feat(docs-hygiene): audit-noise gains three residue shapes (0.20.0)

[kyle-sexton]

Closes #3125

Summary

/code-tidying:audit-comment-residue detects four residue shapes in code; /docs-hygiene:audit-noise owns .md and detected five noise shapes. The two sets did not tile the space — only history-narration had a markdown counterpart (citation). plan-reference, conversational-antecedent, and ticket-pr-residue had no detector on either side of the boundary, so a README, rule body, or CLAUDE.md saying "as you asked, retry three times" or "see PR #45 for the rationale" was invisible to the whole fleet. Not a skipped file type — a gap behind an otherwise correct boundary.

The shapes went to audit-noise rather than widening the code skill to .md, and the reason is a treatment conflict, not a preference: on a markdown line the code skill's history-narration and this skill's citation fire together with opposite rulings (citation says relocate to a ## Sources footer, history-narration says delete). Two owners for one line is a precedence problem; one owner per file type is not. The boundary is now explicitly by file type, with the three shape names deliberately shared so one authoring failure keeps one name wherever it lands.

Version note: originally opened as 0.19.0. main shipped its own docs-hygiene 0.19.0 (5aa7d2ca, extract-ssot) while this PR was open, so this is renumbered to 0.20.0 and its CHANGELOG section stacked above main's.

Fix

audit-noise is now an eight-shape classifier: plan-reference and conversational-antecedent at Tier 1, ticket-pr-residue at Tier 2.

The patterns are adapted, not copied, and the adaptation is the substance of the change. The code lib classifies only the extracted comment portion of a line; this one classifies whole prose, where the same words are load-bearing far more often. Measured against this repository's own 1136-file tracked-markdown corpus, four of the sibling's cues had to go:

• per the plan — prefix-matches "per the planning chapter", and a doc citing a plan artifact that still exists is a live cross-reference, not residue
• as planned — a substring of "was planned", so "what was planned, what was done instead" self-matched
• in this change / in this session — ordinary domain vocabulary in an agent-tooling corpus

in this PR survives only with a first-person actor behind it, which separates narration ("in this PR we switch the default") from a live referent ("the files changed in this PR").

ticket-pr-residue's carve-out is restated in markdown terms rather than inherited: a task-list checklist item (- [ ] … #123) and a TODO(#123)-family marker are never flagged, because both denote outstanding tracked work — the reference is the actionable part of the line — which is what the code skill's sanctioned-TODO exception is actually about. Nothing further is carved out: the sanctioned home for a provenance citation is a ## Sources / ## History footer, and the existing section exemptions already skip those (as they skip CHANGELOG.md, fenced blocks, and frontmatter) before any shape runs. An inline parenthetical (… (tracked in #482)) stays Tier 2 on purpose, so a reviewer rules on it rather than the scanner.

Both tier tables (audit_noise_shape_tier and audit_noise_shape_tier_into) are updated — a shape added to only one silently falls through to Tier 3.

Review feedback incorporated

Three findings from the automated reviewers were addressed in 3345108b and 825ebb1b:

• Standalone as requested was missed. Added, but not as the bare cue suggested — bare would match inside "the change was requested", the same substring bug this skill's changelog records for as planned. It is boundary-guarded and matched only as a clause-final adverbial, which also spares the live attribution as requested by the client.
• The in exemption was too broad. as we discussed in yesterday's meeting was being suppressed. in now defers to a document-locator predicate — does the reference name a written locus a reader can still open (§, #anchor, section/chapter, a named durable doc, a link or path), or the conversation the sentence came from? Tracker nouns are deliberately excluded, so this shape cannot launder a line ticket-pr-residue owns. Follower comparison is now case-normalized.
• Contracted actors escaped both patterns. we've, we'll, I'm etc. now match — in both the ASCII apostrophe and the U+2019 typographic form, as literal alternatives rather than a character class (a bracket class decomposes U+2019 under a C locale; there is prior art for this in claude-config/.../instruction-scan.sh:116). The two call sites take deliberately different contraction sets: conversational-antecedent takes only 've|'d, because admitting 're would flag "do it as you're asked".

