fix: ListenBrainz similarity is stored relative to the seed's best, and the candidate arms keep tier order (#6177)

LB's similarity scores are counts (16 to 1445 for recordings, to 10,875
for artists), and they reached a scorer that reads similarity as 0 to 1.
One edge added 64 to 5,800 points under Songs-like, so likes, skips and
recency stopped mattering for any candidate with an edge.

- ResolveListenBrainzTrackEdges divides each edge by the seed's best
  score in its whole cached answer. The cache keeps the raw answer.
- The worker stores matched and unmatched artist rows over the seed's
  top artist.
- Migration 0096 rescales the rows already stored.
- The candidate arms hold sim_score bands so tier 1 stays above tier 2:
  lb_similar 0.5 + 0.5 x edge, tag_overlap halved, likes_overlap 0.3.
  Otherwise a seed's weaker LB edges fell below a shared genre tag.

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

internal/db/dbq/recommendation.sql.go

@@ -834,7 +834,7 @@ excluded_ids AS (
     WHERE q.user_id = $1
 ),
 lb_similar AS (
-    SELECT ts.track_b_id AS track_id, ts.score AS sim_score
+    SELECT ts.track_b_id AS track_id, 0.5 + 0.5 * ts.score AS sim_score
     FROM track_similarity ts
     WHERE ts.track_a_id = $2
       AND ts.source = 'listenbrainz'
@@ -864,8 +864,8 @@ similar_artists AS (
 ),
 tag_overlap AS (
     SELECT t.id AS track_id,
-           (count(DISTINCT trim(g))::float8
-            / GREATEST((SELECT count(*) FROM seed_tags), 1)) AS sim_score
+           0.5 * (count(DISTINCT trim(g))::float8
+                  / GREATEST((SELECT count(*) FROM seed_tags), 1)) AS sim_score
     FROM tracks t
     LEFT JOIN LATERAL regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g_split(g) ON true
     WHERE trim(g_split.g) IN (SELECT tag FROM seed_tags)
@@ -877,7 +877,7 @@ tag_overlap AS (
     LIMIT $7
 ),
 likes_overlap AS (
-    SELECT gl.track_id, 0.6::float8 AS sim_score
+    SELECT gl.track_id, 0.3::float8 AS sim_score
     FROM general_likes gl
     WHERE gl.user_id = $1
       AND gl.track_id NOT IN (SELECT id FROM excluded_ids)
@@ -1025,6 +1025,20 @@ type LoadRadioCandidatesV2Row struct {
 // 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.
+//
+// sim_score BANDS keep the tiers in order (rule 131, #6177). Similarity edges
+// are stored relative to the seed's best match, 0 to 1, so the arms are placed
+// in bands rather than compared raw:
+//
+//	tier 1  lb_similar                      0.5 + 0.5 × edge   (0.5, 1.0]
+//	tier 2  similar_artists, coplay_artists 0.5 × edge         [0, 0.5]
+//	        tag_overlap                     0.5 × share of the seed's tags
+//	        likes_overlap                   0.3, flat
+//	tier 3  taste_overlap, random_fill      0
+//
+// Without the bands, the seed's weaker LB edges fell below a shared genre tag
+// and the flat likes score, and the one arm that measures the seed lost to
+// the ones that guess.
 func (q *Queries) LoadRadioCandidatesV2(ctx context.Context, arg LoadRadioCandidatesV2Params) ([]LoadRadioCandidatesV2Row, error) {
 	rows, err := q.db.Query(ctx, loadRadioCandidatesV2,
 		arg.UserID,

internal/db/dbq/similarity.sql.go

@@ -260,10 +260,19 @@ func (q *Queries) RecordTrackSimilarityFetch(ctx context.Context, arg RecordTrac
 }
 
