review-codelisted

# Review Code Risk-appropriate review pressure on a finished change: multi-axis read, adversarial review for high-risk diffs. ## Iron Rule <EXTREMELY-IMPORTANT> 1. NEVER merge code on a high-risk surface (auth, billing, payments, credits, migration, data deletion, secrets, tokens, crypto, permissions, security) without an adversarial fresh-context review. 2. NEVER let the author — or, for the adversarial pass, the author's own model — be the final reviewer of their own change. - The external adversarial review runs in a **CLI** different from the Lead's, on that CLI's own default model (same vendor is fine — harness, context and defaults decorrelate). - The vertical fallback (same CLI, stronger tier) and an inline advisor (it advises the author inside the author's context) never satisfy it; both only raise the Lead floor and are recorded as a limitation. 3. NEVER skip review because "tests pass". Tests prove the assertion, not the design. 4. Findings before fixes — the whole round's findings, never the first report's. Severity-ordered list; no silent rewrite. 5. The author MUST verify findings against the codebase before implementing. No performative agreement, no blind implementation; clarify unclear items before partial implementation — findings may be linked. </EXTREMELY-IMPORTANT> ## Skip when - Pure docs / typo / whitespace · the user explicitly accepts the change unreviewed · R1: ≤5 lines, single file, zero logic (user-facing string text alone counts as zero