fix(databases): complete the dead restartNumber reindex-recovery trigger

[heskew]

Summary

table() re-triggers a secondary-index backfill when the persisted attribute descriptor's restartNumber is older than the current restart generation. This is the PID-reuse-robust complement to the indexingPID check: it recovers an index left incomplete by a worker that crashed mid-backfill, even when the replacement worker reuses the old PID. The bug: restartNumber was never persisted onto the descriptor, so the comparison was always undefined < N (false) and the trigger was dead.

This persists restartNumber symmetrically with indexingPID — set on backfill start, carried across metadata-only changes during an in-progress build, cleared on clean completion (retained on park/crash).

Resolves #1359 (part of #1354).

Where to look

• resources/databases.ts — table() (set/carry + the two trigger sites) and runIndexing() (clear on clean completion).
• Operand decision (worth a look): the trigger compares against workerData?.restartNumber ?? manageThreads.restartNumber — the owning worker's stable startup generation, falling back to the maintained manageThreads counter on the main thread. Cross-model review (Codex) flagged that stamping the mutable manageThreads.restartNumber is wrong during a worker's shutdown/drain window: the global counter has already advanced to the replacement generation, so an index built then would be stamped with the replacement's generation and the replacement worker would see an equal generation (+ a possibly-reused PID) and never recover it. Using the spawn generation avoids that.

Back-compat

Existing descriptors have no persisted restartNumber → undefined < N is false, so upgrading does not trigger a reindex storm. On the main thread the generation is 1, so 1 < 1 is false (inert until an actual restart bumps the counter).

Tests

unitTests/resources/indexRestartNumber.test.js (RocksDB-only): (1) restartNumber is persisted on a parked/incomplete index and cleared after a clean completion; (2) a backfill re-triggers when the persisted generation is older than the current one, and not when equal. Re-trigger is asserted via indexingOperation promise identity (the promise carries across resetDatabases, so truthiness is unreliable — mirrors the existing F1 test).

---

🤖 Generated with Claude Code (Opus 4.8)

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[claude]

Reviewed; no blockers found.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[kriszyp]

Dead restartNumber path correctly wired up, good crash-recovery test coverage. LGTM! 🟢

Reviewed by Claude Sonnet 4.6