perf(release-prepare): plan, prep, post and the signing setup as code

[potiuk]

Summary

Fifth PR of the release-management stack; it builds on #1618.

• Scripts move. release-prepare's three skill-local scripts (prev_tag, next_dev_version, category_x) move unchanged into tools/release-management.
• New commands cover what the agent did by hand:
  • verdict: Step 0 train coverage.
  • plan-query / plan: the planning issue.
  • versions, notice and changelog (with the 90% coverage check).
  • archive-mode / archive-review: the 2e source-archive review.
  • prep-pr and post-pr.
  • signing-eligibility / signing-render: the automated-signing A1 gate and the A4/A5 drafts.
• Size. SKILL.md goes from 5,532 to 5,027 tokens, and prep.md from 290 to 209 lines; the 2e review moves to archive-review.md at the skill root.

Security notes

• PR titles are external content. They are rendered as inert text in the planning issue and the changelog: HTML is escaped and line breaks are folded. A title holding <!-- … --> can no longer be hidden from GitHub's renderer while a later agent still reads it.
• The 2e listings now write to $TMPDIR, not /tmp, which broke under the sandbox.
• signing-render refuses key material or a signing step. Step A still never files the Infra ticket, sends mail or adds a key.

Behaviour changes worth a look

• Step 0 is fail-closed. It blocks without GitHub access, on a planning issue missing its label, and on a release line it cannot read until --train-confirmed is passed.
• Changelog categories.
  • A new optional key, changelog_label_categories (LABEL=SECTION, documented in the config template), maps labels to sections. Unset, a label is sorted by its words.
  • prep-pr refuses coverage below 90% until --assign classifies the rest, matching the existing "ask the RM to classify before the PR is opened".
• Planning-issue body: groups are sorted alphabetically with unlabelled PRs last, and numbering runs across groups.

Type of change

• Skill change — eval fixtures updated below
• Python package

Test plan

• prek run --all-files passes on the stack tip.
• tools/release-management: pytest / ruff / mypy, including the moved script tests and a real-load integration test.
• Evals: the scripted cases are now pytest cases named after each former case. They are step-0 1/2/4, step-1-plan 1/2, step-2-prep 1/2, and the whole step-14-post suite. The kept judgement cases pass 6/6, against a baseline of 15/15.
• non-asf-profile-smoke: both of its prepare cases pass. Two of its other cases fail on main independently of this stack; see the notes below.

Notes for reviewers

non-asf-profile-smoke has two failures that reproduce on runs without this change:

• step-reviewer-routing case-2. reviewer-routing was split in perf(pr-management): run mentor, reviewer-routing and pre-first-pr-check rules as code #1583, and this smoke case still grades the old prose step.
• step-contributor-governance-intake case-2. The model answers in prose instead of JSON.

Both are follow-ups outside this stack.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 5). I reviewed the diff and the test results; commits carry Generated-by:.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GTPsZ5aEE5bUVYqp47Dpvv