a month.
+//   - A long stage holds the scan run in flight, and a manual rescan answers 409
+//     for as long as it runs.
+//
+// Progress is read live (FingerprintCoverage, the admin gauge) rather than from a
+// scan_runs tally: the work spans many passes with no single run to attach to.
+
+// fingerprintBackfillTick is how often the worker looks for work. Once the
+// library has caught up, a tick is one indexed query; mostly the hour bounds how
+// long a file that timed out on a slow mount waits before it is tried again.
+const fingerprintBackfillTick = time.Hour
+
+// fingerprintBackfillBatch is how many tracks one query hands the worker. Small,
+// so tracks the scan adds mid-pass are not stuck behind one enormous page.
+const fingerprintBackfillBatch = 50
+
+// fingerprintBackfillConcurrency is the shipped value of the concurrency setting
+// (#3913): how many files are decoded at once. Two is deliberately low: fpcalc
+// and the stream hash compete with playback transcoding for CPU and with
+// streaming for the mount, and a backfill that makes playback stutter is worse
+// than one that takes longer.
+const fingerprintBackfillConcurrency = 2
+
+// BackfillFingerprintsResult tallies one pass.
+type BackfillFingerprintsResult struct {
+	Processed     int
+	Fingerprinted int // both halves stored
+	Rejected      int // stored with a NULL half: a tool refused the file (settled)
+	Inconclusive  int // nothing stored; tried again on a later pass
+}
+
+func (r *BackfillFingerprintsResult) add(o fingerprintOutcome) {
+	r.Processed++
+	switch o {
+	case outcomeFingerprinted:
+		r.Fingerprinted++
+	case outcomeRejected:
+		r.Rejected++
+	default:
+		r.Inconclusive++
+	}
+}
+
+// FingerprintBackfillWorker fingerprints the tracks the scan never will.
+type FingerprintBackfillWorker struct {
+	pool     *pgxpool.Pool
+	logger   *slog.Logger
+	settings *FingerprintSettingsService
+	tick     time.Duration
+	batch    int32
+	// fingerprint is a field for the same reason as Scanner.fingerprint: an
+	// integration test pins which tracks a pass touches, not what the tools print.
+	fingerprint func(ctx context.Context, path string, opts fingerprintOptions) fingerprintResult
+}
+
+// NewFingerprintBackfillWorker builds a worker with the production cadence.
+// settings is shared with the scanner and the admin API; nil runs on defaults.
+func NewFingerprintBackfillWorker(
+	pool *pgxpool.Pool, logger *slog.Logger, settings *FingerprintSettingsService,
+) *FingerprintBackfillWorker {
+	return &FingerprintBackfillWorker{
+		pool:        pool,
+		logger:      logger,
+		settings:    settings,
+		tick:        fingerprintBackfillTick,
+		batch:       fingerprintBackfillBatch,
+		fingerprint: computeFingerprint,
+	}
+}
+
+// Run blocks until ctx is cancelled: one pass at start, so a fresh deploy does
+// not sit idle for an hour, then one per tick.
+func (w *FingerprintBackfillWorker) Run(ctx context.Context) {
+	w.runOnce(ctx)
+	t := time.NewTicker(w.tick)
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			w.runOnce(ctx)
+		}
+	}
+}
+
+// runOnce contains a pass so that nothing it does — an error, a panic — can stop
+// the next tick from firing (rule 157).
+func (w *FingerprintBackfillWorker) runOnce(ctx context.Context) {
+	defer func() {
+		if r := recover(); r != nil {
+			w.logger.Error("fingerprint backfill: pass panicked", "panic", r)
+		}
+	}()
+	res, err := w.pass(ctx)
+	if err != nil && ctx.Err() == nil {
+		w.logger.Warn("fingerprint backfill: pass failed", "err", err, "processed", res.Processed)
+	}
+	if res.Processed > 0 {
+		w.logger.Info("fingerprint backfill: pass complete",
+			"processed", res.Processed, "fingerprinted", res.Fingerprinted,
+			"rejected", res.Rejected, "inconclusive", res.Inconclusive)
+	}
+}
+
+// pass walks every track needing a fingerprint once, keyset-paged on id. The
+// cursor is what lets a pass end: an inconclusive attempt writes no row, so a
+// file that keeps timing out would otherwise be listed again immediately and
+// retried forever within the pass.
