Fail fully-clean when any reviewer's latest is not-clean

[dem-extra1]

Closes #2274.

If one review is all-clear and another raises blocking issues, nits, or any other flagged items, do not merge, even with mwc active. ARD every item from every review, then request fresh reviews.

This encodes that in the existing fully-clean / mwc path, not a parallel process:

• scripts/check-pr-fully-clean.py criterion 4 is per-reviewer. A later all-clear from a different reviewer does not supersede a standing not-clean. Claude and Antigravity share github-actions[bot], so identity is the first-line agent marker; Jules is keyed on login.
• ## Nits / **Nits** / ## Non-blocking count as findings even when ### Verdict says Ready for merge, including at an earlier SHA.
• skills/mwc/SKILL.md Scope Limit, shared/workflow/fully-clean.md, skills/ardi/SKILL.md, and the Strict Merge sections in AGENTS.md / CLAUDE.md state the same veto.

Same-reviewer iterate still works: a later clean from the same reviewer supersedes. A later GitHub APPROVED clears that author's own earlier CHANGES_REQUESTED (Copilot/Jules), not a sibling bot's.

Independent of open PRs that also touch some of these files (#2256, #2272, #2243, #2254); this targets main.

  

[cursor]

Fallback self-review after claude-review skipped (API credential or quota unavailable). This is the dispatched adversarial-reviewer report for HEAD e5ab4a915d9034e45dbd1e07a5fd7192b553c3bb. It does not satisfy mwc.

Summary of Changes

The branch makes fully-clean per-reviewer: a later all-clear from one reviewer does not clear another reviewer's standing not-clean, including nits and including under mwc. check-pr-fully-clean.py keys shared-login bots on the leftmost agent marker on the first non-empty line, treats ## Nits / ## Non-blocking as findings even beside a Ready verdict, and fails when any reviewer's latest is not-clean. Docs and the ardi / merge-it / mwc skills are updated to match. python3 scripts/test_check_pr_fully_clean.py reported 202 passed, 0 failed.

Findings

No actionable findings identified.

Verdict: Ready for merge

Reviewed-Commit: e5ab4a9

Posted by Cursor (AI agent) --- not written by a human.

[cursor]

Driving this PR to clean --- please hold off until done.

Posted by Cursor (AI agent) --- not written by a human.

[cursor]

Fallback self-review after claude-review skipped (API credential or quota unavailable). This is the dispatched adversarial-reviewer report for HEAD 1e5786395b463c9c5a3c23dcfb8e502201ac54b7. It does not satisfy mwc.

Summary of Changes

HEAD 1e5786395b463c9c5a3c23dcfb8e502201ac54b7 is a merge of origin/main (13061c91, #2255) onto d9b2f210. The unique file set versus origin/main is the same 14 files; those 14 are byte-identical to d9b2f210. The #2255 paths match origin/main and are not in this PR's unique diff.

The unique change makes fully-clean criterion 4 per-reviewer rather than globally last: check-pr-fully-clean.py groups by _reviewer_identity (exclusive bot login first, else the leftmost known agent marker on the first non-empty line, else login), fails when any reviewer's latest is not-clean (nits included), and only lets a later GitHub APPROVED clear that same identity. Docs and mwc/ardi/merge-it follow that contract. skills/ardi/SKILL.md fetches issue comments by body marker (**Claude finished|### Verdict|Antigravity Agent Report) and emits every match, not | last.

Confirmed in this tree: per-reviewer grouping and _approval_clears; the ardi body-marker fetch (jq -s / .[][] holds on one- and two-page shapes); the unmarked shared-login residual (docstring plus test: unmarked github-actions does not clear marked Claude). python3 scripts/test_check_pr_fully_clean.py: 204 passed, 0 failed.

Findings

No actionable findings identified.

Verdict: Ready for merge

Reviewed-Commit: 1e57863

Posted by Cursor (AI agent) --- not written by a human.

[cursor]

Driving this PR to clean --- please hold off until done.

Posted by Cursor (AI agent) --- not written by a human.

[cursor]

Driving this PR to clean --- please hold off until done.

Posted by Cursor (AI agent) --- not written by a human.

[cursor]

Driving this PR to clean --- please hold off until done.

Posted by Cursor (AI agent) --- not written by a human.

