Prune /setup and /submit-for-review; fix permission-audit cd false positive

[sebastientaggart]

Follow-on to #206, extending the prune/keep test to the two largest remaining skills and fixing a permission-audit false positive surfaced during the audit. As anticipated in the ticket, the reduction here is modest — these files are mostly gates and policy (keep-side), not output templates.

The defect (fixed)

permissions.yaml listed cd in a single flat commands: list, but sync.py deliberately excluded it from generated allow rules via a hardcoded if c != 'cd'. /setup's Phase 2 read that same flat list and reported Bash(cd:*) as a missing allow rule on every project, every run — contradicting the skill's own hard rule to prefer false negatives, and appearing during first-run setup.

Fixed by splitting the data rather than hardcoding the exception twice:

• permissions.yaml now has commands: (emitted as allow rules and validated) and validate_only: (validated but never emitted — cd).
• sync.py's emitter reads commands:; the if c != 'cd' special case is removed — the exclusion is now a property of the data. A new _validated_commands() helper feeds the union to both validators, so cd stays legal in skill code blocks.
• setup.md Phase 2 reads commands: only, so it structurally cannot report cd as missing.

The exclusion lives in one place; the emitter, validators, and setup audit can no longer drift. Verified by simulating Phase 2 against this project: zero missing rules, no Bash(cd:*). Five regression tests lock in the split.

Pruning

• setup.md 491 → 392 (−20%): removed the fixed health-summary block, permission-report block, both signing confirmation boxes, and the label-list/change-preview boxes; deduped the doubled gh label list re-fetch and the starter-label baseline. Kept: every confirmation gate, state-detection ordering + stop-at-first-failure, the profile→values table and inference rules, the REVIEW_EFFORT policy note, and all hard rules.
• submit-for-review.md 455 → 422 (−7%): four document templates (pr_body, resolution_comment, followup_body, followup_link_comment) replaced with descriptions of required content (keeping the resolution comment's non-technical-audience rule); the no-heredoc warning that appeared twice in immediate succession is now singular. Kept: the CC review contract and finding tags, all four verdict tiers and routing, the sensitive-area gate, Step 9's gating conditions, the branch guard and working-tree drift restore, the make -n check gate, and every closing-keyword note.

Reduction recorded

Combined the pair is 946 → 814 (−14%), below the notes' ~20–25% estimate. setup.md hit its ~380 target; submit-for-review.md reduced less because its listed prune surface (four small templates + one warning) totals ~33 lines — everything else is keep-side policy. Cutting further would remove the review contract or verdict tiers, which the ticket explicitly rules out. This is the modest-reduction outcome the ticket anticipated, not a shortfall to be closed by cutting deeper.

Verification

• ./sync.py --validate clean (placeholder, permission, command-shape); all four adapters regenerated.
• Test suite: 128 passing (+5 for the permission split).
• Keep-list audited across both skills; no gate, ordering rule, profile mapping, verdict tier, review policy, prompt-avoidance instruction, or platform-behaviour note lost.
• /submit-for-review exercised end to end by this PR. /setup's configured-path Phase 2 verified above; its unreachable first-run and partial-setup states reviewed by reading the rendered output.

Closes #208

[sebastientaggart]

Code review — /code-review medium + Code Cannon policy

Reviewed the diff across the 8 finder angles, with attention on the real sync.py changes (the commands: / validate_only: split) alongside the two skill prunes.

Sensitive-area gate: considered, not triggered. The change touches permission-allowlist generation, but the emitted allowlist is byte-identical to before (cd was already excluded via the old hardcode) — no new command is blessed and no runtime permission surface changes. It is a refactor of where the exclusion lives, not what gets allowed.

Correctness: the sync.py change is behavior-preserving. Both validators (validate_permissions, validate_command_shapes) previously read commands: (which included cd) and now read the commands + validate_only union — so cd remains a recognized command exactly as before. Only the emitter changed, and its output is unchanged (verified: 11 rules, no Bash(cd:*)). The removed if c != 'cd' special case is re-established by cd no longer being in commands:. Well covered by the five new TestPermissionCommandSplit tests.

Findings

• [WARNING] docs/index.md:126 — the recommended .claude/settings.json example led with "Bash(cd:*)", and /setup's Phase 2 points users to this doc "for a full example." That taught users to add the exact allow rule Code Cannon deliberately never emits and this PR's audit fix (correctly) will never request — an inconsistency between two parts of the project, caught by no test. Fixed in d636fe0 by removing the line and adding a one-sentence note explaining why cd is omitted.

No correctness, keep-list, or altitude findings on the skill prunes: the keep-list (confirmation gates, state-detection ordering, profile mapping, review contract, all four verdict tiers, sensitive-area gate, Step 9 gating, branch guard, closing-keyword notes) was audited present, and the removed content was display formatting and duplicated templates only.

Verdict: APPROVE