chore(release): v0.3.1

Reviewed-on: #8

.gitea/workflows/build-and-publish.yaml

@@ -45,7 +45,8 @@ jobs:
 
       # The suite runs inside the image, against the interpreter that ships,
       # rather than against whatever the runner happens to provide. A failing
-      # test fails the build. Layers are shared with the push build below.
+      # test fails the build. The runtime stage below is built from the same
+      # daemon afterwards, so its layers are already in cache.
       - name: Run the test suite inside the image
         run: docker build --target test -t music-curator:test .
 
@@ -124,9 +125,6 @@ jobs:
           echo "release=${release}" >> "$GITHUB_OUTPUT"
           echo "Computed bump=${bump}, release=${release}, base=${base}"
 
-      - name: Set up Buildx
-        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5 # v4
-
       - name: Log in to the Gitea container registry
         if: github.event_name != 'pull_request'
         uses: docker/login-action@650006c6eb7dba73a995cc03b0b2d7f5ca915bee # v4
@@ -135,21 +133,33 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.PACKAGES_TOKEN }}
 
-      - name: Build and push
+      # Plain `docker build` rather than buildx. buildx boots its own buildkit
-        uses: docker/build-push-action@f9f3042f7e2789586610d6e8b85c8f03e5195baf # v7
+      # in a container with a cache of its own, so it shared nothing with the
-        with:
+      # test build above and rebuilt the image from the base image up -- two
-          context: .
+      # full builds per run. It earns that cost when building for several
-          # Without this the last stage in the Dockerfile -- the test stage --
+      # platforms; this only ever targets the amd64 NAS, so it does not.
-          # would be what gets published.
+      #
-          target: runtime
+      # `--target runtime` is a strict prefix of the test stage, so every layer
-          # The NAS is the only host this runs on. Building arm64 as well would
+      # is already in the daemon's cache and this resolves in seconds.
-          # mean emulating it under QEMU for no consumer.
+      - name: Build the runtime image
-          platforms: linux/amd64
+        run: |
-          push: ${{ github.event_name != 'pull_request' }}
+          set -euo pipefail
-          tags: ${{ steps.version.outputs.tags }}
+          tags=()
-          labels: |
+          while IFS= read -r tag; do
-            org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}
+            [ -n "$tag" ] && tags+=(-t "$tag")
-            org.opencontainers.image.revision=${{ github.sha }}
+          done <<< "${{ steps.version.outputs.tags }}"
+          docker build --target runtime \
+            --label "org.opencontainers.image.source=${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}" \
+            --label "org.opencontainers.image.revision=${GITHUB_SHA}" \
+            "${tags[@]}" .
+
+      - name: Push
+        if: github.event_name != 'pull_request'
+        run: |
+          set -euo pipefail
+          while IFS= read -r tag; do
+            [ -n "$tag" ] && docker push "$tag"
+          done <<< "${{ steps.version.outputs.tags }}"
 
       # Record the release: write the computed version into pyproject.toml, then
       # commit and tag it, so the packaging metadata always matches the release

README.md

@@ -29,7 +29,14 @@ Two tiers, and no third.
 | `name` | Normalised artist and title | Everything the first tier could not carry |
 | `none` | — | Recorded as a miss, never guessed at |
 
