fix(ai-slop): fix ten audit-run defects, re-enable rule-em-dash, and purge the corpus

[kyle-sexton]

Closes #3987

Summary

Two pieces of work land together here, because the second is what the first
asked for.

The audit-run defects (#3987). A repo-wide /ai-slop:audit run surfaced ten
defects in the ai-slop plugin and a contradiction in this repository's own
config, which had disabled the em-dash rule as "house style". Em dashes are not
this repository's house style, so the rule is re-enabled and the plugin's own
tree becomes the first surface that passes the rule it ships. Decisions were
reached in an interview with the maintainer and recorded in
docs/topics/ai-slop-plugin-scope/PLAN.md.

The corpus purge (#2891). Re-enabling the rule turned 25,558 findings across
1,003 files from a policy question into a backlog. The maintainer's call was that
the whole purge lands in this pull request rather than one per plugin, so it does.
The campaign runs surface by surface: each plugin's markdown is rewritten, its
manifest description cleaned, its released changelog entries corrected in place
and named in a new release entry, and its tree declared in
scripts/em-dash-purged-paths.txt so the gate defends it from that moment on.

Every rewrite follows plugins/ai-slop/skills/audit/reference/rewrite-guide.md:
an em dash becomes a comma, a period, a colon where a definition or list follows,
or a restructured sentence. Never parentheses, never an en dash, never a spaced
hyphen, never a semicolon. Verbatim third-party quotations keep their dashes
inside an ignore marker naming what is quoted; the repository's own prose is
never fenced to silence a finding.

Fix

Plugin (ai-slop 0.5.13 to 0.6.0)

• detect.sh: the quotation exemption carries an open double-quoted span across a
  soft line break (the plugin's own CHANGELOG produced two false
  rule-challenges-conclusion findings without this); the bare noun underscore
  leaves the default vocabulary; every Summary row splits declined= into
  declined_marker, declined_quote, declined_config.
• emit-findings.sh: --from is repeatable and per-rule counts are summed across
  chunks (the old script kept the last chunk's counts and called a rule "no
  result" if any chunk had zero); ## Surfaces now states files scanned, chunk
  count, whole files declined, and every config-disabled rule.
• SKILL.md: the pre-computed config block no longer truncates disabled_rules
  and rule_allowed_paths (head -8 cut them off); the rubric pass covers every
  file in scope and fans out per the new context/rubric-fanout.md with resumable
  per-batch result files.
• persist-findings.md: the guard file is created with the Write tool (the
  guardrails hook blocks a shell redirect); the Surfaces description matches the
  script.
• README: names /ai-slop:audit as the invocation; a bare /ai-slop is not a
  command.
• The plugin's own prose passes its own audit, and whole-file ignore markers on
  catalog.md and rewrite-guide.md are removed: the quotation exemption already
  covers their examples.

Repository config and gates

• .claude/ai-slop.json: rule-em-dash leaves disabled_rules; catalog.md
  leaves excluded_paths (46 of its 47 em-dash lines were its own prose); the
  comment states the current rule and its reason. Curly-quote and emoji rules stay
  disabled on their owner rulings; vendor trees and eval fixtures stay excluded,
  and those two exclusions are the only ones kept.
• scripts/em-dash-purged-paths.txt grows with the campaign, from 103 declared
  paths at the first commit to 383 covering 1,194 files. A glob
  matching nothing fails the gate, so a declaration cannot outlive the surface it
  names.
• .claude/rules/vendor-docs-are-not-style.md: vendored upstream docs are
  reference material, not a style precedent for this repository's own instruction
  surfaces.

Corpus surfaces purged, and what is left

The campaign is complete. All 77 plugins are declared, alongside the convention
documents under docs/conventions/, the specifications under docs/specs/, the
root-level instruction surfaces, and every plugin manifest in the marketplace.
docs/CATALOG.md and docs/SKILL-CHEAT-SHEET.md both report zero em-dash
findings, which is the proof that no manifest description and no skill summary
still carries one.

A repo-wide detector run over every tracked markdown file now reports 817
findings in 36 files, down from 25,558 in 1,003. Every one of the 36 sits in an
exclusion this repository recorded before the campaign began, and nothing is
unaccounted for:

• 33 files (724 findings) in docs/adr/** and docs/upstream/**, which
  scripts/em-dash-purged-paths.txt has declared out of scope since before this
  branch: decision records are a historical account rather than prose to
  restyle, and the upstream tree is third-party text under the repository's own
  vendor-docs-are-not-style rule.
• 3 files (93 findings) holding the standards-contract: the convention and its
  two byte-identical plugin copies, explained below.