Verification

Issue probes reproduced before and after. At dff0942, probe 1 (the three orphaned shapes) gave T1=0 T2=0 T3=0 and probe 2 (control) gave T1=3, all citation — exactly as reported. Post-change:

• Probe 1 → T1=4 T2=2 T3=0: two conversational-antecedent (T1), two plan-reference (T1), two ticket-pr-residue (T2)
• Probe 2 → T1=3 T2=0 T3=0, still exactly three citation, nothing new

Corpus false-positive sweep over git ls-files '*.md' minus **/evals/fixtures/** and CHANGELOG.md (1136 files, matching the skill's own documented repo-wide defaults):

	Findings	T1	T2
Baseline (dff0942)	43	9	34
First-draft patterns	75	33	42
Shipped	55	13	42

The middle row is the point of the exercise: a close prose port of the code lib added 32 findings, 20 of them false positives. The sweep is what produced the four dropped cues above. All 12 new findings were read individually: 7 true positives, 3 self-matches on the sibling skill's own shape-definition table rows (dismissible under the already-documented "shape-definition example matching its own pattern" ground), and 2 false positives, both accepted with reasons rather than patched — see "known limitations" below.

Both later commits were held to the same bar. The in-locator change is precision-only: the corpus finding stream is byte-identical before and after, verified by a line-level differ running the old and new predicate over every line of all 1136 files (disagreements=0), not by comparing totals. The contraction change adds no findings either — a direct grep confirms the corpus currently contains zero contracted instances of either shape, so that recall gap was real but latent.

Gates, at the current head: detect.test.sh 102/102 (from 56 originally; +46), shellcheck --rcfile=.shellcheckrc clean, shfmt -d no diff, check-shell-portability.sh clean in CI-faithful mode, scripts/validate-plugins.sh pass, scripts/check-hook-exec-form.sh pass, check-changelog-parity.sh --check and --check-order pass, markdownlint-cli2 0 issues in 55 files, typos, editorconfig-checker, and check-skill.sh PASS (0 errors, 1 pre-existing warning).

Two CI failures were found and fixed during review, both in the skill's frontmatter metadata.summary: hook-exec-form-gate rejected an unquoted scalar containing ": ", and plugin-gate then rejected the quoted form because scripts/cheatsheet-config.mjs requires a plain unquoted scalar. The value is now colon-free and unquoted at 97 codepoints, checked against summaryError() directly.

Related

• Refs #3117 — the issue's ## Related: default scope section proposes generalizing a scope-resolution order across the code-tidying scanners. Deliberately not implemented here: it is a separate concern whose subjects are #3117 and audit-comment-residue, and audit-noise already has the pagination and confirm-then-run repo-wide offer that issue holds up as the model.
• Three frontmatter readers disagree, and that let a malformed value reach CI twice. skill-quality's regex reader accepted frontmatter that a strict YAML parse rejected; then the cheat-sheet guard rejected what both others accepted. No single local check catches both states. This is the sharpest follow-up candidate here and is not fixed in this PR.
• Known limitation, not filed: three of the new findings are self-matches on plugins/code-tidying/skills/audit-comment-residue/SKILL.md:38-40 — its own shape-definition rows, whose examples use plain double quotes rather than backticks, so the inline-code strip does not remove them. The one-line fix is backticking those three cells, which is outside this PR's plugin scope. This skill's own new rows use backticks precisely so they do not self-match (verified: zero findings anywhere under plugins/docs-hygiene/).
• Pre-existing defect surfaced, not caused, by this change: docs/conventions/finding-suppression/README.md:64 flags from inside a fenced block. detect.sh's fence toggle treats the inner ``` of a ````-wrapped nested fence as a close, so lines 47–65 are scanned as if outside a fence. This affects all eight shapes; fixing the fence tracker is its own change with its own regression surface.
• One accepted Tier 1 false positive: plugins/playbooks/skills/fable-5/context/communication.md:120 — "per your request" sits inside a double-quoted template of what the agent should say to the user. The only structural fix is stripping double-quoted spans the way inline-code spans are stripped, which would change strip semantics for all eight shapes and suppress real residue in quoted prose. The consumer-side remedies (backtick it, or markdown-discipline-ignore-line) already exist. This is the finding most worth a reviewer's second opinion.
• Deferred recall gap: under, at, and on carry the identical latent weakness in had — as we agreed on Tuesday, as we decided at the standup are residue that still stands the shape down. Needs its own corpus measurement (on especially is high-traffic prose) and should not ride this change.

