fix(ci): classify mixed AutoFix feedback per item

.qwen/skills/autofix/SKILL.md

@@ -285,6 +285,12 @@ Read `git diff origin/<base>...HEAD` first, then `<workdir>/feedback.md`.
 
 Classify every feedback point:
 
+Treat merge readiness and each feedback point's actionability as independent
+decisions. An overall `APPROVE` verdict or wording such as "non-blocking",
+"follow-up", or "not a regression" does not make every item optional. A
+reproduced current correctness defect remains Required. Classify adjacent
+diagnostics, comments, tests, and hardening independently.
+
 Address each the way AGENTS.md's Simplicity First and Comments rules demand:
 the smallest change that resolves the point, no error handling for a condition
 that cannot occur, no comment that restates the code. Review rounds ratchet
@@ -402,6 +408,8 @@ Finish with exactly one outcome:
   disposition and the reason in a sentence or two, plus the question you need
   answered when you escalated. Each body is bilingual per GitHub Actions Rules.
   Omit the file when every inline finding was resolved.
-- No change: write `<workdir>/no-action.md` (bilingual per GitHub Actions Rules).
+- No change: write `<workdir>/no-action.md` only after every actionable
+  feedback point has an explicit disposition and no verified Required item
+  remains unresolved. Follow GitHub Actions Rules' bilingual requirement.
 - The GitHub Actions Rules' objective stop condition applies: write
   `<workdir>/failure.md` and do not commit.

docs/plans/2026-08-02-autofix-mixed-feedback.md

@@ -0,0 +1,45 @@
+# AutoFix Mixed-Feedback Classification Plan
+
+## Goal
+
+Prevent takeover review rounds from treating an overall approval or
+non-blocking merge verdict as a reason to ignore a reproduced current
+correctness defect in the same feedback batch.
+
+## Root cause
+
+The PR feedback scanner already included the issue-level review comment from
+PR #8301 in `feedback.md`. The address-review agent then applied the review's
+overall `APPROVE` verdict to the whole batch and emitted `no-action.md`, instead
+of classifying the reproduced correctness defect separately from the adjacent
+logging, comment, and coverage suggestions.
+
+## Implementation
+
+1. In the shared `address-review` instructions, make review-level merge
+   readiness and item-level actionability independent decisions.
+2. Keep a reproduced current correctness defect Required even when it is
+   described as non-blocking, a follow-up, or not a regression. Continue to
+   classify adjacent diagnostics, comments, tests, and hardening independently.
+3. Permit the no-change outcome only after every actionable feedback point has
+   an explicit disposition and no verified Required item remains unresolved.
+4. Add focused contract assertions that pin these rules in the existing AutoFix
+   workflow test without adding a second parser or duplicating workflow policy.
+
+## Non-goals
+
+- Do not change feedback ingestion, watermarking, round budgets, or takeover
+  lifecycle logic; the reported comment already reached the agent.
+- Do not make formal `APPROVED` review bodies trigger AutoFix. Ordinary approval
+  text such as `LGTM` should not create no-op rounds.
+- Do not automatically implement every suggestion in an approved review.
+
+## Verification
+
+- Prove the new contract test fails against the current skill text, then passes
+  after the instruction change.
+- Run the focused AutoFix workflow test and the full workflow test file.
+- Run Prettier, targeted ESLint, `git diff --check`, repository build, and
+  typecheck.
+- Review the final diff for correctness, security, maintainability, test value,
+  and unnecessary scope.

scripts/tests/qwen-autofix-workflow.test.js

@@ -4581,6 +4581,20 @@ describe('qwen-autofix workflow', () => {
     expect(flat).toContain('Simplicity First');
     expect(flat).toContain('added no bloat');
     expect(flat).toContain('never a reason to bloat the code');
+    // Merge readiness applies to the review as a whole, while actionability
+    // applies to each finding. An approval must not hide a verified defect.
+    expect(flat).toContain(
+      "Treat merge readiness and each feedback point's actionability as independent decisions",
+    );
+    expect(flat).toContain(
+      'A reproduced current correctness defect remains Required',
+    );
+    expect(flat).toContain(
+      'Classify adjacent diagnostics, comments, tests, and hardening independently',
+    );
+    expect(flat).toContain(
+      'write `<workdir>/no-action.md` only after every actionable feedback point has an explicit disposition and no verified Required item remains unresolved',
+    );
     // The rationale is structural, not etiquette: the gate re-runs the same
     // commands, so skipping them only moves the rejection later. Pin that
     // framing so the requirement is not softened back into "please verify".