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/recommendation/candidates.go

@@ -127,8 +127,9 @@ func DefaultCandidateSourceLimits() CandidateSourceLimits {
 //	tier 3  taste_overlap, random_fill
 //	                        seed-independent. The floor, and nothing more.
 //
-// The tiering is enforced by SCORE rather than by a fallback ladder: tier-3
-// arms carry sim_score 0, so under SongsLike weights (SimilarityWeight 4.0,
+// The tiering is enforced by SCORE rather than by a fallback ladder: each tier
+// holds its own sim_score band (recommendation.sql; tier 1 above 0.5, tier 2
+// at or below it, #6177), and tier-3 arms carry sim_score 0, so under SongsLike weights (SimilarityWeight 4.0,
 // everything seed-independent demoted) they rank below any real match and
 // surface only when tiers 1–2 cannot fill the mix. That is the rule's
 // "fill from tier 1 first, reach down only when a tier cannot fill".
@@ -159,19 +160,19 @@ func DefaultCandidateSourceLimits() CandidateSourceLimits {
 // surface only when the closer tiers cannot fill the mix.
 //
 // likes_overlap is cut hardest of the tier-2 arms for a specific reason: its
-// SQL assigns a FLAT 0.6 sim_score (recommendation.sql) rather than measuring
+// SQL assigns a FLAT 0.3 sim_score (recommendation.sql) rather than measuring
 // anything. It is a collaborative signal wearing similarity's clothes, and a
-// raised SimilarityWeight amplifies it — if real ListenBrainz scores commonly
-// land below 0.6 it would outrank genuine matches. Halved pending the
-// fill-rate measurement in #3879; the honest fix is to stop it claiming a
-// similarity score it never computed.
+// raised SimilarityWeight amplifies it. The tier bands keep it below every
+// ListenBrainz match (#6177); halved pending the fill-rate measurement in
+// #3879, and the honest fix is still to stop it claiming a similarity score
+// it never computed.
 func SongsLikeCandidateSourceLimits() CandidateSourceLimits {
 	return CandidateSourceLimits{
 		LBSimilar:     60, // tier 1 — doubled; the only arm that measures the seed
 		SimilarArtist: 40, // tier 2 — raised; growing is always safe
 		TagOverlap:    20, // tier 2
 		UserCoplay:    20, // tier 2
-		LikesOverlap:  10, // tier 2, halved — flat 0.6 sim_score, see above
+		LikesOverlap:  10, // tier 2, halved — flat 0.3 sim_score, see above
 		TasteOverlap:  10, // tier 3 floor — halved, not removed
 		RandomFill:    10, // tier 3 floor — cut hard, never to zero
 	}