---

Generated by Claude Code

[claude]

Claude finished @kyle-sexton's task in 1m 18s —— View job

---

Claude's todo list

• Gather context (PR diff, files changed)
• Invoke /review:security-review and follow its instructions
• Post final findings

Security review

No security-relevant surface in this diff. The change adds three regex-based prose classifiers (plan-reference, conversational-antecedent, ticket-pr-residue) to the audit-noise docs-hygiene skill's bash library (noise-shapes.sh), extends the corresponding test suite, and updates docs/changelog/eval fixtures. All new logic is pure pattern-matching over local markdown text (bash [[ =~ ]] regex, string trims) — no eval, no execution of scanned file content, no shell-injection-prone command construction, no network calls, no credential/token handling, and no .github/workflows changes in this diff.

Skip-gate condition 2 applies (change has no security-relevant surface after reading the diff) — reporting nothing further per the lane's high-signal bar.

[github-actions]

Last security-reviewed head: c6601cc4e4b9d8e4d6d62b80e528423d6c106f46. On the next push, the relevance gate compares only the commits since this SHA; delete this comment to force a full re-review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4d45c36f27

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @kyle-sexton's task in 53s —— View job

---

Tasks

• Gather context (PR diff, files changed)
• Invoke /review:security-review and follow its instructions
• Post final findings

Security review

Skip gate (condition 3) applies. The prior review on this PR already covered the full diff at 4d45c36 and found no security-relevant surface (pure regex-based prose classification in a bash library, docs, changelog, and eval fixtures — no eval, no execution of scanned content, no shell-injection-prone command construction, no network/credential handling, no .github/workflows changes).

The only change since that head is the single commit 376d98a, which touches one file:

--- a/plugins/docs-hygiene/skills/audit-noise/SKILL.md
+++ b/plugins/docs-hygiene/skills/audit-noise/SKILL.md
@@ -7,7 +7,7 @@
 metadata:
   workflow-stage: anytime
-  summary: Classify markdown noise: citations, ghost refs, meta-commentary, plan/conversational/tracker residue
+  summary: "Classify markdown noise: citations, ghost refs, meta-commentary, plan/conversational/tracker residue"

This wraps a YAML frontmatter summary: value in double quotes so it parses correctly (colons in the unquoted scalar were ambiguous to the YAML parser). It is a pure documentation/metadata fix with no security-relevant surface — no logic change, no new input handling, no workflow permission changes.

No security issues found. Nothing further to report per the lane's high-signal bar.

[claude]

Claude finished @kyle-sexton's task in 4m 19s —— View job

---

Code review (/review:code-review)

• Gather context (PR diff via gh pr diff, changed files, both commits)
• Read noise-shapes.sh, detect.sh's scan loop, SKILL.md, evals, and fixtures
• Trace the new regexes against the shipped test fixtures by hand (fence/frontmatter/section-exemption ordering, tier tables, carve-out logic)
• Post findings

Findings

One minor, non-blocking observation — posted inline on noise-shapes.sh:141: the new we|I first-person-actor checks in audit_noise_line_has_plan_reference (line 141) and audit_noise_line_has_conversational_antecedent (line 116) require a literal space right after the pronoun, so contractions ("we've", "we're", "I've") silently miss detection. Given how deliberately this PR tightened precision elsewhere, this looked worth flagging, but it may be an acceptable/deliberate recall gap rather than a defect.