[d-morrison]

The prior hold-off claim expired (no push or comment since 15:03 UTC, past the 2-hour convention), so taking this over: driving this PR to clean --- please hold off until done. A fresh Claude review has been dispatched at the current head.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Pushed 33133fc; the push-triggered review will evaluate the new head.

ARD for the two rounds since taking this PR over:

#	Round	Finding	Disposition
1	claude-review at 1e57863	NOT_CLEAN_NEGATION_SUFFIX false-positives on bold-wrapped **None.** under ## Nits/## Non-blocking, permanently recording a clean round as a standing not-clean	Addressed --- emphasis markers now tolerated before the negation, with regression tests for the three bold forms.
2	pre-push adversarial round	The first fix's shared-prefix widening extended the unguarded nothing/0/no ... branches to bold text, so a bold "Nothing major, but [real finding]" was swallowed --- the dangerous direction	Addressed --- alternation split: emphasis tolerance only for the whole-negation alternatives (none/n/a); the risky branches keep their plain prefix. The round's four counterexamples are now negative-control tests.
3	pre-push adversarial round	No negative-direction tests accompanied the first fix	Addressed --- four negative controls added (211/211 passing).
4	pre-push adversarial round (root-cause note)	The plain-text nothing/0/no ... branches were always unguarded --- predates this PR	Deferred --- tracked as a separate checker hardening in the issue filed today: #2357.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Correction to the ARD table above: the deferred checker-hardening issue is #2359, not #2357 --- the comment was written before the issue's number was assigned.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[d-morrison]

Pushed 25a2694. ARD for the adversarial round at 33133fc (verdict: Ready for merge, three non-blocking observations):

#	Finding	Disposition
1	Unmarked shared-login reviewers still collapse into one identity bucket (disclosed residual of the fixed bug class)	Deferred --- tracked in #2362 with the reviewer's reproducer and candidate fixes; not testable until the design changes.
2	Illustrative jq marker match case-sensitive, unlike the checker itself	Addressed --- "i" flag added in 25a2694.
3	Whole-history finding scan now admits any bot's headings (safe direction, but undocumented/untested widening)	Deferred --- same issue #2362, second item.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[github-actions]

Claude finished review — View run

---

Confirmed: that's the only change since the last reviewed commit.

Summary

PR #2298 makes check-pr-fully-clean.py's criterion 4 evaluate every reviewer's latest verdict independently (via new _reviewer_identity() / _approval_clears() helpers) instead of only the globally-last comment, so a later all-clear from one reviewer no longer masks a standing not-clean from another, including under mwc. It also extends the shared finding-pattern scan to treat ## Nits / **Nits** / ## Non-blocking headings as findings even beside a "Ready for merge" verdict. Docs across AGENTS.md, CLAUDE.md, memories/preferences.md, several shared/workflow/*.md files, and skills/ardi|mwc|merge-it/SKILL.md are updated to state the same veto rule.

Findings

None. The prior review round (comment posted 2026-08-26T22:15:26Z, #2298 (comment) — reviewed commit 33133fc7434cbc2340c3c7c7be30cdecc7a90b95) verified the previously-reported bold-negation false positive was fixed and gave Ready for merge, with the full test suite passing (211/211).

Since that commit, exactly one further commit landed: 25a26942 ("fix: case-insensitive jq marker match in ardi recipe"), which changes a single line in skills/ardi/SKILL.md — adding jq's "i" (case-insensitive) flag to an illustrative test(...) command, so the documented command matches the Python checker's own re.IGNORECASE behavior for verdict-marker detection. I reproduced the fix directly:

echo '{"body":"### VERDICT clean"}' | jq -s '[.[] | select(.body | test("### Verdict"; "i"))]'   # matches
echo '{"body":"### VERDICT clean"}' | jq -s '[.[] | select(.body | test("### Verdict"))]'          # empty, without the flag

This is a correct, minimal, doc-only change with no functional/code impact — no new logic bugs, no CLAUDE.md violations, and no hallucinated references introduced.

Verdict

Ready for merge — the prior verdict (#2298 (comment), reviewed commit 33133fc7) stands; the only change since then is a verified, correct one-line doc fix with no functional impact.

Reviewed commit: 25a2694

Reviewed commit: 25a2694

[github-actions]

💰 Cost: $1.1922 (review) — run

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.

_{2026-08-26 22:37 UTC}