feat: genre mixes pick specific genres that don't share songs (#6933)

The shelf had four mixes built largely from the same lo-fi beats: about
1,250 tracks carry Hip Hop, Lo-Fi, Instrumental, Downtempo and four more
tags at once, and the near-duplicate guard compared artists, not songs, so
each tag passed as its own channel.

Genre selection now works on songs:
- ListGenreMixCandidates returns each genre's track ids instead of its
  artist ids.
- A broad genre, one whose songs mostly also carry clearly narrower
  candidates (Rock over Pop Punk and Ska Punk), is considered after the
  specific genres of its tier.
- A genre is kept only when at least half its songs are not already in a
  kept genre.
- Each mix leaves out songs tagged with a genre kept before it
  (ListGenreMixTracks exclude_genres), so a song is in one mix at most.

The tiers, the minimum shelf and the minimum mix length are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/db/dbq/genre_mixes.sql.go

@@ -17,7 +17,7 @@ WITH library_genres AS (
     SELECT lower(trim(g.g))           AS genre_key,
            min(trim(g.g))::text       AS label,
            COUNT(DISTINCT t.id)        AS track_count,
-           array_agg(DISTINCT t.artist_id)::uuid[] AS artist_ids
+           array_agg(DISTINCT t.id)::uuid[] AS track_ids
       FROM tracks t
       CROSS JOIN LATERAL regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
      WHERE t.missing_since IS NULL
@@ -59,7 +59,7 @@ SELECT lg.genre_key::text AS genre_key,
        (CASE WHEN ta.genre_key IS NOT NULL THEN ta.weight
              WHEN h.genre_key  IS NOT NULL THEN h.plays::float8
              ELSE lg.track_count::float8 END)::float8 AS score,
-       lg.artist_ids
+       lg.track_ids
   FROM library_genres lg
   LEFT JOIN taste   ta ON ta.genre_key = lg.genre_key
   LEFT JOIN history h  ON h.genre_key  = lg.genre_key
@@ -68,11 +68,11 @@ SELECT lg.genre_key::text AS genre_key,
 `
 
 type ListGenreMixCandidatesRow struct {
-	GenreKey  string
-	Label     string
-	Tier      int32
-	Score     float64
-	ArtistIds []pgtype.UUID
+	GenreKey string
+	Label    string
+	Tier     int32
+	Score    float64
+	TrackIds []pgtype.UUID
 }
 
 // Candidate queries for the genre mixes: one daily system playlist per genre
@@ -89,8 +89,10 @@ type ListGenreMixCandidatesRow struct {
 //	        has decayed past its window
 //	tier 3: any other library genre, scored by its size
 //
-// artist_ids lets the caller skip a genre whose artists mostly repeat one
-// already picked (Indie Rock after Alternative Rock).
+// track_ids lets the caller tell how much music two genres share: a genre
+// whose songs mostly sit in genres already picked adds no new channel, and a
+// broad genre ("Rock") whose songs mostly carry narrower ones ("Pop Punk",
+// "Ska Punk") gives way to them.
 // $1 user_id.
 func (q *Queries) ListGenreMixCandidates(ctx context.Context, userID pgtype.UUID) ([]ListGenreMixCandidatesRow, error) {
 	rows, err := q.db.Query(ctx, listGenreMixCandidates, userID)
@@ -106,7 +108,7 @@ func (q *Queries) ListGenreMixCandidates(ctx context.Context, userID pgtype.UUID
 			&i.Label,
 			&i.Tier,
 			&i.Score,
-			&i.ArtistIds,
+			&i.TrackIds,
 		); err != nil {
 			return nil, err
 		}
@@ -189,6 +191,11 @@ eligible AS (
              FROM regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
             WHERE lower(trim(g.g)) = $3::text
        )
+       AND NOT EXISTS (
+           SELECT 1
+             FROM regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
+            WHERE lower(trim(g.g)) = ANY($4::text[])
+       )
        AND NOT EXISTS (
            SELECT 1 FROM quarantine q
             WHERE q.user_id = $1 AND q.track_id = t.id
@@ -207,9 +214,10 @@ SELECT r.id, r.album_id, r.artist_id, r.familiar::bool AS familiar
 `
 
