perf(release-audit-report, release-archive-sweep, release-keys-sync): decide and render as code

[potiuk]

Summary

Last PR of the release-management stack; it builds on the prepare PR.

• Scripts move. The last three skills' scripts move unchanged into tools/release-management: render_record → audit_report.record, retention → archive_sweep, check_key → keys_sync.
• audit-report gets verdict, gather (the mail-archive search plan; the search itself stays with the agent), voters (roster handles through vote-tally's binding resolution, never citing an unresolved voter), check, record and pr. The PR step moves, byte-identical with its generated adversarial block, to audit-log-pr.md at the skill root.
• archive-sweep gets verdict, listing, trains, retention and render. The render covers svn mv, gh release delete (with the permanent-deletion warning), aws s3 mv, self-hosted, and atr (no command).
• keys-sync no longer finds the key by text scan:
  • A new read-only release-verify list-keys lists fingerprints and UIDs in the existing throwaway GNUPGHOME. It uses no agent, no keyring and no network, and refuses private-key material.
  • keys-sync verdict classifies that output as proceed / noop / blocked.
  • render draws the KEYS block, the svn / git / generic sequence and the upload reminder.
• Size: audit-report goes from 6,657 to 5,136 tokens, archive-sweep from 4,573 to 4,165 and keys-sync from 4,871 to 4,584.

Security notes

• Every rendered command quotes its values with the shared helpers.
• release-management still runs nothing. The only gpg call is release-verify list-keys, with the same isolation guarantees as its other checks.

Behaviour changes worth a look

• keys-sync matches parsed fingerprints, primary or subkey. A fingerprint that appears only in a KEYS comment no longer counts as present.
• keys-sync fails closed when it cannot compare (no keyserver key or UID). noop means the keyserver's primary UID is among the KEYS entry's UIDs.
• The svn blocks in all three skills lose their \ # comment continuation lines, which broke a pasted command.
• archive-sweep: gh release list adds --exclude-drafts. The self-hosted backend needs --command-template, since no config key holds that command.
• audit-report refuses a planning-issue URL that isn't canonical, and escapes markdown in a hostile product name.

Type of change

• Skill change — eval fixtures updated below
• Python package (tools/release-management, tools/release-verify)

Test plan

• prek run --all-files passes on the stack tip.
• tools/release-management and tools/release-verify: pytest / ruff / mypy, with three real-load integration tests. The gpg-based list-keys tests ran.
• Evals: the scripted cases are now pytest cases named after each former case, with byte-identical expected records. The kept judgement and injection cases pass: audit-report 4/4, archive-sweep 2/2, keys-sync 2/2. The baselines were 9/9, 10/10 and 9/9.

---

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