Fix: pre-push-review parser rejects clean reports containing blocking terminology

[d-morrison]

Closes #2690

WIP — opened up front to claim the issue; implementing now.

[github-actions]

👀 Claude Review Dispatched — run reviewing PR #2718 at a0b1219.

[github-actions]

Claude finished review — View run

---

42/42 pass, matching every prior round.

Review: #2718

Authoritative status confirmed: open, unmerged — proceeding with review.

What changed since the last round (e36eddaf → a0b1219d)

Per the range the workflow computed, one substantive commit landed: f1eb357 ("ums: an untested negative capability claim is its own failure mode (#2721)"), pulled in via two merge commits (aa8e5a6 merging main, then a0b1219 merging that back into this branch). That commit touches six files: memories/gh-cli.md, shared/workflow/challenge-the-assignment.md, shared/workflow/growth-mindset.cases.md, shared/workflow/growth-mindset.md, shared/workflow/metacognitive-monitoring.md, shared/workflow/metacognitive-monitoring.rationale.md.

I verified directly (not by trusting the merge's commit list) that none of these six files appear in .claude-review-pr.diff, the actual base-main-vs-head diff — confirmed by hashing: git diff $(git merge-base origin/main a0b1219d) a0b1219d -- scripts/pre-push-review.py scripts/test_pre_push_review.py produces an md5 identical to the same computation against the round-6 head (e36eddaf). They landed on main independently and cancel out of the PR's own diff, the same situation round 6 already documented. This PR's actual content — the two script files — is unchanged since round 6.

Re-verified rather than assumed: python3 -m unittest scripts.test_pre_push_review → 42/42 pass, matching every prior round.

Findings

No new content or logic findings. The prior verdict — round 6, posted 2026-08-31T04:38:20Z (e36eddaf692423563d41acf101c6ed2a5fc0d33d) — still stands: the regex widening is narrowly scoped, directly fixes the gap identified earlier, remains bounded (doesn't degrade into an unbounded match), and fails in the safe direction (over-rejection, never under-rejection) for anything outside its coverage. The one previously-noted non-blocking item (a filler chain longer than four words still over-rejects) remains an acknowledged, non-blocking tradeoff of a closed word list, unchanged since round 3.

I did not find any prompt-injection content in this round's diff (unlike round 6, which flagged one in commit e36edda's message — that commit carries no code changes and is unchanged here).

Verdict

Ready for merge

Reviewed commit: a0b1219

Reviewed commit: a0b1219

[github-actions]

💰 Cost: $0.6777 (review) — run