feat: autotag scoring overhaul, dump-based explore index, and lyrics search

Consolidates in-progress work across autotag, explore, and library:

- autotag: beets/Picard-informed scoring engine — ID-first matching, VA
  handling, recommendation tiers, and a merged distance/rank cascade, with
  an eval harness for regression tracking.
- explore: offline MusicBrainz dump import/incremental refresh replaces the
  legacy tier crawl; index-first local search with fuzzy matching and a
  dedicated ranker; disk-free guards for dump downloads.
- library: artist-credit extraction and matching.
- lyrics: owned-library lyric search (FTS) with LRCLIB backfill.

Also: rewrite README to be user-focused, and migrate upstream to
git.ljones.me/yonlu/yellowjacket.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/database/sql/schemas/lyrics_index.sql

@@ -0,0 +1,15 @@
+-- Contentless FTS5 index over recording lyrics, enabling
+-- "search by a lyric fragment → find the song".  The rowid is
+-- recordings.id.  Only recordings with non-empty lyrics are indexed.
+--
+-- content='' means the lyric text itself is NOT stored a second time
+-- (it already lives in recordings.lyrics); the index keeps only the
+-- tokenised inverted index, so it stays compact even for large
+-- libraries.  contentless_delete=1 lets us delete/reinsert a single
+-- row when a track's lyrics change (scan update or LRCLIB backfill).
+CREATE VIRTUAL TABLE IF NOT EXISTS lyrics_index USING fts5(
+    lyrics,
+    content='',
+    contentless_delete=1,
+    tokenize='unicode61 remove_diacritics 2'
+);

backend/database/sql/schemas/playlists.sql

@@ -3,6 +3,7 @@ CREATE TABLE IF NOT EXISTS playlists (
     name TEXT NOT NULL,
     is_smart INTEGER NOT NULL DEFAULT 0,
     smart_rules TEXT,
+    smart_snapshot_at DATETIME,
     created_at DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP,
     updated_at DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP
 );

backend/database/sql/schemas/tagging_candidates.sql

@@ -0,0 +1,18 @@
+-- tagging_candidates durably stores the scored candidate list for a
+-- tagging group so it survives process restarts.  Without it, only the
+-- top score (a single REAL on tagging_items) is persisted; the full
+-- candidate list — releases, tracks, alignments — is recomputed every
+-- session, re-hitting MusicBrainz whenever the short-TTL http_cache has
+-- expired.  The blob is written once when a group is first scored and
+-- read back on every subsequent open.
+--
+-- candidates holds the JSON-encoded []autotag.Candidate.  ON DELETE
+-- CASCADE ties the blob's lifetime to its tagging_items row: when a
+-- group's tracks change, the scan path deletes the old group_key row
+-- (and SQLite, with foreign_keys = ON, drops the stale blob with it).
+CREATE TABLE IF NOT EXISTS tagging_candidates (
+  group_key   TEXT PRIMARY KEY,
+  candidates  TEXT NOT NULL,
+  computed_at DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  FOREIGN KEY(group_key) REFERENCES tagging_items(group_key) ON DELETE CASCADE
+);