docs(knowledge): land the 9-slice publisher-profile amendment bundle

[kyle-sexton]

Closes #3014

Summary

Lands the interview-ratified amendment bundle for plugins/knowledge/skills/docpage-digest/context/anthropic-docs-profile.md from the 9-slice cloud-fleet corpus run. Knowledge plugin 0.13.3 → 0.13.4.

Ten contract rules, one PR, no verified-slice retags.

Fix

Applied the issue's ten ratified rules in the existing profile voice:

1. mixed is the closed four-surface letter rule (API request parameter, endpoint, SDK call, model ID) with worked examples (dangerouslyDisableSandbox, prUrlTemplate / skipWebFetchPreflight, apiKeyHelper).
2. Vocabulary evidence burden binds Summary / Implications / candidate text, not only claim rows.
3. code.claude.com raw-md artifact register (Documentation-Index banner, theme={null}, hard-tab expansion, \&-escaped URLs); digest layer reproduce-never-repair; reader-facing repair-and-disclose.
4. Hedge-preservation wrong-footer trap (server-managed-settings' own hedge travels instead of the hallucination footer).
5. Subsection-level inheritance satisfies row-local basis when anchor-correct and mechanically recoverable.
6. Settings-style claim = whole table row including Example cell; slice retags stay graduation-time only.
7. tag-exempt (consumer-surface) is a documented-subject test, not a hosting test.
8. Bare "See X" is navigation-pointer; a directive pointer is guidance.
9. Hard rule: no consuming-org context in digest sections.
10. Ranked doc queue recorded, not dispatched (/docs/en/permissions, /docs/en/self-hosted-environments); corpus expansion otherwise STOPPED.

Bumped plugins/knowledge/.claude-plugin/plugin.json and added a 0.13.4 CHANGELOG heading, keeping every prior heading.

Test plan

• No profile-specific unit tests exist (docpage-digest/evals/evals.json covers pipeline shape, not these contract sentences).
• scripts/check-changelog-parity.sh --check and --check-order pass; newest heading matches the bumped manifest.
• npx markdownlint-cli2 on the two markdown files: 0 issues.
• plugin.json parses as JSON; ranked queue URLs (code.claude.com/docs/en/permissions, .../self-hosted-environments) return HTTP 200.
• Prior CHANGELOG headings from 0.13.3 downward are retained.

Verification

Same evidence as the test plan. Changelog parity and markdownlint ran locally against the three-file diff before push. Slice retags named in the issue were deliberately not applied.

Related

Refs the 9-slice interview + dual-validator pass recorded in the issue body (items 2, 3, 4, 7, 8 confirmed by both validators; 1, 5, 6, 9 human-ratified). Does not take #3015.

[claude]

Claude finished @kyle-sexton's task in 3m 39s —— View job

---

Code review complete

• Gather context (PR diff, changed files, CLAUDE.md)
• Apply skip gate checks (PR open, non-draft, meaningful review surface — proceeded with full review)
• Review plugin.json version bump
• Review CHANGELOG.md entry
• Review anthropic-docs-profile.md contract-rule additions
• Post findings (inline + summary)

Summary

This PR lands a documentation/content amendment to
anthropic-docs-profile.md
(a rules corpus for the docpage-digest pipeline), bumping knowledge 0.13.3 → 0.13.4. No
executable code changed, so the review focused on internal consistency of the new/modified rules,
cross-references, and the mechanical parity between plugin.json, CHANGELOG.md, and the profile
diff.

No blocking findings. Checks performed:

• Version/changelog parity: plugin.json version (0.13.4) matches the new ## [0.13.4]
  changelog heading; prior headings (0.13.3 downward) are untouched. Changelog prose accurately
  summarizes the diff (four-surface boundary rewrite, Example-cell rule, subsection-level
  inheritance carve-out, wrong-footer trap, doc-queue STOPPED note, etc.) — I did not execute
  scripts/check-changelog-parity.sh myself (no shell access in this lane), so that part of the
  author's test plan is author-claimed and unverified by me.
• Internal consistency of the new rules: checked the reworked cc-applicable/mixed
  four-surface boundary
  (lines 175–202)
  against the pre-existing "third-party APIs count as API surfaces" rule
  (line 231):
  the prUrlTemplate/github.com example (bare hostname, not a callable API endpoint) doesn't
  contradict the no-vendor-exemption rule (which applies once a row actually names a third-party
  API call) — no conflict found.
  • The new "subsection-level inheritance" carve-out
    (lines 213–217)
    is explicitly flagged in-text as an overrule of the stricter "same basis as claim N" reading in
    the same bullet, so the tension is intentional and disclosed rather than an oversight.
  • New anchor link and cross-references (#archive-reading-conventions, the upstream-drift
    README.md fetch-route anchor) resolve correctly.
• Doc queue additions: independently fetched both newly-ranked URLs
  (code.claude.com/docs/en/permissions, .../self-hosted-environments) — both are live, real
  pages (not 404s), and the permissions page does show the Documentation-Index banner artifact
  the new "Fetch channel" bullet describes. No duplicate queue entries introduced.
• plugin.json: diff is a single-line version bump; JSON structure otherwise untouched.

