fix(hooks): anchor fence-marker detection to line start in has_duplicate_waiver

[tvna]

Summary

has_duplicate_waiver() in hooks/gitapex_check_pr_duplicate_issue.py misdetected a fence-shaped Duplicate-PR-waiver: reason (e.g. ~~~) as opening an unterminated code fence, silently stripping the reason before its own regex ever saw it. Ports the exact, already-adversarially-hardened fix issue #1432/PR #1440 applied to the sibling copy of this same defect class.

Facts

• has_duplicate_waiver("Some PR description.\n\nDuplicate-PR-waiver: ~~~\n") returned False (expected True), root-caused to _UNTERMINATED_FENCE_RE/_FENCE_RE matching a fence marker anywhere in the text rather than only at the start of its own line -- confirmed reproducing identically and deterministically on origin/main (commit 2ce6c2f5) via direct function call, independent of any PR.
• This is currently a required pytest CI check failure (a derandomize=True Hypothesis test, the same failing example every run) on PR feat(drafting-a-skill): close Dimension 4-8 gaps + redivide edit responsibility #1632, which does not touch this file at all -- diagnosed there, disclosed via a PR comment, and filed here as issue fix(hooks): has_duplicate_waiver misdetects a fence-shaped waiver reason as an unterminated code fence #1714 rather than widening that PR.
• Issue fix(drafting-issues): has_dedup_disclosure misdetects a fence-shaped Dedup reason as an unterminated code fence #1432/PR fix(drafting-issues): anchor fence-marker detection to line start #1440 already found and fixed the identical bug class in the sibling copy (skills/drafting-issues/scripts/gitapex_check_acm_present.py's own _FENCE_RE/_UNTERMINATED_FENCE_RE, which this file's own copies previously mirrored verbatim) -- that fix's own two adversarial-review rounds additionally found and closed a list-item/blockquote-prefixed-fence inversion bug and a ReDoS shape (CWE-1333) a naive line-start anchor alone would have reintroduced. This PR ports that same, already-hardened _CONTAINER_PREFIX construction rather than re-deriving it.

Assumptions

• None beyond what issue fix(hooks): has_duplicate_waiver misdetects a fence-shaped waiver reason as an unterminated code fence #1714's own Acceptance Criteria Map states; this PR follows that ACM's single row directly.

Risk / blast radius

Scoped to hooks/gitapex_check_pr_duplicate_issue.py and its own test file only. This hook only guards mcp__github__create_pull_request (denying a duplicate-issue PR without a waiver) -- widening what counts as a valid waiver reason is a narrowing-the-false-positive-rate change, not a new false-negative: every previously-detected waiver and every previously-stripped genuine fenced block is re-verified unchanged by the existing test suite plus new regression tests.

Rollback

Revert this PR's merge commit; no schema, data, or external-service change.

Verification

Restating issue #1714's own Acceptance Criteria Map:

Criterion	Proof method	Result
has_duplicate_waiver correctly recognizes a Duplicate-PR-waiver: line's own disclosed reason even when that reason starts with or contains fence-marker characters that do not actually open a real fence, and every already-adversarially-verified case from issue #1432/PR #1440's own fix holds identically	Ported the exact _CONTAINER_PREFIX-based construction from PR #1440 into this file's own _FENCE_RE/_UNTERMINATED_FENCE_RE; regression tests mirroring that fix's own defeat-test set (fence-shaped reason, list-item/blockquote/ordered/+-fenced rejection, bare-CR-fenced rejection, ReDoS wall-clock budget), strengthened across Step 8's own three review rounds below after two were found vacuous and fixed	Done. uv run --frozen python3 -m pytest -q --no-cov hooks/test_gitapex_check_pr_duplicate_issue.py tests/test_gitapex_check_pr_duplicate_issue_properties.py: 50/50 passed, including the previously-failing property test. ruff check/ruff format --check/mypy: clean. All 45 gitapex_gate_local_preflight.py gates passed on commit 2529c6e4.

