Self-review: Remote sub-tasks hear remote state changes between turns

The self-review node's proof. One row per round, per reference/reviewing.md. The operator's policy (the-loop critic policy) is 2 self rounds and 1 critic round, stopping on a round with no new findings and escalating on a repeat.

Review cycles

Round	Reviewer	Outcome	Findings → disposition	Link
1	[claude/claude-opus-5-5] (a fresh-context sub-agent, read-only)	new findings	9 code findings and 5 test gaps. Will fix: 1–7, 9 and the test gaps. Won't fix: 8 (refusal timing; ids are not secret, recorded in requirements)	PR comment
2	[claude/claude-opus-5-5] (a fresh-context sub-agent, read-only, told round 1's findings)	new findings	4 findings, all will-fix: (1) an IndexError after a failed fetch on the last recheck; (2) a cancel lost across a rollover; (3) a stale claim from a push during a turn; (4) the no-spin test could not tell the fix from the bug. Fixing the test gaps (follow_remote_task unit tests) also showed that a missing agent was retryable, which defeated round 1's fix 3; it is now non-retryable.	PR comment

Round 2 was the last self round the operator's policy allows (selfReviewCount: 2), so the review moved to the critic with round 2's fixes applied. Each round's fixes were committed together (921f679 for round 1, and the round-2 commit), not one commit per finding: the fixes touched the same functions. The PR comments carry the per-finding dispositions.