No other issues found. Specifically checked and confirmed sound:

• audit_noise_shape_tier and audit_noise_shape_tier_into are updated in lockstep (both tables list the three new shapes with matching tiers) — the PR body correctly calls out that a shape added to only one silently falls to Tier 3, and that isn't the case here.
• The ticket-pr-residue tracked-work carve-out (audit_noise_line_is_tracked_work, checked before the bare-provenance patterns) matches the documented behavior for task-list items and TODO(#123)-family markers, including the case where a carved-out line also contains a bare reference elsewhere on the same line.
• Section-exemption / frontmatter / fence handling in detect.sh runs before any shape check, so the three new shapes correctly inherit the existing ## Sources/## History/CHANGELOG.md/fence/frontmatter exemptions without needing shape-specific carve-outs.
• The dropped code-sibling cues (per the plan, as planned, in this change/in this session) and the tightened in this PR / as we discussed conditions all check out against the new detect.test.sh fixtures (prose-residue, carve-out, exemption, and negative cases) by manual trace.
• Version bump (0.18.3 → 0.19.0) and the second commit's frontmatter-quoting fix are both correctly reflected in the current SKILL.md.

Not independently re-run (this lane's Bash access is scoped to gh pr diff/view/comment/review, Read, Glob, Grep): detect.test.sh, shellcheck, shfmt, markdownlint-cli2, typos, editorconfig-checker, check-changelog-parity.sh, check-skill.sh. The 79/79, gate-clean, and corpus-sweep numbers in the PR description are author-claimed and unverified by me.

[github-actions]

Claude has reviewed this PR 4 times. The lane skips further automatic reviews after 5; deleting this comment resets the count.

[claude]

Claude finished @kyle-sexton's task in 1m 17s —— View job

---

Tasks

• Gather context (PR diff, files changed)
• Invoke /review:security-review and follow its instructions
• Post final findings

Security review

Skip gate (condition 2/3) applies — no security-relevant surface, including in the commits since the last review.

The prior two reviews on this PR covered the diff through 376d98a (the frontmatter-quoting fix) and found no security-relevant surface. Two more commits have landed since then, both addressing review feedback on the same regex-based classifiers:

• 3345108 — tightens the conversational-antecedent shape's in exemption to a document-locator check, and adds a bounded as requested clause-final match
• 825ebb1 — admits contracted first-person pronouns (we've, we'd, I'm, …) in plan-reference and conversational-antecedent, matching both the ASCII apostrophe and the U+2019 typographic form as literal alternatives (not a bracket class, to stay correct under a C locale)

I read the full diff of noise-shapes.sh since 376d98a. Both commits stay entirely within the same pattern: bash [[ =~ ]] regex matching and string trims over an already-locally-read markdown line. No eval, no execution of scanned file content, no command construction from untrusted input, no network or credential handling, and no changes anywhere in this PR's diff under .github/workflows. The rest of the changed-file set (CHANGELOG.md, SKILL.md, detect.test.sh, eval fixtures) is docs/test-fixture content with no logic of its own.

No security issues found. Nothing further to report per the lane's high-signal bar.

[claude]

Claude finished @kyle-sexton's task in 4m 9s —— View job

---

Tasks

• Gather context (PR diff via gh pr diff, changed files, all 5 commits)
• Invoke /review:code-review and follow its instructions
• Post final findings

Code review (/review:code-review)

Skip-gate check: this head (a24899d) has new commits since the last /review:code-review pass (which reviewed 4d45c36) — 3345108 and 825ebb1 land the fixes for both findings raised by that pass and by Codex's review, so condition 4 (an applicable prior review on this exact head) doesn't hold. Did a fresh pass over the full diff.

Scope: noise-shapes.sh's three new shape functions (audit_noise_line_has_conversational_antecedent, audit_noise_line_has_plan_reference, audit_noise_line_has_ticket_pr_residue, audit_noise_follower_is_document_locator, audit_noise_line_is_tracked_work), their wiring into audit_noise_detect_shapes_into and both tier tables, detect.test.sh's new fixtures (11–11f), SKILL.md, CHANGELOG.md, evals, and the two eval fixture files.