+//
+// Settings are read before every batch, so a save takes effect within a batch
+// rather than an hour (#3913): switching fingerprinting off ends the pass, a new
+// concurrency applies to the next batch, and a new length restarts the walk from
+// the top at that length, because every row written at the old one went stale
+// the moment it changed.
+func (w *FingerprintBackfillWorker) pass(ctx context.Context) (BackfillFingerprintsResult, error) {
+	q := dbq.New(w.pool)
+	var (
+		res BackfillFingerprintsResult
+		mu  sync.Mutex
+	)
+	// The all-zero uuid sorts before every real id. Valid must be true: a NULL
+	// cursor would make "id > NULL" match nothing and every pass a silent no-op.
+	start := pgtype.UUID{Valid: true}
+	after := start
+	lengthSec := w.settings.Get().ChromaprintLengthSec
+	for {
+		if err := ctx.Err(); err != nil {
+			return res, err
+		}
+		cfg := w.settings.Get()
+		if !cfg.Enabled {
+			return res, nil
+		}
+		if cfg.ChromaprintLengthSec != lengthSec {
+			lengthSec, after = cfg.ChromaprintLengthSec, start
+		}
+		rows, err := q.ListTracksNeedingFingerprint(ctx, dbq.ListTracksNeedingFingerprintParams{
+			CurrentVersion:       fingerprintVersion,
+			ChromaprintLengthSec: lengthSec,
+			AfterID:              after,
+			BatchLimit:           w.batch,
+		})
+		if err != nil {
+			return res, fmt.Errorf("list tracks needing fingerprint: %w", err)
+		}
+		if len(rows) == 0 {
+			return res, nil
+		}
+
+		opts := fingerprintOptions{lengthSec: lengthSec, chromaprint: true}
+		// Validation keeps concurrency at one or more; the floor guards a zero
+		// that would block the first send for ever.
+		sem := make(chan struct{}, max(1, int(cfg.BackfillConcurrency)))
+		var wg sync.WaitGroup
+		for _, row := range rows {
+			if ctx.Err() != nil {
+				break
+			}
+			sem <- struct{}{}
+			wg.Add(1)
+			go func(trackID pgtype.UUID, path string) {
+				defer wg.Done()
+				defer func() { <-sem }()
+				defer func() {
+					if r := recover(); r != nil {
+						w.logger.Error("fingerprint backfill: track panicked", "path", path, "panic", r)
+					}
+				}()
+				outcome := storeFingerprint(ctx, q, w.logger, trackID, path, w.fingerprintFile(ctx, path, opts), lengthSec)
+				mu.Lock()
+				res.add(outcome)
+				mu.Unlock()
+			}(row.ID, row.FilePath)
+		}
+		wg.Wait()
+		after = rows[len(rows)-1].ID
+	}
+}
+
+func (w *FingerprintBackfillWorker) fingerprintFile(ctx context.Context, path string, opts fingerprintOptions) fingerprintResult {
+	if w.fingerprint == nil {
+		return computeFingerprint(ctx, path, opts)
+	}
+	return w.fingerprint(ctx, path, opts)
+}
+
+// FingerprintCoverage reports how much of the library carries a current
+// fingerprint, for the admin gauge. It lives here, beside the backfill, so the
+// version and length it counts against are the ones the backfill writes.
