perf(review): shorter reviewer reports; draft verdict-independent records during review

[cskwork]

In two measured compact runs (taskdesk complete <id>, Codex gpt-6.1-sol high, lazymode 4), proof→ship was the largest span: 302 s of 462 s and 251 s of 552 s. It was mostly the fresh reviewer, whose reports ran to about 2,900 words, followed by 63–90 s of author paperwork after the verdict. Between 13% and 17% of each report restated recipe checks the receipt already proves, and 12–13% was narrative under adversary "Checked and clean".

Change (no check removed)

• verifier.md: a current receipt is cited once for the recipe checks it covers (path, recorded_at, VERIFY and coverage lines). A recipe check is listed on its own when it failed or when its log was read, and commands run outside the receipt are still listed individually. Lens 1's duties to inspect actual logs and validate the live receipt are unchanged.
• adversary.md: "Checked and clean" is one line per area. Per-file rows and the Blocking/Non-blocking sections are unchanged.
• build skill: while reviewers run, the author drafts qa-guide.md, harvest.md and the parts of evidence.md outside Verification and Adversarial code review, so ship only adds the reports and approves. A finding or gap revises the drafts.
• Word budgets: raised by exactly the added words (verifier 2,420→2,456, adversary 818→824, build 1,337→1,374). These are new instructions, not drift.

Validation

• Adversarial semantic review (Codex gpt-6.1-sol high, read-only): NO FINDINGS.
  • Receipt citation does not replace inspection: roles/verifier.md:19, and gates/approve.sh:158 rejects failed or stale verification.
  • Provenance, per-file rows and VERDICTs remain required: roles/verifier.md:174, roles/adversary.md:87.
  • Drafts must still match final evidence: skills/5-ship/SKILL.md:50, templates/qa-guide.md:38, AGENTS.md:221.
• Local macOS: gates/selftest.sh SELFTEST PASS, and git diff --check is clean.

Speed is not yet shown. A paired 3+3 comparison against main (isolated HOME/CODEX_HOME) was started but is not complete. This PR claims no time saving.