feat(skills): add adversarial-review and gate issue-ship on it

[paveg]

What

A skill that tries to refute what a change claims about itself, and an issue-ship gate that refuses to open a PR until it passes.

File	Change
skills/adversarial-review/SKILL.md	new — extract claims from the diff / PR body / commit messages / subagent report into | claim | source | kind |, derive the implicit claims an attacker would test (Step 1b), route by kind, adjudicate, emit ADVERSARIAL: PASS or FAIL (n refuted, m unverifiable)
skills/issue-ship/SKILL.md	Step 4 runs it after the implementer returns; FAIL sends refuted claims and evidence to a fresh implementer dispatch; Step 5 never starts on FAIL

Routing by claim kind — existing pieces do the checking, the skill only orchestrates:

kind	check
behavior	one skeptic agent per claim, parallel, cap 6 by blast radius
doc	verify-doc-claims on the touched md files, table taken as-is
test	break one condition in a throwaway worktree, suite must go red
count	enumerate before/after, account for every item by name

Subagent verdicts stay hypotheses: a REFUTED is accepted only after the main session opens the cited evidence.

Step 1b (second commit; the third only unwraps the description onto one line) adds the attacker stance: stated claims are chosen by the author, so the author's blind spots never reach the table. Five angles — trust boundary, error paths, ordering, touched callers, removal — each become an implicit claim with source: implicit and ride the same routing under the same skeptic cap. A hypothetical that never ran is a row without evidence and does not count against the change. Security attack surface stays with pentest-parallel-prs.

Design choices

• Mutation for test claims runs in the main session in a git worktree add throwaway, because skeptic is read-only. Alternative was marking such claims UNVERIFIABLE.
• The cap of 6 skeptics is derived from workflow.md's ~10K-token spawn cost, not measured.
• Description trimmed 490 → 298 chars (context-diet/scripts/measure.py) before commit; routing detail lives in the body.

Not in this PR

Behavioral verification of the skill (a run through empirical-prompt-tuning, or an end-to-end issue-ship pass). just test covers format and lint only.

🤖 Generated with Claude Code

[changeset-bot]

⚠️ No Changeset found

Latest commit: 101b12d

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[paveg]

Test results

• Target commit: 101b12deccd5a9c3079562d1934a276d9f0dba02

Check	Command	Result
Lint + render + format + hook/skill-script tests	just test	✓ All checks passed!
Always-on cost of the new description	context-diet/scripts/measure.py	332 chars (trimmed from 490 before the first commit; +34 for the attacker-stance wording)
Commit signature	git log -1 --show-signature	Good "git" signature (ED25519)

Not covered

• The skill has not been executed: no empirical-prompt-tuning pass, no end-to-end issue-ship run through the new gate.
• The throwaway-worktree mutation step for test claims is specified, not exercised.