 type ListGenreMixTracksParams struct {
-	UserID   pgtype.UUID
-	DateSeed string
-	GenreKey string
+	UserID        pgtype.UUID
+	DateSeed      string
+	GenreKey      string
+	ExcludeGenres []string
 }
 
 type ListGenreMixTracksRow struct {
@@ -244,8 +252,16 @@ type ListGenreMixTracksRow struct {
 // A track the user keeps skipping stays out: twice-skipped and skipped more
 // often than played through. The second half of that test keeps a favourite
 // with a couple of stray skips in.
+//
+// exclude_genres keeps each song in one mix: a track carrying a genre that
+// already has a mix today is left to that mix.
 func (q *Queries) ListGenreMixTracks(ctx context.Context, arg ListGenreMixTracksParams) ([]ListGenreMixTracksRow, error) {
-	rows, err := q.db.Query(ctx, listGenreMixTracks, arg.UserID, arg.DateSeed, arg.GenreKey)
+	rows, err := q.db.Query(ctx, listGenreMixTracks,
+		arg.UserID,
+		arg.DateSeed,
+		arg.GenreKey,
+		arg.ExcludeGenres,
+	)
 	if err != nil {
 		return nil, err
 	}

internal/db/queries/genre_mixes.sql

@@ -14,14 +14,16 @@
 --           has decayed past its window
 --   tier 3: any other library genre, scored by its size
 --
--- artist_ids lets the caller skip a genre whose artists mostly repeat one
--- already picked (Indie Rock after Alternative Rock).
+-- track_ids lets the caller tell how much music two genres share: a genre
+-- whose songs mostly sit in genres already picked adds no new channel, and a
+-- broad genre ("Rock") whose songs mostly carry narrower ones ("Pop Punk",
+-- "Ska Punk") gives way to them.
 -- $1 user_id.
 WITH library_genres AS (
     SELECT lower(trim(g.g))           AS genre_key,
            min(trim(g.g))::text       AS label,
            COUNT(DISTINCT t.id)        AS track_count,
-           array_agg(DISTINCT t.artist_id)::uuid[] AS artist_ids
+           array_agg(DISTINCT t.id)::uuid[] AS track_ids
       FROM tracks t
       CROSS JOIN LATERAL regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
      WHERE t.missing_since IS NULL
@@ -63,7 +65,7 @@ SELECT lg.genre_key::text AS genre_key,
        (CASE WHEN ta.genre_key IS NOT NULL THEN ta.weight
              WHEN h.genre_key  IS NOT NULL THEN h.plays::float8
              ELSE lg.track_count::float8 END)::float8 AS score,
-       lg.artist_ids
+       lg.track_ids
   FROM library_genres lg
   LEFT JOIN taste   ta ON ta.genre_key = lg.genre_key
   LEFT JOIN history h  ON h.genre_key  = lg.genre_key
@@ -96,6 +98,9 @@ SELECT lg.genre_key::text AS genre_key,
 -- A track the user keeps skipping stays out: twice-skipped and skipped more
 -- often than played through. The second half of that test keeps a favourite
 -- with a couple of stray skips in.
+--
+-- exclude_genres keeps each song in one mix: a track carrying a genre that
+-- already has a mix today is left to that mix.
 WITH play_counts AS (
     SELECT pe.track_id, COUNT(*) AS c
       FROM play_events pe
@@ -166,6 +171,11 @@ eligible AS (
              FROM regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
             WHERE lower(trim(g.g)) = sqlc.arg(genre_key)::text
        )
+       AND NOT EXISTS (
+           SELECT 1
+             FROM regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g(g)
+            WHERE lower(trim(g.g)) = ANY(sqlc.arg(exclude_genres)::text[])
+       )
        AND NOT EXISTS (
            SELECT 1 FROM quarantine q
             WHERE q.user_id = sqlc.arg(user_id) AND q.track_id = t.id