v1.6.4: cover-art identification + messy-name recovery

[DynamycSound]

Cover-art identification (new)

A file with no usable artist anywhere ("Bounce", artist Unknown) is textually indistinguishable from every other same-titled song — that's how it saved Russian lyrics from an unrelated "BOUNCE". Its embedded cover art, however, is distinguishable, and the user confirmed the file's thumbnail matches the real song's cover.

CoverIdentity.discover() searches iTunes/Deezer widely (up to 25 rows per service via the new LastResortAPI.identityCandidates(), vs. top-3 in the canonicalizer — "Bounce" by Voyage sits at ~#18) with the bare title, hashes each result's album art (existing aHash from CoverArtCompare), and treats a perceptual match + compatible runtime (±15s) as the file's identity. The discovered identity is injected as:

• the leading query candidate (new MatchStrategy.COVER_IDENTITY) — the search runs "Voyage Bounce" instead of a blind "Bounce" (LRCLib has it, synced, verified live), and
• an artist guess the wrong-singer veto can finally use.

Wired into both batch (LyricsUtils) and single-song (LyricsFetchViewModel). Bounded: ≤2 queries, ≤12 cover downloads per song, only runs when there are no artist guesses (or on demand, below).

Cover second opinion for runtime coincidences

"CHECK" by WAV3POP got Portuguese lyrics from a same-duration "Check" by The Boy — it survived v1.6.3's veto through the exact-runtime escape. In batch, a hit whose artist agrees with none of the artist guesses now triggers cover identification before being trusted: if the file's art identifies a different artist, the hit is skipped. No art / no identification keeps existing behaviour (the escape is load-bearing for junk-artist rips). Verified live: Deezer has "CHECK" by WAV3POP (140s, with art); LRCLib has no WAV3POP entry → honest no-lyrics instead of an impostor.

Messy-name recovery (each verified against the live catalogues)

• ????/U+FFFD runs (chars the filesystem couldn't store) stripped from queries — "DEVITO - FLEX ????" → "Taj flex" by Devito is on LRCLib.
• Unknown bracket junk (incl. an unterminated trailing clause left by earlier cleanup) falls back to a bare-title candidate — "CRNI CERAK - CC v1.5.0: background batch download, animated progress screen, matching and stability fixes #2 (JUŽNI VETAR 2022)" → "CC v1.5.0: background batch download, animated progress screen, matching and stability fixes #2" is on LRCLib.
• Every collab part becomes a candidate + veto guess, not just the first — "GRCA X FOX - BELE ŠARE" is indexed under Fox (GRCA is the guest); Fox must not look like a stranger.
• YouTube channel decorations (- Topic, TV, Official…) stripped into an extra tag-artist reading — "CHECK" / "WAV3POP - Topic" → artist "WAV3POP" reaches providers. Raw value always kept too.

Version

1.6.3 (163) → 1.6.4 (164), release notes rewritten.

Testing

• 5 new tests in MessyNameRecoveryTest (all four reported name shapes + the no-guess/no-veto invariant) — pass.
• WrongSingerVetoTest (8) and BadMetadataMatchTest (7) — pass, no precision/recall regressions.
• Full matching suite green except the pre-existing live-network SuicideboysMatchTest.loot (LRCLib data drift, fails on unmodified master).
• :app:compileDebugKotlin clean.
• Live API probes for every claim above (LRCLib/Deezer/iTunes) run during development.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QADiJzVyNycc1w28sirc1a

---

Generated by Claude Code