-The normalisation is the load-bearing part, because the two sides disagree in
+The name tier does almost all of the work. A recording MBID is exact when it
+lands, but MusicBrainz holds a separate recording per release, and Last.fm and
+Lidarr rarely pick the same one: on a real library, two thirds of scrobbles
+carry a recording id and barely a twentieth of them join on it. The report
+counts how many carried an id and matched on name anyway, which is the measure
+of that disagreement.
+
+So the normalisation is the load-bearing part, because the two sides disagree in
 predictable ways. It folds case and accents, drops guest credits (`Yellowcard
 feat. Tay Jardine` against a tag of `Yellowcard`), strips a trailing
 version suffix (`(Remastered 2011)`, `- Live`), expands `&`, and removes a
@@ -102,10 +109,22 @@ matcher. The line to watch is:
 unmatched by an artist the library holds: N pairs, M plays
 ```
 
-That is a track that was played, sitting beside a file it should have matched.
+That is a track that was played by an artist the library holds. It is then
-Those are the matcher's real misses, and every one is a candidate for being
+split three ways, because owning an artist is a weak proxy for owning a track
-wrongly called cold in stage four. The report lists the worst fifteen by play
+and a shared title is a weak proxy for a shared song:
-count so they can be eyeballed.
+
+- **the library's own title credits the scrobbled artist** — `Voodoo People
+  (Pendulum Remix)` against a play credited to Pendulum. Same song, filed
+  under the original artist. These are the genuine misses.
+- **the same title under an unrelated artist** — a collision, not a miss.
+  Across fifty thousand tracks these are constant: `Everyday` is Rusko and
+  also Def Leppard, `Kaleidoscope` is Delta Heavy and also Chappell Roan.
+  Matching on title alone would be far worse than missing them, which is why
+  there is no such tier.
+- **the title is nowhere in the library** — never bought.
+
+Only the first is worth chasing. Counting all three as matcher failures
+overstates the problem and would over-block the cull.
 
 ## How the ingest works
 

music_curator.py

@@ -141,6 +141,11 @@ CREATE TABLE IF NOT EXISTS lidarr_track (
 );
 CREATE INDEX IF NOT EXISTS lidarr_track_recording ON lidarr_track (recording_mbid);
 CREATE INDEX IF NOT EXISTS lidarr_track_norm ON lidarr_track (norm_artist, norm_title);
+-- Separate from the composite above, which a lookup by title alone cannot use:
+-- its leading column is the artist. The report searches by title on its own, and
+-- without this it scans every track for every unmatched key -- forty-seven
+-- seconds on a library of eighty-four thousand.
+CREATE INDEX IF NOT EXISTS lidarr_track_title ON lidarr_track (norm_title);
 CREATE INDEX IF NOT EXISTS lidarr_track_album ON lidarr_track (album_id);
 
 -- One row per distinct thing listened to, with the verdict on whether it could
@@ -187,18 +192,28 @@ VERSION_WORDS = (
     "anniversary",
     "reissue",
     "instrumental",
+    # Drum and bass and its neighbours mark versions their own way.
+    "vip",
+    "bootleg",
+    "rework",
+    "extended",
 )
 _VERSIONS = "|".join(VERSION_WORDS)
 BRACKETED_VERSION = re.compile(
     rf"\s*[\(\[][^\)\]]*\b(?:{_VERSIONS})\b[^\)\]]*[\)\]]\s*$", re.IGNORECASE
 )
-TRAILING_VERSION = re.compile(rf"\s+-\s+[^-]*\b(?:{_VERSIONS})\b.*$", re.IGNORECASE)
+# The suffix is matched lazily rather than as a run of non-hyphens, because the
+# thing being stripped frequently contains hyphens of its own -- "Gold Dust -
+# Shy FX Re-Edit", "Back To Your Roots - Friction & K-Tee Remix".
+TRAILING_VERSION = re.compile(rf"\s+-\s+.*?\b(?:{_VERSIONS})\b.*$", re.IGNORECASE)
 
-# Last.fm routinely carries the guest credit in the artist field where the file
+# Last.fm routinely carries the guest credit where the file tag holds only the
-# tag holds only the primary artist -- "Yellowcard feat. Tay Jardine" against a
+# primary artist -- "Yellowcard feat. Tay Jardine" against a tag of
-# tag of "Yellowcard". `with` is deliberately absent: it appears in far too many
+# "Yellowcard", or "Self vs Self (feat. In Flames)" against "Self vs Self". The
-# real titles to cut on sight.
+# opening bracket has to be allowed for: requiring whitespace immediately before
-GUEST_CREDIT = re.compile(r"\s+(?:feat|ft|featuring)\b.*$", re.IGNORECASE)
+# the word misses every bracketed credit, which is most of them. `with` is
+# deliberately absent -- it appears in far too many real titles to cut on sight.
+GUEST_CREDIT = re.compile(r"[\s(\[]+(?:feat|ft|featuring)\b.*$", re.IGNORECASE)
 
