fix(docs-hygiene): correct audit-noise's stated negation limitation (0.21.2)

[kyle-sexton]

No linked issue

Deliberately not Closes #3195. This corrects what the skill body says; #3195 remains open for the paragraph-accumulation fix that changes what the detector does.

Summary

0.21.1 (#3202) added a gate requiring a negation candidate's line to close its own sentence. That silently changed what happens to a soft-wrapped prohibition, and the "known limitation" bullet in SKILL.md was left describing 0.21.0 — so main currently ships a skill body making two false claims about its own detector.

Measured against the shipped detector at 7dea5a84:

Do not use markdown;
compose flowing prose instead.

Do not use markdown
in the summary body.

Never emit a bare summary.

Line	SKILL.md claims	Actual
Do not use markdown; (positive paired on next line)	"is reported"	not reported — ; never terminates a sentence, so the line reaches no verdict
Do not use markdown (no positive anywhere)	—	not reported
Never emit a bare summary.	flags	flags

1. "is reported" is false. The false positive the bullet documents no longer occurs.
2. "The error direction is a false positive, never a silent withhold" is inverted. A hard-wrapped prohibition with no positive alternative anywhere in its sentence is now missed entirely.

Fix

Rewrite the "known limitation" bullet in plugins/docs-hygiene/skills/audit-noise/SKILL.md to describe the behavior that actually ships, and bump to 0.21.2 with a CHANGELOG entry. Documentation only — no detector, emitter, or test behavior changes.

The rewritten bullet does three things the old one did not:

• States the real mechanism. The gate requires a line to close its own sentence, so a soft-wrapped prohibition reaches no verdict at all — rather than the old claim that it is reported in pieces.
• Names the error direction correctly, as a silent withhold. This is the load-bearing half. A silent withhold is the one failure mode admission test 2 in docs/conventions/detector-findings/README.md asks this rule set to fail away from, and in a repo that hard-wraps its prose — this one — it takes every prohibition long enough to wrap. The old text told a reader the opposite: that the gap cost reviewer attention rather than coverage, which is exactly the reassurance that would stop someone from prioritizing it.
• Marks it as the shape's one departure from fail-safe-toward-emitting, and points at fix(docs-hygiene): audit-noise's negation shape misreads sentences that markdown soft-wraps #3195, which now carries the revised acceptance criteria.

A second correction, recorded rather than silently fixed

While re-measuring I found the merged #3202's body table says the narrowing took the negation count 1053 → 31. That is wrong: 31 was the intermediate line-gate build, superseded when the per-sentence gate landed. The shipped figure is 69, which is what CHANGELOG.md 0.21.1 and the corpus re-measure both say.

Re-run on the identical 85-file corpus (plugins/docs-hygiene + plugins/source-control tracked markdown, minus evals/fixtures and CHANGELOG.md) at 7dea5a84:

Shape	Count
negation	69
ghost-ref	3
ticket-pr-residue	2
citation	2
scope-meta	1
plan-reference	1
enum-list	1

The other shapes are byte-identical to the claim (3/2/2/1/1/1), so the narrowing itself behaved as described — only the headline number in that PR body was stale. Issue #3201's body carries the same 31. Both are merged/closed records; the correction is noted here and on #3195 rather than by rewriting them.

Verification

Gate	Result
detect.test.sh	All 133 checks passed
emit-findings.test.sh	All 35 checks passed
check-skill.sh (source checkout)	PASS — 0 errors; all 8 base-ref trigger phrases preserved
check-changelog-parity.sh --check / --check-bump origin/main / --check-order	PASS
check-detector-findings-crosswalk.sh	PASS
markdownlint	0 issues
typos	clean
corpus re-measure, 85 files	negation 69; other shapes 3/2/2/1/1/1

The reproduction above was run against the merged detector, not a local build, so the "Actual" column reflects what main ships today.

Related

• fix(docs-hygiene): narrow audit-noise's negation shape to imperatives (0.21.1) #3202 — shipped the 0.21.1 gate whose side effect this documents; its body's 1053 → 31 is corrected above.
• fix(docs-hygiene): audit-noise's negation shape misreads sentences that markdown soft-wraps #3195 — the deferred paragraph-accumulation fix; its acceptance criteria and error direction were revised in light of this measurement.
• fix(docs-hygiene): audit-noise's negation shape is unusably broad as shipped in 0.21.0 #3201 — closed by fix(docs-hygiene): narrow audit-noise's negation shape to imperatives (0.21.1) #3202; its body carries the same superseded 31.
• feat(docs-hygiene): negation-without-positive detector on audit-noise, wired to the findings relay (0.21.0) #3194 — introduced the shape and the original limitation text.

[claude]

Claude finished @kyle-sexton's task in 42s —— View job

---

Security Review

• Gather context (PR diff, files changed)
• Run /review:security-review procedure
• Analyze changed files for security implications
• Post findings

No security issues found.

This PR is documentation-only: a plugin version bump (0.21.1 → 0.21.2 in plugin.json), a CHANGELOG entry, and a corrected prose bullet in SKILL.md describing a known limitation of the audit-noise detector's negation shape.

None of the changed content is executable — no detector script (detect.sh), emitter, test, or GitHub Actions workflow logic is touched, and no config/permissions were widened. There is no injection surface, no credential handling, and no trust-boundary change here; this is a pure fix to what the skill's markdown documentation claims about existing (unchanged) detector behavior. Skip-gate condition 2 applies (no security-relevant surface after reading the diff).

[github-actions]

Last security-reviewed head: 597e1efd8287dfbab7bc3d74a9710beab6605766. On the next push, the relevance gate compares only the commits since this SHA; delete this comment to force a full re-review.

[claude]

Claude finished @kyle-sexton's task in 30s —— View job

---

/review:code-review — skipped

• Gather context (pre-fetched PR data + plugins/review/skills/code-review/SKILL.md)
• Apply skip gate

Result: This PR is already merged (PR State: MERGED). The skill's skip gate stops before any deep review when "PR is closed or not open," so no inline comments were posted and no review findings were generated.

If a fresh review is wanted, it would need to run against an open PR or another branch/target — let me know and I can do that instead.

[github-actions]

Claude has reviewed this PR 1 time. The lane skips further automatic reviews after 5; deleting this comment resets the count.