docs(docs-hygiene): unify audit-derivability fork-mechanism wording

[kyle-sexton]

Closes #1082

Summary

The audit-derivability skill warns against confirming a deletion with a contaminated context, and
names the fork mechanism at three sites in SKILL.md. Those three sites had drifted apart: two gave
the same fact in two different voices, and the third named none of it.

• Hard Rules bullet (was line 111): is unrelated → is the opposite. The spot-test bullet
  (line 56), context/rubric.md:134, and evals/evals.json:84 already used "the opposite," so this
  bullet was the sole outlier of four sites — aligning it required no change to the other three.
• Gotchas self-grade bullet (was line ~119): previously the one bare non-fork subagent
  mention. It now names the Agent tool's fork subagent type as the forbidden mechanism, which
  evals/evals.json:88 already grades on. It does not repeat the context: fork explanation —
  a Gotchas summary restating the full disambiguation a third time would be duplication on an
  agent-facing surface, exactly what this plugin exists to remove.

"Opposite" is also the more accurate word, verified against current official docs this session:

• Agent-tool fork — "A fork inherits the full parent conversation instead of starting fresh"
  (tools reference)
• Skill context: fork — "It won't have access to your conversation history"
  (skills)

Same word, inverse context behavior. "Unrelated" understated precisely the trap the disambiguation
exists to prevent.

docs-hygiene bumped 0.8.6 → 0.8.7 with the matching CHANGELOG entry.

Test plan

Run from the PR worktree against origin/main:

• scripts/check-changed-skills.sh origin/main — PASS, 0 errors, 0 warnings; confirms all 8
  base-ref trigger phrases preserved (the frontmatter description is untouched) and SKILL.md at
  142/500 lines.
• scripts/check-skill-portability.sh origin/main — no unexcused coupling tokens.
• scripts/check-changelog-parity.sh --check-bump origin/main — the version bump has its
  ## [0.8.7] entry.
• scripts/validate-plugins.sh — all plugin manifests and the catalog validate.
• markdownlint-cli2 --config .markdownlint-cli2.jsonc "plugins/docs-hygiene/**/*.md" — 36 files,
  0 errors.
• scripts/check-silent-skips.sh, scripts/check-skill-leaf-names.sh — clean.

No behavioral surface changes: prose only, plus the version/CHANGELOG pair.

Related

• docs(docs-hygiene): audit-derivability fork-mechanism wording drift (line 56 vs 111) + Gotchas line ~119 still undisambiguated #1082 — the item this closes; a review-deferral from PR docs(docs-hygiene): clarify audit-derivability Hard Rules line 111 fork-mechanism reference #1067, itself a follow-up to docs(docs-hygiene): audit-derivability SKILL.md Hard Rules line 111 'never by a fork' reads ambiguous without line 56's context #1062 /
  docs(docs-hygiene): fix inverted fork-mechanism attribution in audit-derivability rubric+evals #1058 / docs: audit-derivability rubric/evals invert the context-fork inheritance claim #1053, which is the lineage that produced the two-voice drift.
• docs(plugin-audit-port): "context: fork inherits degraded history" rationale is false — a forked skill has no history #1268 — touches a related "fork inherits degraded history" claim in a different plugin.
  Deliberately out of scope here; this PR stays inside
  plugins/docs-hygiene/skills/audit-derivability/.
• docs(docs-hygiene): audit-derivability asserts "context: fork is the opposite" as settled while #1258 contests it (4 sites) #1317 — deferred review finding from this PR (Codex, P2), classified VALID (defer). sweep-all-disciplines: fork subagents silently do not inherit the conversation, and two of four claimed they did #1258 reports
  the Agent tool's fork subagent type NOT inheriting the conversation at runtime, against its
  documentation, which would make "the opposite" an overclaim at all four sites. Deferred because a
  partial hedge reintroduces exactly the drift docs(docs-hygiene): audit-derivability fork-mechanism wording drift (line 56 vs 111) + Gotchas line ~119 still undisambiguated #1082 removes, one of the four sites is a graded
  evals.json criterion rather than prose, and sweep-all-disciplines: fork subagents silently do not inherit the conversation, and two of four claimed they did #1258 is itself unreduced and needs-human. The
  Hard Rule is unaffected — "never fork for the spot-test" holds under either branch.
• context/rubric.md and evals/evals.json were read as the tie-breaker for which phrasing is
  canonical, and are intentionally not modified — they already carry the winning wording.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NE8W4XPPVfZVCGBW2GYgyH

[claude]

Claude encountered an error after 1s —— View job

---

I'll analyze this and get back to you.

[claude]

Claude encountered an error after 1s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Warning

Automated review did not complete — this is an infrastructure failure, not a review verdict.

Treat any Claude comment on this PR (including a placeholder like "I'll analyze this and get back to you") as incomplete, not "no findings."

• Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/30160800194
• Last SDK result: {"subtype":"success","is_error":true,"num_turns":1,"duration_ms":582,"total_cost_usd":0}

Re-running the job, or pushing a new commit, will retry the review.

[github-actions]

Warning

Automated security review did not complete — this is an infrastructure failure, not a review verdict.

Treat any Claude comment on this PR (including a placeholder like "I'll analyze this and get back to you") as incomplete, not "no findings."

• Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/30160800163
• Last SDK result: {"subtype":"success","is_error":true,"num_turns":1,"duration_ms":350,"total_cost_usd":0}

Re-running the job, or pushing a new commit, will retry the review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bc2b43fede

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 encountered an error after 1s —— View job

---

I'll analyze this and get back to you.

[claude]

Claude encountered an error after 0s —— View job

---

I'll analyze this and get back to you.

[kyle-sexton]

🤖 Autonomous lane note (AI-generated) — read before merging.

CI is green (26/26 SUCCESS, MERGEABLE), but the AI review verdict on this PR is absent, not clean.

claude-review and claude-security-review both failed inside the SDK (is_error: true, 1s and 0s), and the repo's own infra-status gate posted its warning: "Treat any Claude comment on this PR ... as incomplete, not 'no findings.'" I re-ran both workflows (runs 30160800194, 30160800163); both failed identically on the retry, so this is not a transient blip.

It is not branch-owned. 4 of the 5 most recent open PRs (#1323, #1322, #1320, #1319) carry the same warning, so this is an environment/credential-level failure across the repo, not something this diff caused or can fix. Filed separately rather than papered over here.

Note that the surrounding job reports review / review pass — the workflow succeeds while the review inside it errors, so a green check row is not evidence a review ran. Worth knowing for any lane that gates on check status alone.

What did review this PR: Codex posted one P2 finding, which was validated against the actual sources and classified VALID (defer) → #1317, with a threaded reply and the reasoning recorded there. So this PR is not unreviewed — but it has had no Claude-side review, and the merger should weigh that.