+func FingerprintCoverage(
+	ctx context.Context, pool *pgxpool.Pool, cfg FingerprintSettings,
+) (dbq.GetFingerprintCoverageRow, error) {
+	return dbq.New(pool).GetFingerprintCoverage(ctx, dbq.GetFingerprintCoverageParams{
+		CurrentVersion:       fingerprintVersion,
+		ChromaprintLengthSec: cfg.ChromaprintLengthSec,
+	})
+}
@@ -0,0 +1,226 @@
+package library
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"maps"
+	"path/filepath"
+	"sync"
+	"testing"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+)
+
+// TestFingerprintBackfill_Integration pins which tracks a pass touches, that a
+// pass ends, and that the coverage gauge counts what the pass wrote.
+func TestFingerprintBackfill_Integration(t *testing.T) {
+	pool := newPool(t)
+	ctx := context.Background()
+	q := dbq.New(pool)
+	dir := t.TempDir()
+
+	_, album, artist := seedTrack(t, pool, filepath.Join(dir, "unfingerprinted.mp3"))
+	addTrack := func(name string) dbq.Track {
+		t.Helper()
+		tr, err := q.UpsertTrack(ctx, dbq.UpsertTrackParams{
+			Title: name, AlbumID: album.ID, ArtistID: artist.ID,
+			DurationMs: 1000, FilePath: filepath.Join(dir, name+".mp3"), FileSize: 100, FileFormat: "mp3",
+		})
+		if err != nil {
+			t.Fatalf("track %s: %v", name, err)
+		}
+		return tr
+	}
+	current := addTrack("current")
+	stale := addTrack("stale")
+	missing := addTrack("missing")
+
+	sum := bytes.Repeat([]byte{0xCD}, 32)
+	for _, seed := range []struct {
+		track   dbq.Track
+		version int16
+	}{
+		{current, fingerprintVersion},
+		{stale, fingerprintVersion - 1},
+	} {
+		if err := q.UpsertTrackFingerprint(ctx, dbq.UpsertTrackFingerprintParams{
+			TrackID: seed.track.ID, AudioStreamSha256: sum, Chromaprint: []int32{1},
+			FingerprintVersion: seed.version, ChromaprintLengthSec: defaultChromaprintLengthSec,
+		}); err != nil {
+			t.Fatalf("seed fingerprint: %v", err)
+		}
+	}
+	if _, err := pool.Exec(ctx, "UPDATE tracks SET missing_since = now() WHERE id = $1", missing.ID); err != nil {
+		t.Fatalf("mark missing: %v", err)
+	}
+
+	var mu sync.Mutex
+	calls := map[string]int{}
+	settings, err := NewFingerprintSettingsService(ctx, pool)
+	if err != nil {
+		t.Fatalf("fingerprint settings: %v", err)
+	}
+	t.Cleanup(func() {
+		if _, err := settings.Set(context.Background(), DefaultFingerprintSettings); err != nil {
+			t.Errorf("restore fingerprint settings: %v", err)
+		}
+	})
+	w := NewFingerprintBackfillWorker(pool, slog.New(slog.NewTextHandler(io.Discard, nil)), settings)
+	// A batch of one forces the keyset cursor across several queries in a pass.
+	w.batch = 1
+	lengths := map[int32]int{}
+	w.fingerprint = func(_ context.Context, path string, opts fingerprintOptions) fingerprintResult {
+		name := filepath.Base(path)
+		mu.Lock()
+		calls[name]++
+		lengths[opts.lengthSec]++
+		mu.Unlock()
+		switch name {
+		case "stall.mp3":
+			return fingerprintResult{streamSHA256: sum, printErr: fmt.Errorf("fpcalc: %w", errFingerprintTimeout)}
+		case "rejected.mp3":
+			return fingerprintResult{hashErr: errors.New("ffmpeg exited 1"), printErr: errors.New("fpcalc exited 2")}
+		default:
+			return fingerprintResult{streamSHA256: sum, chromaprint: []int32{7, -7}}
+		}
+	}
+	callCount := func(name string) int {
+		mu.Lock()
+		defer mu.Unlock()
+		return calls[name]
+	}
+
+	// 1. Only the track with no row and the stale one are fingerprinted — never
+	// the current one, never the missing one.
