Summary

A source's last_commit bookmark can point at a commit that no longer exists in the source repo (orphaned by a history rewrite / force-push / branch cleanup). When this happens, gbrain sync silently degrades to a full repo walk on every run, forever, never advancing last_sync_at. On a large brain with a cross-region DB this never completes (see #1958), so the source goes permanently stale with zero error surfaced.

Repro

1. Source default has sources.last_commit = <X>.
2. The source repo's history is rewritten so <X> is no longer reachable (e.g. a master→main consolidation that drops <X>; force-push; squash; branch delete).
3. Run gbrain sync --source default.

Observed: sync reports Large sync (NNN,NNN files) and walks the entire tree every run. git diff <X> HEAD fails internally because <X> is not a valid object. last_commit / last_sync_at never advance. No error, no warning — just eternal "in progress" that the cron timeout keeps killing.

Expected: sync detects the unreachable bookmark and either (a) reseeds to the nearest reachable ancestor by date and diffs forward, or (b) does a one-time full reconcile and then stamps the bookmark to HEAD — and in both cases emits a clear [sync] bookmark <X> unreachable (history rewritten) — re-pinning log line.

Root cause

The incremental path resolves lastCommit = readSyncAnchor(...) and then git diff lastCommit..HEAD. There is handling for a checkpoint target that's no longer reachable (merge-base --is-ancestor → discard + re-pin to HEAD), but no equivalent guard for the persisted last_commit bookmark itself. An unreachable bookmark falls through to the full-walk fallback every time instead of being detected and re-pinned.

Suggested fix

In the incremental sync entrypoint, before diffing:

if (lastCommit && !isAncestor(repoPath, lastCommit, headCommit)) {
  slog(`[sync] last_commit ${lastCommit.slice(0,8)} unreachable (history rewritten) — re-pinning to HEAD and reconciling via content_hash`);
  lastCommit = null;            // forces a content-hash reconcile, not an eternal blind re-walk
  // (content_hash already short-circuits unchanged files at ~10ms each)
}

Mirror the existing checkpoint-target reachability check (merge-base --is-ancestor) for the bookmark. Optionally also reseed to the nearest reachable commit ≤ last_sync_at to minimize the reconcile set.

Real-world impact (this brain)

• last_commit pointed at cd8732419775d93f982248820cd42b8c46892df6, orphaned by a master→main cleanup on 2026-05-31.
• Every sync for 4 days walked all ~150K files and committed nothing; last_sync_at froze at 2026-06-04 03:39.
• Manual fix that worked: reseed last_commit to a reachable commit at the same date → sync diffed only ~1,510 changed files and completed, stamping last_sync_at.
• Compounds with Sync can never complete: DB is cross-region (us-east-1) from compute (us-west-2) — 71ms/query × ~90 queries/file = 6.5s per 2KB page → full sync ~14 days → last_sync never commits → permanent staleness #1958 (cross-region latency makes the blind full-walk physically unable to finish under any cron timeout).

Related

• Sync bookmark permanently stuck when importer throws on non-string frontmatter title (content-sanity.ts:379) — silent indexing outage #1939 — sibling silent-staleness mode (poison file freezes the bookmark).
• Sync can never complete: DB is cross-region (us-east-1) from compute (us-west-2) — 71ms/query × ~90 queries/file = 6.5s per 2KB page → full sync ~14 days → last_sync never commits → permanent staleness #1958 — cross-region latency (why the full-walk fallback can never finish here).