 const resolveListenBrainzTrackEdges = `-- name: ResolveListenBrainzTrackEdges :exec
-WITH want AS (
+WITH best AS (
+    SELECT seed_track_id, max(score) AS score
+    FROM listenbrainz_similar_recordings
+    WHERE $1::uuid[] IS NULL
+       OR seed_track_id = ANY($1::uuid[])
+    GROUP BY seed_track_id
+),
+want AS (
     SELECT DISTINCT ON (c.seed_track_id, c.recording_mbid)
-           c.seed_track_id AS track_a_id, t.id AS track_b_id, c.score
+           c.seed_track_id AS track_a_id, t.id AS track_b_id,
+           CASE WHEN b.score > 0 THEN c.score / b.score ELSE 0 END AS score
     FROM listenbrainz_similar_recordings c
+    JOIN best b ON b.seed_track_id = c.seed_track_id
     JOIN tracks t ON t.mbid = c.recording_mbid
                  AND t.missing_since IS NULL
                  AND t.id <> c.seed_track_id
@@ -299,6 +308,13 @@ WHERE ts.source = 'listenbrainz'
 // pool would offer the same song twice. No cap per seed: LB's answer is already
 // at most 50.
 //
+// The score is RELATIVE: each edge's LB score over the seed's best score in its
+// whole cached answer, so 1.0 is the seed's closest recording and the rest fall
+// below it (#6177). LB's raw scores are counts (16 to 1445 on the operator's
+// library) and reached the scorer, which reads similarity as 0 to 1, so one
+// edge outweighed likes, skips and recency together. The cache keeps the raw
+// answer; only the edge is scaled.
+//
 // Only edges of seeds the cache owns are removed, those with a fetch row.
 // Edges written before the cache existed, or copied onto a survivor by a fold
 // (merge.sql), stay until their seed is fetched.

internal/db/migrations/0096_similarity_relative_scores.down.sql

@@ -0,0 +1,3 @@
+-- The raw scores are not kept for artists; the worker rewrites every seed with
+-- relative scores on its next fetch either way. Nothing to undo.
+SELECT 1;

internal/db/migrations/0096_similarity_relative_scores.up.sql

@@ -0,0 +1,51 @@
+-- #6177: ListenBrainz similarity scores were stored raw. They are counts (16 to
+-- 1445 for recordings, 16 to 10,875 for artists on the operator's library), and
+-- every reader treats similarity as 0 to 1. Rescale the stored rows relative to
+-- each seed's best score, as the worker now writes them.
+--
+-- A seed's best is taken over everything stored for it: the cached LB answer
+-- and its edges for tracks, matched and unmatched rows for artists. A seed
+-- whose best is already at most 1 is left alone, so a second run changes
+-- nothing.
+
+WITH best AS (
+    SELECT seed_id, max(score) AS score
+    FROM (
+        SELECT track_a_id AS seed_id, score FROM track_similarity WHERE source = 'listenbrainz'
+        UNION ALL
+        SELECT seed_track_id, score FROM listenbrainz_similar_recordings
+    ) s
+    GROUP BY seed_id
+)
+UPDATE track_similarity ts
+   SET score = ts.score / b.score
+  FROM best b
+ WHERE ts.source = 'listenbrainz'
+   AND ts.track_a_id = b.seed_id
+   AND b.score > 1;
+
+-- One statement, so both updates divide by the best as it stood before either.
+WITH best AS (
+    SELECT seed_id, max(score) AS score
+    FROM (
+        SELECT artist_a_id AS seed_id, score FROM artist_similarity WHERE source = 'listenbrainz'
+        UNION ALL
+        SELECT seed_artist_id, score FROM artist_similarity_unmatched WHERE source = 'listenbrainz'
+    ) s
+    GROUP BY seed_id
+),
+matched AS (
+    UPDATE artist_similarity a
+       SET score = a.score / b.score
+      FROM best b
+     WHERE a.source = 'listenbrainz'
+       AND a.artist_a_id = b.seed_id
+       AND b.score > 1
+    RETURNING 1
+)
+UPDATE artist_similarity_unmatched u
+   SET score = u.score / b.score
+  FROM best b
+ WHERE u.source = 'listenbrainz'
+   AND u.seed_artist_id = b.seed_id
+   AND b.score > 1;