 LEADING_ARTICLE = re.compile(r"^the\s+")
 # Deleted rather than spaced, so "Don't" and "Dont" agree. Every other mark
@@ -1078,18 +1093,107 @@ def coverage_report(store):
             row["plays"],
         )
 
+    # Why the mbid tier performs the way it does. A recording id on both sides
+    # that still fails to join means the two disagree about which recording the
+    # song is -- MusicBrainz holds a separate recording per release, and Last.fm
+    # and Lidarr need not have picked the same one. That is not a fault to fix
+    # in the matcher; it is the reason the name tier has to carry the load.
+    library_with_mbid = store.scalar(
+        "SELECT COUNT(*) FROM lidarr_track WHERE recording_mbid IS NOT NULL"
+    )
+    logger.info(
+        "library tracks carrying a recording MBID: %d of %d (%.1f%%)",
+        library_with_mbid,
+        tracks,
+        100 * library_with_mbid / tracks,
+    )
+    disagreed = store.connection.execute(
+        "SELECT COUNT(*) AS pairs, COALESCE(SUM(plays), 0) AS plays FROM scrobble_key"
+        " WHERE track_mbid IS NOT NULL AND method = 'name'"
+    ).fetchone()
+    logger.info(
+        "carried a recording MBID, joined on name instead: %d pairs, %d plays"
+        " -- both sides know the song, they disagree on which recording it is",
+        disagreed["pairs"],
+        disagreed["plays"],
+    )
+
     suspect = store.connection.execute(
         "SELECT COUNT(*) AS pairs, COALESCE(SUM(plays), 0) AS plays FROM scrobble_key k"
         " WHERE k.track_id IS NULL"
         "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
     ).fetchone()
     logger.info(
-        "unmatched by an artist the library holds: %d pairs, %d plays -- these are the"
+        "unmatched by an artist the library holds: %d pairs, %d plays",
-        " matcher's misses, not music you do not own",
         suspect["pairs"],
         suspect["plays"],
     )
 
+    # Owning an artist is a weak proxy for owning a track, so that figure alone
+    # overstates the matcher's failings. Split it -- but not on the title alone.
+    # Across fifty thousand tracks, titles collide constantly: "Everyday" is
+    # Rusko and also Def Leppard, "Kaleidoscope" is Delta Heavy and also
+    # Chappell Roan. Matching those would be worse than missing them.
+    #
+    # The signal for a genuine attribution miss is that the library's own title
+    # credits the artist the scrobble is filed under -- "Voodoo People (Pendulum
+    # Remix)" against a play credited to Pendulum. The normalised title has that
+    # suffix stripped, which is exactly what let them meet, so the raw one has to
+    # be searched for the name.
+    attribution = store.connection.execute(
+        "SELECT COUNT(*) AS pairs, COALESCE(SUM(plays), 0) AS plays FROM scrobble_key k"
+        " WHERE k.track_id IS NULL"
+        "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
+        "   AND EXISTS (SELECT 1 FROM lidarr_track t"
+        "                WHERE t.norm_title = k.norm_track"
+        "                  AND instr(lower(t.title), lower(k.artist)) > 0)"
+    ).fetchone()
+    collision = store.connection.execute(
+        "SELECT COUNT(*) AS pairs, COALESCE(SUM(plays), 0) AS plays FROM scrobble_key k"
+        " WHERE k.track_id IS NULL"
+        "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
+        "   AND EXISTS (SELECT 1 FROM lidarr_track t WHERE t.norm_title = k.norm_track)"
+    ).fetchone()
+    logger.info(
+        "  the library's title credits the scrobbled artist: %d pairs, %d plays"
+        " -- remixes and guest spots, and the genuine misses",
+        attribution["pairs"],
+        attribution["plays"],
+    )
+    logger.info(
+        "  same title under an unrelated artist: %d pairs, %d plays"
+        " -- title collisions, not misses; matching these would be a mistake",
+        collision["pairs"] - attribution["pairs"],
+        collision["plays"] - attribution["plays"],
+    )
+    logger.info(
+        "  the rest, %d pairs, %d plays: you own the artist but not the track",
+        suspect["pairs"] - collision["pairs"],
+        suspect["plays"] - collision["plays"],
+    )
+
+    mismatched = store.connection.execute(
+        "SELECT k.artist, k.track, k.plays,"
+        "       (SELECT t.title FROM lidarr_track t"
+        "         WHERE t.norm_title = k.norm_track"
+        "           AND instr(lower(t.title), lower(k.artist)) > 0 LIMIT 1) AS library_title"
+        "  FROM scrobble_key k"
+        " WHERE k.track_id IS NULL"
+        "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
