review: treat a new negative assertion as a decision, not as coverage - #2861

An assertion that a behavior is ABSENT (`not in`, `assertNotIn`,
`not.toContain`, `pytest.raises` on a call that used to succeed) is part of the
behavior change, not test coverage — and it arrives in the diff wearing
coverage's clothes, so scope-creep detection walks past it.

Three things the review now does:

- Scope check: a new negative assertion needs the stated-intent line that
  authorizes the removal. With nothing to cite, it is scope creep signed as
  coverage.
- The testing specialist states the revert cost in the finding. Once merged,
  restoring the old behavior makes these tests fail, so whoever reverts looks
  like the author of the regression.
- The one that is easy to miss: a PRE-EXISTING negative assertion emptied by
  this diff. A rename or removal makes a standing `assert "<old token>" not in
  output` true forever, so it stops guarding without ever going red. Grep the
  diff for renamed or deleted identifiers, then grep the suite for negative
  assertions naming the old token — including the ones inside tests that assert
  something else.

The scope-check bullet is edited in `scripts/resolvers/review.ts`, which feeds
both `/review` and `/ship`; `review/SKILL.md` and `ship/sections/plan-completion.md`
are the regenerated output. `review/specialists/testing.md` is hand-maintained.

Kept deliberately terse in the skeleton: `/review` sits close to its
`maxSkeletonBytes` ratchet, so the always-loaded bullet is one line and the
detail that earns its keep lives in the specialist, which loads on demand.
Parity suite: 21/21.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>