+	res, err := w.pass(ctx)
+	if err != nil {
+		t.Fatalf("first pass: %v", err)
+	}
+	if res.Processed != 2 || res.Fingerprinted != 2 {
+		t.Fatalf("first pass = %+v, want 2 processed, 2 fingerprinted", res)
+	}
+	for name, want := range map[string]int{
+		"unfingerprinted.mp3": 1, "stale.mp3": 1, "current.mp3": 0, "missing.mp3": 0,
+	} {
+		if got := callCount(name); got != want {
+			t.Errorf("%s fingerprinted %d times, want %d", name, got, want)
+		}
+	}
+
+	// 2. A pass after a complete one is a no-op. A backfill that redoes its work
+	// every hour is the expensive way this could be wrong.
+	res, err = w.pass(ctx)
+	if err != nil {
+		t.Fatalf("second pass: %v", err)
+	}
+	if res.Processed != 0 {
+		t.Fatalf("second pass processed %d tracks, want 0", res.Processed)
+	}
+
+	// 3. An inconclusive file is tried exactly once and the pass ENDS. Without the
+	// keyset cursor it would be re-listed immediately and this call would never
+	// return.
+	addTrack("stall")
+	addTrack("rejected")
+	res, err = w.pass(ctx)
+	if err != nil {
+		t.Fatalf("third pass: %v", err)
+	}
+	if res.Processed != 2 || res.Inconclusive != 1 || res.Rejected != 1 {
+		t.Fatalf("third pass = %+v, want 2 processed, 1 inconclusive, 1 rejected", res)
+	}
+	if got := callCount("stall.mp3"); got != 1 {
+		t.Fatalf("stalling file tried %d times in one pass, want exactly 1", got)
+	}
+
+	// 4. The gauge counts what the passes wrote, and its buckets add up.
+	cov, err := FingerprintCoverage(ctx, pool, settings.Get())
+	if err != nil {
+		t.Fatalf("coverage: %v", err)
+	}
+	// Five present tracks: unfingerprinted, current, stale, stall, rejected.
+	// The missing track is not counted.
+	if cov.Total != 5 || cov.Fingerprinted != 3 || cov.Rejected != 1 || cov.Pending != 1 {
+		t.Errorf("coverage = %+v, want total 5, fingerprinted 3, rejected 1, pending 1", cov)
+	}
+	if cov.Fingerprinted+cov.Rejected+cov.Pending != cov.Total {
+		t.Errorf("coverage buckets %+v do not sum to the total", cov)
+	}
+
+	// 5. Changing the length (#3913) makes every stored row stale at once — the
+	// gauge shows the whole library pending before the backfill has touched a
+	// file — and the next pass re-derives each at the new length. The failure
+	// this prevents is invisible from the UI: prints at two lengths that silently
+	// never match.
+	shorter := DefaultFingerprintSettings
+	shorter.ChromaprintLengthSec = 60
+	if _, err := settings.Set(ctx, shorter); err != nil {
+		t.Fatalf("change length: %v", err)
+	}
+	cov, err = FingerprintCoverage(ctx, pool, settings.Get())
+	if err != nil {
+		t.Fatalf("coverage after length change: %v", err)
+	}
+	if cov.Total != 5 || cov.Pending != 5 {
+		t.Fatalf("coverage after length change = %+v, want all 5 tracks pending", cov)
+	}
+	mu.Lock()
+	clear(lengths)
+	mu.Unlock()
+	res, err = w.pass(ctx)
+	if err != nil {
+		t.Fatalf("new-length pass: %v", err)
+	}
+	if res.Processed != 5 {
+		t.Fatalf("new-length pass = %+v, want all 5 present tracks re-derived", res)
+	}
+	mu.Lock()
+	asked := maps.Clone(lengths)
+	mu.Unlock()
+	if len(asked) != 1 || asked[60] != 5 {
+		t.Fatalf("new-length pass asked for lengths %v, want 60s for all 5", asked)
+	}
+	var atOldLength int
+	if err := pool.QueryRow(ctx,
+		"SELECT count(*) FROM track_fingerprints WHERE chromaprint_length_sec <> 60").Scan(&atOldLength); err != nil {
+		t.Fatalf("count old-length rows: %v", err)
+	}
+	if atOldLength != 0 {
+		t.Fatalf("%d fingerprints are still at the old length after a complete pass", atOldLength)
+	}
+
+	// 6. Switched off, the backfill does nothing, even with work waiting.