+        "   AND EXISTS (SELECT 1 FROM lidarr_track t"
+        "                WHERE t.norm_title = k.norm_track"
+        "                  AND instr(lower(t.title), lower(k.artist)) > 0)"
+        " ORDER BY k.plays DESC, k.artist LIMIT 10"
+    ).fetchall()
+    for position, row in enumerate(mismatched, start=1):
+        logger.info(
+            "  attribution %2d: %-45s %4d plays, library has %r",
+            position,
+            f"{row['artist']} - {row['track']}"[:45],
+            row["plays"],
+            row["library_title"],
+        )
+
     with_files = store.scalar("SELECT COUNT(*) FROM lidarr_track WHERE has_file = 1")
     played = store.scalar(
         "SELECT COUNT(DISTINCT k.track_id) FROM scrobble_key k"

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "music-curator"
-version = "0.2.3"
+version = "0.3.1"
 description = "Ingest a Last.fm listening history and curate a music library from it"
 readme = "README.md"
 requires-python = ">=3.11"

tests/test_music_curator.py

@@ -55,8 +55,22 @@ LIBRARY = [
     {
         "name": "The Prodigy",
         "albums": [
-            {"title": "The Fat of the Land", "tracks": [{"title": "Breathe (Remastered)"}]}
+            {
+                "title": "The Fat of the Land",
+                "tracks": [
+                    {"title": "Breathe (Remastered)"},
+                    # Credits its remixer in the title, which is the only signal
+                    # separating a real attribution miss from a title collision.
+                    {"title": "Voodoo People (Pendulum Remix)"},
                 ],
+            }
+        ],
+    },
+    # Held by the library in its own right, which is what puts its scrobbles
+    # inside the "artist the library holds" filter at all.
+    {
+        "name": "Pendulum",
+        "albums": [{"title": "Immersion", "tracks": [{"title": "Watercolour"}]}],
     },
 ]
 
@@ -393,7 +407,7 @@ def test_keys_carry_the_play_count_and_the_span(tmp_path):
 def test_re_indexing_drops_what_lidarr_no_longer_has(tmp_path):
     """The index is Lidarr's mirror, not an accumulation of everything ever seen."""
     store = indexed(tmp_path, [scrobble_of("AC/DC", "Hells Bells")])
-    assert store.scalar("SELECT COUNT(*) FROM lidarr_artist") == 3
+    assert store.scalar("SELECT COUNT(*) FROM lidarr_artist") == 4
 
     music_curator.index_library(
         music_curator.Lidarr("http://lidarr", "key", transport=FakeLidarr(LIBRARY[:1])), store
@@ -512,7 +526,7 @@ def test_albums_come_from_the_unfiltered_endpoint(tmp_path):
 
     album_calls = [query for path, query in api.calls if path == "album"]
     assert album_calls == [{}]
-    assert store.scalar("SELECT COUNT(*) FROM lidarr_album") == 3
+    assert store.scalar("SELECT COUNT(*) FROM lidarr_album") == 4
 
 
 def test_a_bad_album_falls_back_to_asking_per_artist(tmp_path):
@@ -526,7 +540,7 @@ def test_a_bad_album_falls_back_to_asking_per_artist(tmp_path):
 
     assert store.scalar("SELECT COUNT(*) FROM lidarr_album WHERE artist_id = 2") == 0
     # The other two artists keep their albums.
-    assert store.scalar("SELECT COUNT(*) FROM lidarr_album") == 2
+    assert store.scalar("SELECT COUNT(*) FROM lidarr_album") == 3
     assert store.get_state("index_albums_skipped") == "1"
 
 
@@ -549,10 +563,10 @@ def test_an_artist_lidarr_cannot_serve_does_not_kill_the_index(tmp_path):
 
     music_curator.index_library(music_curator.Lidarr("http://lidarr", "key", transport=api), store)
 
-    assert store.scalar("SELECT COUNT(*) FROM lidarr_artist") == 3
+    assert store.scalar("SELECT COUNT(*) FROM lidarr_artist") == 4
     assert store.scalar("SELECT COUNT(*) FROM lidarr_track WHERE artist_id = 2") == 0
     # The other two artists are indexed in full.
-    assert store.scalar("SELECT COUNT(*) FROM lidarr_track") == 3
+    assert store.scalar("SELECT COUNT(*) FROM lidarr_track") == 5
     assert store.get_state("index_skipped") == "1"
 
 
@@ -687,3 +701,127 @@ def test_lidarr_talks_to_a_real_server_through_keep_alive(http_server):
         client.transport.close()
 
     assert server.connections == 1
+
+
+# Titles taken verbatim from a real coverage report's unmatched list. Each one
+# was a genuine miss before the normaliser handled it.
+@pytest.mark.parametrize(
+    ("scrobbled", "tagged"),
+    [
+        ("Self vs Self (feat. In Flames)", "Self vs Self"),
+        ("Grime Battle of Hastings (feat. The Town Crier)", "Grime Battle of Hastings"),
+        ("Gold Dust - Shy FX Re-Edit", "Gold Dust"),
+        ("Back To Your Roots - Friction & K-Tee Remix", "Back To Your Roots"),