internal/db/queries/recommendation.sql

@@ -62,6 +62,18 @@ WHERE t.id <> $2
 -- 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.
+--
+-- sim_score BANDS keep the tiers in order (rule 131, #6177). Similarity edges
+-- are stored relative to the seed's best match, 0 to 1, so the arms are placed
+-- in bands rather than compared raw:
+--   tier 1  lb_similar                      0.5 + 0.5 × edge   (0.5, 1.0]
+--   tier 2  similar_artists, coplay_artists 0.5 × edge         [0, 0.5]
+--           tag_overlap                     0.5 × share of the seed's tags
+--           likes_overlap                   0.3, flat
+--   tier 3  taste_overlap, random_fill      0
+-- Without the bands, the seed's weaker LB edges fell below a shared genre tag
+-- and the flat likes score, and the one arm that measures the seed lost to
+-- the ones that guess.
 
 WITH
 seed_artist AS (
@@ -87,7 +99,7 @@ excluded_ids AS (
     WHERE q.user_id = $1
 ),
 lb_similar AS (
-    SELECT ts.track_b_id AS track_id, ts.score AS sim_score
+    SELECT ts.track_b_id AS track_id, 0.5 + 0.5 * ts.score AS sim_score
     FROM track_similarity ts
     WHERE ts.track_a_id = $2
       AND ts.source = 'listenbrainz'
@@ -117,8 +129,8 @@ similar_artists AS (
 ),
 tag_overlap AS (
     SELECT t.id AS track_id,
-           (count(DISTINCT trim(g))::float8
-            / GREATEST((SELECT count(*) FROM seed_tags), 1)) AS sim_score
+           0.5 * (count(DISTINCT trim(g))::float8
+                  / GREATEST((SELECT count(*) FROM seed_tags), 1)) AS sim_score
     FROM tracks t
     LEFT JOIN LATERAL regexp_split_to_table(coalesce(t.genre, ''), '[;,]') AS g_split(g) ON true
     WHERE trim(g_split.g) IN (SELECT tag FROM seed_tags)
@@ -130,7 +142,7 @@ tag_overlap AS (
     LIMIT $7
 ),
 likes_overlap AS (
-    SELECT gl.track_id, 0.6::float8 AS sim_score
+    SELECT gl.track_id, 0.3::float8 AS sim_score
     FROM general_likes gl
     WHERE gl.user_id = $1
       AND gl.track_id NOT IN (SELECT id FROM excluded_ids)

internal/db/queries/similarity.sql

@@ -68,13 +68,29 @@ ON CONFLICT (artist_id) DO UPDATE SET fetched_at = now(), returned = EXCLUDED.re
 -- pool would offer the same song twice. No cap per seed: LB's answer is already
 -- at most 50.
 --
+-- The score is RELATIVE: each edge's LB score over the seed's best score in its
+-- whole cached answer, so 1.0 is the seed's closest recording and the rest fall
+-- below it (#6177). LB's raw scores are counts (16 to 1445 on the operator's
+-- library) and reached the scorer, which reads similarity as 0 to 1, so one
+-- edge outweighed likes, skips and recency together. The cache keeps the raw
+-- answer; only the edge is scaled.
+--
 -- Only edges of seeds the cache owns are removed, those with a fetch row.
 -- Edges written before the cache existed, or copied onto a survivor by a fold
 -- (merge.sql), stay until their seed is fetched.
-WITH want AS (
+WITH best AS (
+    SELECT seed_track_id, max(score) AS score
+    FROM listenbrainz_similar_recordings
+    WHERE sqlc.narg(seed_ids)::uuid[] IS NULL
+       OR seed_track_id = ANY(sqlc.narg(seed_ids)::uuid[])
+    GROUP BY seed_track_id
+),
+want AS (
     SELECT DISTINCT ON (c.seed_track_id, c.recording_mbid)
-           c.seed_track_id AS track_a_id, t.id AS track_b_id, c.score
+           c.seed_track_id AS track_a_id, t.id AS track_b_id,
+           CASE WHEN b.score > 0 THEN c.score / b.score ELSE 0 END AS score
     FROM listenbrainz_similar_recordings c
+    JOIN best b ON b.seed_track_id = c.seed_track_id
     JOIN tracks t ON t.mbid = c.recording_mbid
                  AND t.missing_since IS NULL
                  AND t.id <> c.seed_track_id