+	off := DefaultFingerprintSettings
+	off.Enabled = false
+	if _, err := settings.Set(ctx, off); err != nil {
+		t.Fatalf("switch fingerprinting off: %v", err)
+	}
+	addTrack("later")
+	res, err = w.pass(ctx)
+	if err != nil {
+		t.Fatalf("pass with fingerprinting off: %v", err)
+	}
+	if res.Processed != 0 || callCount("later.mp3") != 0 {
+		t.Fatalf("pass with fingerprinting off = %+v (later.mp3 tried %d times), want nothing done",
+			res, callCount("later.mp3"))
+	}
+}
+
+func TestBackfillFingerprintsResult_Add(t *testing.T) {
+	var r BackfillFingerprintsResult
+	for _, o := range []fingerprintOutcome{
+		outcomeFingerprinted, outcomeFingerprinted, outcomeRejected, outcomeInconclusive, outcomeStoreFailed,
+	} {
+		r.add(o)
+	}
+	// A failed write stored nothing, so like an inconclusive attempt it is
+	// tried again next pass — and counts as such.
+	want := BackfillFingerprintsResult{Processed: 5, Fingerprinted: 2, Rejected: 1, Inconclusive: 2}
+	if r != want {
+		t.Errorf("tally = %+v, want %+v", r, want)
+	}
+}
@@ -0,0 +1,230 @@
+package library
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db"
+)
+
+// TestScanner_FingerprintsOnlyNewOrChangedBytes_Integration pins WHEN the scan
+// fingerprints. The cost of getting it wrong is asymmetric and invisible: a
+// scan that re-fingerprints unchanged files still produces correct rows, just
+// by decoding the entire library on every tag-repair pass.
+//
+// The fingerprinter is stubbed. CI has no real audio, and the tools' output is
+// covered by the parser tests; this covers the scan's decisions.
+func TestScanner_FingerprintsOnlyNewOrChangedBytes_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping scanner integration in -short mode")
+	}
+	dsn := os.Getenv("MINSTREL_TEST_DATABASE_URL")
+	if dsn == "" {
+		t.Skip("MINSTREL_TEST_DATABASE_URL not set")
+	}
+	ctx := context.Background()
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+
+	if err := db.Migrate(dsn, logger); err != nil {
+		t.Fatalf("migrate: %v", err)
+	}
+	pool, err := pgxpool.New(ctx, dsn)
+	if err != nil {
+		t.Fatalf("pool: %v", err)
+	}
+	t.Cleanup(pool.Close)
+	if _, err := pool.Exec(ctx, "TRUNCATE tracks, albums, artists RESTART IDENTITY CASCADE"); err != nil {
+		t.Fatalf("truncate: %v", err)
+	}
+
+	root := t.TempDir()
+	a := filepath.Join(root, "artist/album/01.mp3")
+	b := filepath.Join(root, "artist/album/02.mp3")
+	writeTestMP3(t, a, map[string]string{"TIT2": "One", "TPE1": "Artist", "TALB": "Album", "TRCK": "1"})
+	writeTestMP3(t, b, map[string]string{"TIT2": "Two", "TPE1": "Artist", "TALB": "Album", "TRCK": "2"})
+
+	sum := bytes.Repeat([]byte{0xAB}, 32)
+	chroma := []int32{7, -7, 2147483647}
+	result := fingerprintResult{streamSHA256: sum, chromaprint: chroma}
+	calls := map[string]int{}
+
+	settings, err := NewFingerprintSettingsService(ctx, pool)