+        ("Constellations - Forza Horizon 3 VIP", "Constellations"),
+        ("Everyday (Netsky Remix)", "Everyday"),
+        ("Voodoo People [Pendulum Remix] [Live At Brixton Academy]", "Voodoo People"),
+    ],
+)
+def test_real_unmatched_titles_now_agree_with_their_tags(scrobbled, tagged):
+    assert music_curator.normalise(scrobbled) == music_curator.normalise(tagged)
+
+
+@pytest.mark.parametrize(
+    "title",
+    [
+        "Dancing with Myself",
+        "(Don't Fear) The Reaper",
+        "Live and Let Die",
+        "Radio Ga Ga",
+        "Editors",
+        "Mixed Emotions",
+        "Vipassana",
+    ],
+)
+def test_the_version_words_do_not_eat_ordinary_titles(title):
+    """Every one of these contains a version word and must survive intact."""
+    assert music_curator.normalise(title) == music_curator.normalise(title.lower())
+    assert len(music_curator.normalise(title).split()) == len(title.split())
+
+
+def test_a_bracketed_guest_credit_matches_the_bare_tag(tmp_path):
+    """Whitespace-then-feat misses the bracketed form, which is most of them."""
+    store = indexed(tmp_path, [scrobble_of("Yellowcard", "Here I Am Alive (feat. Someone)")])
+
+    method, track_id = verdict(store, "Yellowcard", "Here I Am Alive (feat. Someone)")
+    assert method == "name"
+    assert track_id is not None
+
+
+def test_misses_are_split_by_whether_the_library_holds_the_title(tmp_path):
+    """Owning an artist is a weak proxy for owning a track. Counting both as
+    matcher failures overstates the problem and would over-block the cull."""
+    store = indexed(
+        tmp_path,
+        [
+            # The library holds "Hells Bells", but under AC/DC, not Yellowcard:
+            # an attribution disagreement, and a real miss.
+            scrobble_of("Yellowcard", "Hells Bells"),
+            # Yellowcard is in the library; this track is not, under any artist.
+            scrobble_of("Yellowcard", "A Single She Never Bought"),
+        ],
+    )
+
+    def count(extra):
+        return store.connection.execute(
+            "SELECT COUNT(*) FROM scrobble_key k WHERE k.track_id IS NULL"
+            "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
+            f"  {extra}"
+        ).fetchone()[0]
+
+    assert count("") == 2
+    assert count("AND EXISTS (SELECT 1 FROM lidarr_track t WHERE t.norm_title = k.norm_track)") == 1
+
+
+def test_the_report_survives_the_attribution_split(tmp_path):
+    store = indexed(tmp_path, [scrobble_of("Yellowcard", "Hells Bells")])
+
+    music_curator.report(store, NOW)
+
+
+def attribution_pairs(store):
+    """Unmatched pairs where the library's own title credits the scrobbled artist."""
+    return [
+        row["artist"]
+        for row in store.connection.execute(
+            "SELECT k.artist FROM scrobble_key k WHERE k.track_id IS NULL"
+            "   AND EXISTS (SELECT 1 FROM lidarr_artist a WHERE a.norm_name = k.norm_artist)"
+            "   AND EXISTS (SELECT 1 FROM lidarr_track t"
+            "                WHERE t.norm_title = k.norm_track"
+            "                  AND instr(lower(t.title), lower(k.artist)) > 0)"
+        )
+    ]
+
+
+def test_a_shared_title_is_not_an_attribution_miss(tmp_path):
+    """Across fifty thousand tracks, titles collide constantly: "Everyday" is
+    Rusko and also Def Leppard. Matching those would be worse than missing."""
+    store = indexed(tmp_path, [scrobble_of("Yellowcard", "Hells Bells")])
+
+    # The library holds "Hells Bells", by AC/DC, and its title says nothing
+    # about Yellowcard. A collision, not a miss.
+    assert verdict(store, "Yellowcard", "Hells Bells") == ("none", None)
+    assert attribution_pairs(store) == []
+
+
+def test_a_remix_credited_in_the_library_title_is_an_attribution_miss(tmp_path):
+    """The library has "Voodoo People (Pendulum Remix)" under The Prodigy; the
+    scrobble credits Pendulum. Same song, different filing."""
+    store = indexed(tmp_path, [scrobble_of("Pendulum", "Voodoo People")])
+
+    assert attribution_pairs(store) == ["Pendulum"]
+
+
+def test_the_title_index_is_used_for_the_report_lookup(tmp_path):
+    """Without it the report scans every track for every unmatched key: forty-
+    seven seconds on a real library."""
+    store = indexed(tmp_path, [])
+
+    plan = "\n".join(
+        row[-1]
+        for row in store.connection.execute(
+            "EXPLAIN QUERY PLAN SELECT 1 FROM scrobble_key k WHERE k.track_id IS NULL"
+            " AND EXISTS (SELECT 1 FROM lidarr_track t WHERE t.norm_title = k.norm_track)"
+        )
+    )
+    assert "lidarr_track_title" in plan, plan