fix(review): answer covered cases on skill prose with the condition, not a patch

[tmchow]

Summary

Follow-up to #1397, which took nine review rounds and 24 bot findings to land a two-condition change. Every finding was valid in isolation; every fix added a case to the previous fix; the block converged only when restated as its conditions. Two guards that should have stopped it existed and did not fire: the AGENTS.md accretion rules live in authoring context read at session start, not in the protocol running when feedback arrives; and the resolver's two-cycle loop cap counted per invocation, which ce-babysit-pr resets every round.

This PR puts the rule where both sides of a review read it, and states it for the whole skill lifecycle:

• AGENTS.md → ## Working on Skills. The former Cross-Model Skill Authoring block and its subsections are restructured into one lifecycle with a shared standard: authoring a new skill (conditions, not procedures; owned mechanisms; prose admission; invocations; loading), editing an existing skill (new: bring the touched block up to the current standard rather than matching the old shape; scope to your change plus what it makes wrong), reviewing a skill change (new, addressed to bots and humans: what a finding is on skills/**, a covered case is not a finding, request a condition or an owning-layer move — never a case), and acting on review feedback (the existing steps, with the covered-case and second-round rules made explicit). Review bots here cite AGENTS.md line-by-line in their findings, so this shapes what gets filed as well as what gets fixed. No net new rules for src/, tests/, scripts/.
• ce-resolve-pr-feedback rubric. The project's review guidance in context frames the verdict, not just the harm veto. New "instruction prose is not code" section: a case the condition already decides → not-addressing quoting it; fix only the condition or the layer; a second round against the same block → restate, not qualify. Ordinary code keeps the ordinary rubric.
• Loop cap counts rounds per PR (from the branch's review-fix commits), so it fires across babysit re-invocations.
• Fixer prompt: on prose files, implement the restatement it was handed; return blocked rather than add a case.
• Learning: docs/solutions/skill-design/prose-review-is-unbounded-answer-with-the-condition.md.

Validation

bun run test: 3128 pass, 0 fail; bun run release:validate in sync. Prose-only change; the behavioral effect is evaluable only by the next skill-file PR review.

Security Disclosure

No security-relevant changes.

Agent Disclosure

• Model: Claude Code · claude-fable-5

[cursor]

PR Summary

^{Cursor Bugbot is generating a summary for commit 35e025f. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 35e025fcb1

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".

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a2a08ae. Configure here.}