+	if err != nil {
+		t.Fatalf("fingerprint settings: %v", err)
+	}
+	if _, err := settings.Set(ctx, DefaultFingerprintSettings); err != nil {
+		t.Fatalf("reset fingerprint settings: %v", err)
+	}
+	t.Cleanup(func() {
+		if _, err := settings.Set(context.Background(), DefaultFingerprintSettings); err != nil {
+			t.Errorf("restore fingerprint settings: %v", err)
+		}
+	})
+
+	var lastOpts fingerprintOptions
+	scanner := New(pool, logger, []string{root}, settings)
+	scanner.fingerprint = func(_ context.Context, path string, opts fingerprintOptions) fingerprintResult {
+		calls[path]++
+		lastOpts = opts
+		return result
+	}
+
+	scan := func(step string) Stats {
+		t.Helper()
+		st, err := scanner.Scan(ctx, nil)
+		if err != nil {
+			t.Fatalf("%s: scan: %v", step, err)
+		}
+		return st
+	}
+	type row struct {
+		sha     []byte
+		chroma  []int32
+		version int16
+		length  int32
+	}
+	stored := func(path string) (row, bool) {
+		t.Helper()
+		var r row
+		err := pool.QueryRow(ctx, `
+			SELECT f.audio_stream_sha256, f.chromaprint, f.fingerprint_version, f.chromaprint_length_sec
+			  FROM track_fingerprints f JOIN tracks t ON t.id = f.track_id
+			 WHERE t.file_path = $1`, path).Scan(&r.sha, &r.chroma, &r.version, &r.length)
+		if errors.Is(err, pgx.ErrNoRows) {
+			return row{}, false
+		}
+		if err != nil {
+			t.Fatalf("read fingerprint for %s: %v", path, err)
+		}
+		return r, true
+	}
+	// A later step moves mtime forward past the row's updated_at, which is
+	// what the scan reads as "these bytes changed".
+	touch := func(path string, ahead time.Duration) {
+		t.Helper()
+		when := time.Now().Add(ahead)
+		if err := os.Chtimes(path, when, when); err != nil {
+			t.Fatalf("chtimes %s: %v", path, err)
+		}
+	}
+
+	// 1. New files are fingerprinted, and stored at the current version.
+	scan("first scan")
+	if calls[a] != 1 || calls[b] != 1 {
+		t.Fatalf("first scan fingerprint calls = %v, want one per file", calls)
+	}
+	got, ok := stored(a)
+	if !ok {
+		t.Fatal("first scan stored no fingerprint")
+	}
+	if !bytes.Equal(got.sha, sum) || !slices.Equal(got.chroma, chroma) || got.version != fingerprintVersion ||
+		got.length != defaultChromaprintLengthSec {
+		t.Fatalf("stored %+v, want sha %x chromaprint %v version %d length %d",
+			got, sum, chroma, fingerprintVersion, defaultChromaprintLengthSec)
+	}
+	if !lastOpts.chromaprint || lastOpts.lengthSec != defaultChromaprintLengthSec {
+		t.Fatalf("first scan asked for %+v, want a chromaprint at the default length", lastOpts)
+	}
+
+	// 2. A tag-repair pass re-reads every unchanged file and must not
+	// fingerprint any of them again.
+	//
+	// The Updated count is what makes this able to fail. Without it, a scan
+	// that simply SKIPPED both files would also leave the call counts at one,
+	// and the assertion would pass without the re-read path ever running.