Generated artifacts and the files that feed them

A file another artifact is generated from, or that a gate parses, is not just
prose. Four such surfaces were fixed at the source rather than in the output:

• scripts/generate-catalog.mjs (the row separator) and docs/CATALOG.md
  regenerated.
• plugins/claude-ops/skills/audit-native-overlap/scripts/overlap.py (the view
  header, the empty-verdict placeholder, and three row templates),
  docs/native-surfaces/records.json, and docs/NATIVE-SURFACES.md regenerated.
  test_overlap.py asserted the old header wording and was updated in the same
  pass.
• Twelve plugin manifests' userConfig descriptions, with each plugin README's
  "Options reference" table regenerated by
  scripts/sync-plugin-options-docs.py.
• The extract-ssot citation contract prescribed an em dash as its citation
  separator. It now prescribes a colon, so the template no longer emits the
  character this repository has purged.

docs/conventions/standards/README.md is deliberately NOT purged and is excluded
by name: it carries a standards-contract semver, and changing one character of
it forces a version bump that makes every consuming repository's index mismatch,
degrading skills to best-effort routing until each one re-runs setup. Thirty-nine
lines of punctuation are not worth that migration.

Anchors

Renaming a heading that carried an em dash changes its anchor. Every cross-file
anchor in the repository resolves: check-anchors.py reports 0 broken across all
1,529 tracked markdown files, including three links to the upstream-drift
convention's fetch-route heading that lost its doubled separator.

Verification

• plugins/ai-slop/skills/audit/scripts/detect.test.sh: all 216 cases pass,
  including new cases for the wrapped-quote exemption, the noun underscore, the
  split declined counts, and multi-chunk aggregation.
• scripts/check-purged-em-dashes.sh: 383 declared paths, 1,194 files scanned, no
  em dashes.
• scripts/check-changelog-parity.sh in --check, --check-bump origin/main,
  and --check-preserved origin/main modes: clean. All 88 changed changelogs
  preserve every version heading they carried at the base ref (3,938 headings
  compared).
• node scripts/generate-catalog.mjs --check, node scripts/generate-cheatsheet.mjs --check, overlap.py generate --check, and
  python3 scripts/sync-plugin-options-docs.py --check: all in sync.
• scripts/affected-tests.sh --run on each shard of the work, and twice over the
  full selection after the second origin/main merge. Four suites fail in this
  container and none of them belongs to this branch: every hook, test, and harness
  file involved is byte-identical to origin/main. Two fail on every run,
  test_save_point.py::test_new_origin_falls_back_to_directory_name and
  cache-content-check.test.sh's process-budget probe, which measures -1
  because the pid-stamped PS4 does not reach the traced shell in this sandbox.
  Two more rotate between runs, which is what marks them as container artifacts
  rather than defects. block-noncanonical-commit.test.sh reported 43 failures on
  the first run and none on the second; every one was a blocked case returning the
  allow status while every allowed case passed, the signature of the guard's
  documented fail-open posture, and the suite passes standalone at 239 of 239 over
  four separate runs. session-event-log.test.sh reported 33 parallel fires
  writing 32 lines on the second run and none on the first, a lost-write race in
  the shared empty-guard heal that this branch does not touch. All four pass in
  CI: lint, all four test-linux shards, test-windows, hook-utils, and
  ci-status are green on the merge head.
• markdownlint-cli2 clean on every edited file; shellcheck, shfmt, and typos
  clean on the edited scripts.
• Each area was checked by a fresh-context semantic-diff verifier blind to the
  rewrite rationale, and the flags it raised were fixed before the area landed.

Related

• Closes ai-slop: 10 audit-run findings (config truncation, chunk aggregation, wrapped-quote exemption, rubric fan-out) and this repo's em-dash posture #3987.
• Refs De-slop campaign: purge em dashes and AI tells from this repo's own corpus via /ai-slop:audit fix #2891: the corpus campaign, which this PR carries rather than defers.

🤖 Generated with Claude Code

https://claude.ai/code/session_011o8Akf3vwDEHy5ATnB1RwF

---

Generated by Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-08T23:07:01.370117Z	a7962e6	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a7962e6c96

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".