Independent review verdict

• Verdict: CLEAN
• Verified commit: 2529c6e

Outer layer (GitHub-native reviewer). GitHub Copilot review re-requested (request_copilot_review) at approximately 2026-09-03T10:37Z. No response posted at request time. Per the operator's explicit instruction, the 30-minute wait for this round was skipped rather than observed -- disclosed as unreachable/skipped for this run, not silently omitted.

Inner layer (reviewing-an-artifact, deferred per its own Step 0 to evaluating-deterministic-gate-quality). This PR's diff is unchanged since the prior round -- still entirely a deterministic gate and its own test file (hooks/gitapex_check_pr_duplicate_issue.py, registered in .gitapex/ssot.json as gate id pr-duplicate-issue) -- so Step 0 again deferred to that specialist. A fresh, isolated Agent dispatch (no memory of the prior rounds, general-purpose subagent with Bash access for live testing) independently re-derived the verdict from the current code rather than trusting the prior round's own recorded result.

Verdict: well-formed and well-placed. Confirmed the gate is correctly wired (.gitapex/ssot.json:279-293, hooks/hooks.json:71, matcher mcp__github__create_pull_request) and that this diff only changes the fence-detection regexes inside an already-correctly-placed gate.

Live-tested directly against the checked-out module (not read-only inference):

• pytest hooks/test_gitapex_check_pr_duplicate_issue.py tests/test_gitapex_check_pr_duplicate_issue_properties.py: 50 passed, target module at 100% line coverage.
• Mid-line (non-line-start) fence-shaped waiver reason: correctly detected as a real waiver (not stripped).
• Genuine line-start fence containing a fabricated waiver: correctly stripped.
• Backreference cross-type check (open ``` / close ~~~): correctly treated as unterminated, not falsely paired.
• List-item, blockquote, ordered-marker, and +-marker fences: fabricated interior waivers correctly stripped; real waivers outside such blocks still detected.
• ReDoS stress tests beyond the PR's own bundled case, across 5 adversarial payload shapes up to ~40KB: no catastrophic-backtracking behavior found in any.
• re.MULTILINE anchor semantics confirmed correct across multi-line spans.

No bypass (fail-open) scenario found; the targeted bug class is closed.

Finding (non-blocking, disclosed, not fixed in this round). _CONTAINER_PREFIX and its two dependent regexes now exist verbatim in two files (hooks/gitapex_check_pr_duplicate_issue.py and skills/drafting-issues/scripts/gitapex_check_acm_present.py) with no automated parity/sync test. This is a pre-existing pattern this PR ports rather than a regression it introduces, and the reviewing agent explicitly characterized it as not blocking. Not treated as a confirmed finding requiring a Step 3 loop-back; recorded here as disclosure for a human reader. A small parity test (asserting the two _CONTAINER_PREFIX values stay identical) is a reasonable follow-up, potentially in scope for issue #1721 (already filed for the sibling-file drift) or a new issue.

Why CLEAN. Zero confirmed findings against the current diff from either layer. The one disclosed finding is explicitly non-blocking and pre-existing, not a defect in this PR's own change. pytest is green. See the archived prior round (PR comment) for the full three-round mutation-testing account that closed the original bug class.