Nothing here rises to a blocking or high-signal finding under this lane's bar (pre-existing content
judgment calls from the linked interview/dual-validator process are out of scope for a code-review
lane, and the specific retag/row claims — e.g. "unit 05 rows 18/37" — reference external corpus
artifacts not included in this diff, so they're not independently checkable here).

Branch: cursor/3014-publisher-profile-cfcf

[claude]

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

---

Security review — skipped (no security-relevant surface)

Task list

• Gather context (PR diff, files changed)
• Run /review:security-review per custom instructions
• Post findings (or clean bill of health)

Result

Reviewed the full diff (gh pr diff 3090) against the skip gate in /review:security-review. This PR touches only:

• plugins/knowledge/.claude-plugin/plugin.json — a version bump (0.13.3 → 0.13.4), no logic change.
• plugins/knowledge/CHANGELOG.md — prose changelog entry.
• plugins/knowledge/skills/docpage-digest/context/anthropic-docs-profile.md — markdown reference content documenting extraction/citation contract rules for a documentation-digest pipeline (tag boundaries, citation conventions, a "reproduce-never-repair" artifact register, a doc queue).

None of this introduces or modifies executable code, scripts, GitHub Actions workflows, authentication/authorization logic, injection surfaces, or secret/token handling — it's pure content and metadata. This matches skip-gate condition 2 ("change has no security-relevant surface after reading the diff"), so per the skill's instructions I'm stopping here with no findings rather than manufacturing generic commentary on prose content.

No security issues found.

[github-actions]

Last security-reviewed head: a4f8419874cb4882da54aec0b9c6863fef8437ea. 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: a4f8419874

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".

[github-actions]

Claude has reviewed this PR 2 times. The lane skips further automatic reviews after 5; deleting this comment resets the count.

[claude]

Claude finished @cursor[bot]'s task in 2m 13s —— View job

---

Code review (/review:code-review)

Tasks

• Gather context (PR diff since last-reviewed head a4f8419, CLAUDE.md, skip gate)
• Review the new commit (b41baf7) that responds to the Codex P2 hard-tab finding
• Re-check plugin.json / CHANGELOG.md parity (unchanged by this commit)
• Re-scan anthropic-docs-profile.md for internal consistency after the amendment
• Post findings

Scope

This is a pull_request synchronize re-run. The only change since the prior review (a4f8419) is commit b41baf7, which edits one paragraph in
anthropic-docs-profile.md
in direct response to the Codex P2 comment on hard-tab expansion. plugin.json and CHANGELOG.md are untouched by this commit.

Findings

No blocking findings.

• Codex P2 (hard-tab / source.* immutability) — addressed. The amended text now reads: "Formatter hooks expand hard tabs on the surfaces they are allowed to touch — never source.*, which stays the unaltered fetch... Digest text reproduces these artifacts byte-exact and never repairs them." This directly closes the gap Codex flagged (formatter-hook tab expansion silently corrupting the immutable snapshot) by carving source.* out as untouchable, consistent with SKILL.md's own immutability rule ("Every source.* file is immutable from this point forward", SKILL.md:123). No residual contradiction with the sibling artifact-register bullet under Archive-reading conventions (anthropic-docs-profile.md:94-98), which still lists "hard-tab expansion" as a standing artifact instance without needing the same caveat (it's describing the artifact class, not re-asserting who may touch source.*).
• Parity re-check (manual, not scripted — see note below): plugin.json version 0.13.4 matches the sole ## [0.13.4] changelog heading; JSON is well-formed; prior headings (0.13.3 downward) are untouched by this commit.
• Structural check: heading hierarchy in the profile file is flat and consistent (single H1 + H2s, no skipped levels); the #archive-reading-conventions anchor referenced at line 68 resolves correctly.

Verification note

This sandbox's Bash tool required interactive approval for scripts/check-changelog-parity.sh, npx markdownlint-cli2, and node -e in this run, so I could not execute them — the parity/JSON checks above were done by manual inspection instead, and markdownlint was not independently re-run. If you want those genuinely automated in this lane, --allowedTools would need to include the specific script/npx/node invocations. Per this lane's rules, the author's own test-plan claims for those commands remain author-claimed and unverified by me.

Branch: cursor/3014-publisher-profile-cfcf