Method: manually traced each regex against its own doc comment and against the detect.test.sh fixture lines it's meant to classify (capture-group indices, the rest/follower handoff between audit_noise_line_has_conversational_antecedent and audit_noise_follower_is_document_locator, the contraction alternations for both straight and curly apostrophes, the tracked-work carve-out ordering in audit_noise_line_has_ticket_pr_residue, and the two tier-table case statements) rather than executing the suite (this lane's Bash access is gh pr diff/view/comment/review, Read, Glob, Grep only — detect.test.sh, shellcheck, etc. were not re-run; the 79/79 and gate-clean numbers in the PR description are author-claimed and unverified by me).

No new findings. Both prior review threads on this PR — Codex's standalone-as requested and blanket-in-exemption findings, and this lane's contracted-actor gap — are fixed in 3345108/825ebb1 and check out against the new fixtures on manual trace:

• The clause-final-adverbial bound on As requested, … (noise-shapes.sh:578) correctly admits As requested, retry… while excluding was requested (boundary check on [Aa]s) and as requested by the client (no trailing clause break).
• audit_noise_follower_is_document_locator (noise-shapes.sh:531) correctly distinguishes as we decided in the ADR (exempted) from as we decided in favor of streaming / as we discussed in yesterday's meeting (still flagged) — traced the rest handoff through both capture-group reassignments.
• The 've/'d-only contraction set on conversational-antecedent vs. the all-five-form set on plan-reference matches the stated rationale (past-participle follower vs. free tense) and the as you're asked negative case correctly falls through (no whitespace immediately after you once the empty optional-group alternative is tried).
• audit_noise_line_is_tracked_work is checked before the bare-provenance patterns in audit_noise_line_has_ticket_pr_residue, and both tier tables (audit_noise_shape_tier, audit_noise_shape_tier_into) list the three new shapes in lockstep.
• Both new shape functions run against the inline-code-stripped line ($stripped, not $unwrapped), consistent with the doc comment's self-match-avoidance claim.

Nothing else stood out as something a careful reviewer would block on — the remaining edges the author already called out as accepted trade-offs (the fable-5 quoted-template false positive, the pre-existing fence-tracking defect, the three code-tidying self-matches) are documented limitations, not regressions in this diff.

[kyle-sexton]

No description provided.

[claude]

Claude finished @kyle-sexton's task in 1m 12s —— View job

---

Tasks

• Gather context (PR diff via git diff origin/main...HEAD, commit history since last review)
• Invoke /review:code-review and follow its instructions
• Post final findings

Code review (/review:code-review)

Skip gate (condition 4) applies — this head's actual diff has already been fully reviewed.

Current head is 7be4bac21a76ec24f12c80b7df27a9560a4a5852. The prior /review:code-review pass (06:08:15Z run) covered head a24899d — the same two shape-fix commits (3345108, 825ebb1) that are still HEAD's tip of this PR's own work — and found no issues.

The only commit between a24899d and 7be4bac is 1aff27e (#3153, an unrelated guardrails fix landed on main), pulled in by the merge commit 7be4bac. Confirmed this adds no new PR surface:

• git diff origin/main...HEAD --stat at 7be4bac shows exactly the same 10 files and the same 480/-18 line count reported in the PR description — no drift from the merge.
• git diff a24899d 7be4bac -- plugins/docs-hygiene/ shows real content (a detect.sh porcelain-parsing fix, CHANGELOG.md 0.19.1 entry), but that content originates from main itself (#3143), not from this branch — it doesn't appear in the origin/main...HEAD diff and isn't in this PR's own changed-file list.

So the reviewable diff is unchanged since the last full pass, and that pass's "no new findings" verdict still applies. Per the skip-gate instruction, posting nothing further.

[claude]

Claude finished @kyle-sexton's task in 1m 6s —— View job

---

Tasks

• Gather context (PR diff, files changed)
• Invoke /review:security-review and follow its instructions
• Post final findings

Security review

Skip gate (condition 3) applies — this head's PR diff is unchanged from the last successful security review.

