feat: concurrency hardening — fine-grained ingest transactions + fast-path DB init + write retry #124

[four-bytes-robby]

Closes #124

Changes

1. Fine-grained per-file transactions in ingestPath() — replaced giant SAVEPOINT (write lock held for minutes) with per-file SAVEPOINTs (write lock held for milliseconds). Other write operations (memory, knowledge) no longer blocked during ingest.
2. Fast-path DB initialization — added getSchemaVersion() + version check in initBrainDatabase(). When DB is already at current schema version, skips createSchema() and runMigrations() entirely.
3. Write retry wrappers — brain_kb_add, brain_kb_record, brain_kb_review, and brain_reindex now use withDbRetry() for SQLITE_BUSY handling.
4. Removed migration helpers from createSchema() — destructive migrateKnowledgeFts() (DROP+CREATE entries_fts) no longer runs on every init. Migration helpers only run in runMigrations() during schema upgrades.

Verification

• ✅ bun run build — success
• ✅ bun test — 247 pass, 0 fail
• ✅ Version bumped to 1.7.3

---

Summary by cubic

Improves database concurrency and startup time. Ingest now uses per-file transactions, DB init short-circuits when the schema is current, and write paths auto-retry to avoid SQLITE_BUSY.

• Refactors
  • Per-file SAVEPOINTs in ingest to avoid long write locks; other writes continue during ingest.
  • Fast-path init: check schema_version via getSchemaVersion(); skip createSchema() and migrations when up to date.
  • Write retries: wrap knowledge writes and reindex in withDbRetry() to handle SQLITE_BUSY.
  • Remove migration helpers from createSchema(); destructive FTS rebuild runs only during migrations.

^{Written for commit cfb08a6. Summary will update on new commits.}