perf(core): stop rebuilding the whole messages_fts index in every incremental finalize

[Cyb3rN8]

Fixes #75.

Root cause and reproduction

Every finalize runs messages_fts('rebuild'), regenerating the FTS index from the whole messages table. Cost is proportional to index size, not change size: measured on a 1.8 GB / 688k-message index, ~10.3 s under unicode61 and ~30.2 s under trigram (#14), out of a ~50 s incremental build. File stat-comparison (8,669 jsonl, 0.07 s) is not the bottleneck — the rebuild is.

The rebuild is redundant on the incremental path: messages_fts is external-content maintained by the ai/au/ad triggers, persist upserts messages via ON CONFLICT DO UPDATE (fires AU) and deletes via plain DELETE (fires AD). No current write shape bypasses the triggers; FTS5 integrity-check (rank=1) passes on the live index.

Change

• Incremental finalize performs the rebuild exactly once per index (heals rows that predate trigger-only maintenance, e.g. written by an INSERT OR REPLACE-era version) and records __messages_fts_synced__ in index_state in the same transaction — an interrupted build redoes the heal; the marker never runs ahead of the work (converges when re-run, per the contributing guide's destructive-work rule).
• The force path keeps its wholesale rebuild as the last line of defence and re-records the marker it wipes with DELETE FROM index_state.
• rebuildMemoryFts is deliberately untouched: memories is tiny, and --attune writes it outside build transactions, so the cheap belt-and-suspenders refresh stays.

Verification

• New tests/incremental-fts-sync-marker.test.mjs:
  • poison-posting probe (an FTS row with no messages counterpart): survives an incremental build (rebuild skipped), cleared by a force build, cleared exactly once for a marker-less index (the heal), with integrity-check after each path;
  • pins the trigger semantics for persist's exact upsert/delete shapes — the writes the wholesale rebuild used to bail out.
• Full suite on this branch head: 469/469 pass, typecheck clean, lint 0 errors (4 pre-existing warnings in unrelated app tests, present on main).
• Real-machine effect (daemon-style refresh loop): incremental build 40–56 s → 10.6–11.4 s with real changes, 2.4 s floor with none; search behavior verified unchanged against LIKE ground truth.

🤖 Generated with Claude Code