Skill audit evidence

• deterministic-gate-quality: RAN -- hooks/gitapex_check_pr_duplicate_issue.py is registered as a deterministic gate (.gitapex/ssot.json gates[].script). Read this diff against skills/evaluating-deterministic-gate-quality/references/dimensions.md dimension 15 (fail-closed default on incomplete or malformed input) specifically, per that dimension's own instruction not to credit it from the bundled test suite's happy-path coverage alone. Independently constructed and ran, live, directly against the current code rather than from memory: has_duplicate_waiver(None) -> False (no crash, falls through to the normal duplicate-citation check rather than silently allowing); a malformed-JSON stdin payload -> error: payload is not valid JSON..., exit 1; a resolving-citation payload with GH_TOKEN/GITHUB_TOKEN unset -> FAIL: cannot verify #1 ... failing closed, exit 1; a wrong-typed body field (int) -> error: ... field 'body' must be a string ..., exit 1; a pathological 200-space-padded unterminated-fence input against the new/changed _FENCE_RE/_UNTERMINATED_FENCE_RE -> completed in 0.0001s (no ReDoS). Every case denies or safely no-ops rather than silently allowing. This PR's own diff changes only the fence-detection regexes and their own test coverage inside has_duplicate_waiver; the pagination-bound, missing-token, and malformed-payload fail-closed paths verified above are all pre-existing and unmodified by this diff. See ## Independent review verdict above for the full three-round mutation-testing account this dimension's own live-testing requirement was additionally satisfied by.
• defeat-test-disclosure: RAN -- every _CONTAINER_PREFIX/_WAIVER_RE branch this file's own tests claim to cover was independently mutation-tested in Step 8's own Round 3 (see ## Independent review verdict above): each of six targeted mutants broke exactly one branch, and exactly the test(s) naming that branch failed -- proof the defeat tests genuinely defeat the logic they claim to, not merely exercise its happy path. Two tests (the bare-CR and blockquote cases) were found vacuous during this same process and rewritten to assert on _strip_fences()'s own output directly rather than through has_duplicate_waiver() alone, closing the gap a first pass had missed.

Checklist

• Tests pass locally
• Docs updated if behavior changed
• Issue number cited in every commit
• If this PR adds/modifies a skills/*/SKILL.md, a docs/superpowers/specs/*.md design doc, a security-relevant skill, or a deterministic checker script under skills/*/scripts/*.py, evals/scripts/*.py, or .github/scripts/*.py (the checklist's own enumerated scope) -- this PR's own files are under hooks/, but hooks/gitapex_check_pr_duplicate_issue.py is independently registered as a deterministic gate (.gitapex/ssot.json), so the ## Skill audit evidence section above discloses deterministic-gate-quality and defeat-test-disclosure as required
• N/A: no evals/*/split.md Kept-edit-log entry added
• N/A: no skills/*/SKILL.md Stop-boundary/branch change

Merge gate: independent review

This PR is also subject to the independent-review-pending required
status check (see .github/workflows/independent-review-pending.yml /
.github/scripts/gitapex_gate_independent_review_pending.py). It stays
pending/failing until a ## Independent review verdict section naming
this PR's current head commit is recorded in this body --
drafting-a-pr-to-merge's own Step 8 records it once its independent
review completes. There is nothing for you to do here now: do not
pre-fill this section yourself, and do not remove this note.

Related Issue

Closes #1714

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Team

Run ID: 25761cab-5fc2-4ec8-8abd-ba1b241566cf

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.67%. Comparing base (45e265e) to head (2529c6e).
⚠️ Report is 17 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1715   +/-   ##
=======================================
  Coverage   99.67%   99.67%           
=======================================
  Files         153      153           
  Lines       24963    24990   +27     
  Branches     3010     3010           
=======================================
+ Hits        24881    24908   +27     
  Misses         82       82           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tvna]

Independent review verdict (archived -- round ending at commit 2529c6e)

• Verdict: CLEAN
• Verified commit: 2529c6e

Outer layer (GitHub-native reviewer). GitHub Copilot review requested (request_copilot_review) at approximately 2026-09-03T05:30Z. No response posted within the 30-minute window (checked well past that window via get_reviews: empty). Per drafting-a-pr-to-merge Step 8's own rule, treated as unreachable for this run -- disclosed, not silently omitted.

