bench: run the review-verbosity arms and promote

[rookslog]

Child of #105. Two arms on the same PRs and model: baseline review prompt vs the fixed per-finding shape. Metric: confirmed findings per thousand output tokens. Counterweight: confirmed findings the compressed arm drops. Arms carry manifests, figures derive from the scorer, promotion is a reviewed act under the measurement design.

[rookslog]

Compact review contract — Opus 5 via claude -p (DRAFT, unmeasured)

Treatment arm for the #109 study. Compresses the per-finding format and the
framing around it. It never caps or filters the finding count: Opus 5 follows
severity filters literally, which depresses recall — so compression targets
shape, not coverage.

The contract (prepend to the review prompt)

Review this diff for defects. Findings are your entire output.

Each finding is one line-group, in this exact shape:
<file>:<line> — <high|med|low> <confirmed|plausible> — <one-sentence defect
claim>. <one-sentence concrete failure scenario>. [fix: <one clause, optional>]

Rules:
- Rank by severity, highest first.
- "confirmed" means you traced the defect in the code shown. "plausible" means
  pattern-matched, not traced. Never present plausible as confirmed.
- Report every finding you believe in, at every severity. Do not filter by
  severity, cap the count, or merge distinct findings.
- No patches unless asked. No preamble, no restatement of what the change
  does, no praise, no summary section, no conclusion, no advice sections.
- If nothing clears your bar, output exactly: No findings above the bar.

Invocation shape (from codex or any harness)

git diff main...HEAD | claude -p "$(cat opus5-compact-review-contract.md)" --model opus

How this gets judged (#109)

Metric: confirmed findings per thousand output tokens, against the current
default-review control. Counterweight: recall — a confirmed finding the
control caught and this arm dropped counts against it, at any severity.
Disposition of each finding comes from the verdict blocks (#108 mines them).