Current head is 7be4bac21a76ec24f12c80b7df27a9560a4a5852. The prior security review pass covered a24899d (through commits 3345108/825ebb1) and found no security-relevant surface. The only commits added since a24899d are 1aff27e (fix(guardrails): land the unlanded trust-gate scan fixes from preserve/stash-0, #3153) and the merge commit 7be4bac itself — both are origin/main history merged into this branch, not new PR commits.

Confirmed via git diff origin/main...HEAD --stat: the PR's actual changeset is still exactly the same 10 files (docs/SKILL-CHEAT-SHEET.md, plugins/docs-hygiene/.claude-plugin/plugin.json, plugins/docs-hygiene/CHANGELOG.md, plugins/docs-hygiene/README.md, plugins/docs-hygiene/skills/audit-noise/SKILL.md, evals.json, two eval fixtures, detect.test.sh, noise-shapes.sh) with the same +480/-18 totals as when last reviewed. 1aff27e touches only plugins/claude-config/plugins/testing/plugins/work-items guardrail files, none of which overlap this PR's scope.

As established in prior passes: this diff is pure bash [[ =~ ]] regex pattern-matching over locally-read markdown text (three new residue-detection shapes in noise-shapes.sh), plus docs/changelog/eval-fixture content. No eval, no execution of scanned file content, no shell-injection-prone command construction, no network or credential handling, and no .github/workflows changes anywhere in this PR's diff.

No security issues found. Nothing further to report per the lane's high-signal bar.

[kyle-sexton]

🔒 Lane claim + escalation (melo-lap-001-merge-20260823T045045Z, autopilot, PR-queue coordinator).

Fresh merge-gate re-check just now shows this PR has moved from BLOCKED (unresolved threads only) to mergeStateStatus=DIRTY (mergeable=CONFLICTING) — churn from #3181/#3178 merging into main just now, touching an overlapping file (plugins/docs-hygiene/skills/audit-noise/scripts/lib/noise-shapes.sh, same file this PR's unresolved threads reference). This is now a genuine merge-conflict PR and out of scope for this pass (a separate pass handles conflicting PRs).

Independent read of the 3 unresolved bot threads: all three are genuinely addressed — kyle-sexton (human) replied to each with the exact fix commit (3345108b, 3345108b, 825ebb1b), all reachable in this branch's history. However this PR is self-authored (kyle-sexton), so --autonomous/--independent-resolver thread resolution refuses by design (self-login guard, no override) — this is the same self-login block flagged in the prior cycle and it is still the blocker; it is not stale.

Leaving for human: resolve the (already-addressed) threads and merge the branch to pick up the current main, or route through the conflict-handling pass.

[claude]

Claude finished @kyle-sexton's task in 1m 17s —— View job

---

Tasks

• Gather context (PR diff via git diff origin/main...HEAD, commit history since last review)
• Invoke /review:security-review and follow its instructions
• Post final findings

Security review

Skip gate (condition 3) applies — this head's PR diff is unchanged from the last successful security review.

Current head is c6601cc4e4b9d8e4d6d62b80e528423d6c106f46. The prior security review pass covered 7be4bac and found no security-relevant surface. The only thing added since then is one merge commit (c6601cc) pulling in 6fcc723 (refactor(scripts): define the list-file format once, and close a third twin fail-open, #3181) from origin/main — not a new commit on this PR's own branch.

Confirmed two ways:

• git diff origin/main...HEAD --stat at c6601cc shows the identical 10-file, +480/-18 changeset previously reviewed — no drift.
• git diff 7be4bac c6601cc -- plugins/docs-hygiene/skills/audit-noise/scripts/lib/noise-shapes.sh plugins/docs-hygiene/skills/audit-noise/scripts/detect.test.sh is empty — the actual regex-matching logic is byte-identical to what was already reviewed.