Inner layer (reviewing-an-artifact, deferred per its own Step 0 to evaluating-deterministic-gate-quality). This PR's diff is entirely a deterministic gate and its own test file (hooks/gitapex_check_pr_duplicate_issue.py, registered in .gitapex/ssot.json), so Step 0 deferred the full review to that specialist rather than running the ordinary five-persona fan-out. Three isolated dispatches (fresh Agent subagent, no memory of authoring this diff, working from a snapshot directory outside this repository's own CLAUDE.md/AGENTS.md ancestry, Read/Grep/Glob/synthetic-local-Bash only) ran across this Step 8 round, each verifying the prior round's own fix from scratch rather than only re-confirming its claim -- the same discipline this PR's own sibling PR #1632 established this session.

Round 1. Verdict: well-formed and well-placed, with two disclosed findings. (1) test_waiver_line_rejected_inside_a_bare_cr_fenced_block's docstring claimed to pin _CONTAINER_PREFIX's own \r alternative, but has_duplicate_waiver() normalizes CR to LF before _strip_fences() ever runs, so that test only exercised normalization, not the regex branch it named -- live-confirmed by rebuilding _CONTAINER_PREFIX without the \r alternative and observing the test still pass. (2) A sibling module this hook imports from, hooks/gitapex_check_pr_issue_acm_disclosure.py, still carries the pre-fix, line-start-unanchored _FENCE_RE -- a real, currently-live instance of the same bug class in a different file/function (citation extraction, not waiver detection), confirmed by direct read. Fixed (in scope): corrected the bare-CR test's docstring and added test_strip_fences_rejects_a_bare_cr_fenced_block_without_normalization, calling _strip_fences() directly to genuinely pin the \r branch -- confirmed live to fail without that branch present. Also added a blockquote-fence regression test to close a coverage gap the same round flagged (the > alternative was live-verified correct but untested). Disclosed, not fixed here (out of scope): filed issue #1721 for the sibling-file drift, per issue #1714's own Acceptance Criteria Map scoping this PR to has_duplicate_waiver() alone -- the same relationship issue #1432 already has to issue #1714 (distinct instances of one bug class, each its own issue).

Round 2 (verifying round 1's own fix, commit d2f0126e). The bare-CR fix was independently re-derived and confirmed closed. The new blockquote test, however, was itself found defective: it asserted only through has_duplicate_waiver(), but that function's own _WAIVER_RE line-prefix (^[ \t]*[-*]?[ \t]*) never matches a >-prefixed line at all, regardless of whether _strip_fences() correctly handles a blockquote-fenced block -- so the test passed vacuously. Live-confirmed by removing the > alternative from _CONTAINER_PREFIX entirely and observing the test still pass. Fixed: rewrote the test to assert on _strip_fences()'s own output directly (matching the pattern the bare-CR fix already established), confirmed locally to fail without the > branch present before pushing (commit 2529c6e4).

Round 3 (verifying round 2's own fix). Independently re-derived and confirmed the blockquote test now genuinely fails when the > branch is removed (a full mutation-tested run: 1 failed, 44 passed, the one failure at exactly the asserted line). Went further than a single re-check: mutation-tested every other claimed _CONTAINER_PREFIX/_WAIVER_RE branch this PR's own tests name (list bullets -/*, ordered 1., +, bare-CR \r, _WAIVER_RE's optional bullet, its \S reason-required guard) -- six separate mutants, each breaking exactly one targeted branch, with exactly the test(s) naming that branch failing and zero collateral or silent passes. No further vacuous test found anywhere in the file. Full suite: 45/45 passed.

Why CLEAN. All findings across three rounds are either fixed and independently re-verified through mutation testing (not merely re-read), or explicitly disclosed and separated into a new issue (#1721) as genuinely out of this PR's own scope per its own Acceptance Criteria Map -- no dimension silently left unaddressed, and no finding carried forward unresolved. skill-audit-disclosure (see ## Skill audit evidence below) is populated and green. pytest is green (the previously-failing property test now passes). All 45 gitapex_gate_local_preflight.py gates pass on commit 2529c6e4.

---

Generated by Claude Code