+	if _, err := pool.Exec(ctx, "UPDATE tracks SET duration_ms = 1000, tag_read_version = 0"); err != nil {
+		t.Fatalf("force tag re-read: %v", err)
+	}
+	if st := scan("tag-repair scan"); st.Updated != 2 || st.Skipped != 0 {
+		t.Fatalf("tag-repair scan stats = %+v, want both files re-read (Updated=2 Skipped=0)", st)
+	}
+	if calls[a] != 1 || calls[b] != 1 {
+		t.Fatalf("tag-repair scan re-fingerprinted unchanged files: calls = %v", calls)
+	}
+	if _, ok := stored(a); !ok {
+		t.Fatal("tag-repair scan dropped a stored fingerprint")
+	}
+
+	// 3. Bytes that changed are fingerprinted again, and only those.
+	touch(a, time.Hour)
+	scan("changed-file scan")
+	if calls[a] != 2 || calls[b] != 1 {
+		t.Fatalf("changed-file scan calls = %v, want a=2 b=1", calls)
+	}
+
+	// 4. A changed file whose attempt is inconclusive loses its old row: that
+	// row describes the previous bytes, and a stall says nothing about the new
+	// ones.
+	result = fingerprintResult{streamSHA256: sum, printErr: fmt.Errorf("fpcalc: %w", errFingerprintTimeout)}
+	touch(a, 2*time.Hour)
+	scan("inconclusive scan")
+	if _, ok := stored(a); ok {
+		t.Fatal("inconclusive attempt left the previous bytes' fingerprint in place")
+	}
+	if _, ok := stored(b); !ok {
+		t.Fatal("inconclusive attempt on one file removed another file's fingerprint")
+	}
+
+	// 5. A file the tools reject gets a row at the current version with both
+	// halves NULL — a verdict, so the backfill does not retry it every boot.
+	result = fingerprintResult{
+		hashErr:  errors.New("ffmpeg exited 1"),
+		printErr: errors.New("fpcalc exited 2"),
+	}
+	touch(a, 3*time.Hour)
+	scan("rejected scan")
+	got, ok = stored(a)
+	if !ok {
+		t.Fatal("a file the tools rejected got no row, so the backfill would retry it forever")
+	}
+	if got.sha != nil || got.chroma != nil || got.version != fingerprintVersion {
+		t.Fatalf("rejected file stored %+v, want both halves NULL at version %d", got, fingerprintVersion)
+	}
+
+	// 6. With fingerprinting off (#3913) the scan decodes nothing. It still asks
+	// for the stream hash — a demux, and what recognises a moved file — but stores
+	// no row, and a changed file's old row goes: it describes bytes that are gone.
+	result = fingerprintResult{streamSHA256: sum, chromaprint: chroma}
+	off := DefaultFingerprintSettings
+	off.Enabled = false
+	if _, err := settings.Set(ctx, off); err != nil {
+		t.Fatalf("switch fingerprinting off: %v", err)
+	}
+	touch(a, 4*time.Hour)
+	scan("fingerprinting-off scan")
+	if calls[a] != 5 || lastOpts.chromaprint {
+		t.Fatalf("fingerprinting-off scan: calls = %v, last options %+v; want a fifth call asking for no chromaprint",
+			calls, lastOpts)
+	}
+	if _, ok := stored(a); ok {
+		t.Fatal("with fingerprinting off, a changed file kept the previous bytes' fingerprint")
+	}
+	if _, ok := stored(b); !ok {
+		t.Fatal("with fingerprinting off, an unchanged file lost its fingerprint")
+	}
+
+	// 7. The length setting reaches the scan, and is stored with the row.
+	longer := DefaultFingerprintSettings
+	longer.ChromaprintLengthSec = 90
+	if _, err := settings.Set(ctx, longer); err != nil {
+		t.Fatalf("change length: %v", err)
+	}
+	touch(a, 5*time.Hour)
+	scan("new-length scan")
+	got, ok = stored(a)
+	if !ok || got.length != 90 || lastOpts.lengthSec != 90 || !lastOpts.chromaprint {
+		t.Fatalf("new-length scan stored %+v (present %v) after asking for %+v; want a chromaprint at 90s", got, ok, lastOpts)
+	}
+}
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`DROP INDEX IF EXISTS play_events_track_idx;`