The only textual movement visible between 7be4bac and c6601cc (a plugin.json em-dash unescape and a new CHANGELOG.md [0.19.2] entry) originates from origin/main itself, not from this branch, and cancels out of the three-dot diff — consistent with 6fcc723 touching only unrelated scripts/ list-parsing files (scripts/lib/read-list.sh and friends), none of which overlap this PR's scope.

As established in every prior pass on this PR: the reviewable diff is pure bash [[ =~ ]] regex pattern-matching over locally-read markdown text (three new residue-detection shapes in noise-shapes.sh), plus docs/changelog/eval-fixture content. No eval, no execution of scanned file content, no shell-injection-prone command construction, no network or credential handling, and no .github/workflows changes anywhere in this PR's diff.

No security issues found. Nothing further to report per the lane's high-signal bar.

[kyle-sexton]

Both blockers from the merge-gate pass are cleared.

Conflict resolved — head is now c6601cc4, a merge of main at 6fcc723a.

One correction to the lane's diagnosis: noise-shapes.sh was not the conflicting file. It auto-merged cleanly. The actual conflicts were plugins/docs-hygiene/.claude-plugin/plugin.json and plugins/docs-hygiene/CHANGELOG.md — the same version/changelog collision this PR has hit on each main advance, this time from #3178's SSOT normalization batch bumping docs-hygiene to 0.19.2.

Resolution: kept our 0.20.0 (still above main's 0.19.2, so no collision), took main's normalized description from #3178 rather than ours — that batch un-escaped — to a literal em dash, and this PR should not revert it — and stacked the changelog 0.20.0 → 0.19.2 → 0.19.1 → 0.19.0.

Re-verified after the merge: validate-plugins.sh, check-changelog-parity.sh --check and --check-order, check-hook-exec-form.sh, detect.test.sh 107/107, check-skill.sh, markdownlint-cli2 (0 issues, 55 files), and all four regression probes unchanged. git diff origin/main...HEAD is still exactly the same 10 files at +480/-18, so the merge introduced no drift.

Threads resolved. The three bot threads are now marked resolved. The lane's read was correct — all three were genuinely addressed, each with a reply naming its fix commit (3345108b, 3345108b, 825ebb1b), and I verified the fixes behaviorally rather than by reference alone before replying. The self-login guard that blocked --autonomous resolution is doing its job for an independent resolver; these are resolved here by the lane that authored the fixes, which is where that responsibility sits.

No human action needed on either count. The PR should re-evaluate as mergeable once CI completes on c6601cc4.

---

Generated by Claude Code

[claude]

Claude finished @kyle-sexton's task in 1m 43s —— View job

---

Tasks

• Gather context (PR diff via git diff origin/main...HEAD, commit history since last review)
• Invoke /review:code-review and follow its instructions
• Post final findings

Code review (/review:code-review)

Skip gate (condition 4) applies — this head's actual diff has already been fully reviewed.

Current head is c6601cc4e4b9d8e4d6d62b80e528423d6c106f46. Confirmed git diff origin/main...HEAD --stat is still exactly the same 10 files with the same +480/-18 totals reported in the PR description, and traced what changed since the last /review:code-review pass on this PR (7be4bac, 08:36:38Z run, which itself carried forward the no-new-findings verdict from a24899d's full pass):

• git diff 7be4bac c6601cc is empty for 8 of the 10 PR files, including noise-shapes.sh and detect.test.sh — byte-identical to what was already reviewed.
• The only two files with any diff between those heads are plugin.json and CHANGELOG.md, and in both cases the change is main's own churn arriving via the merge commit, not new work on this branch: plugin.json's only change in this PR's actual diff (origin/main...HEAD) is the version bump 0.19.2 → 0.20.0; CHANGELOG.md's PR diff is still exactly the [0.20.0] section, now stacked above main's own newly-landed [0.19.2] entry rather than [0.19.1].

So the reviewable content is unchanged since the last full pass, and that pass's "no new findings" verdict still applies. Per the skip-gate instruction, posting nothing further beyond this confirmation.
· branch claude/work-items-integration-cuy22l