review-buildlisted

# compass:review-build (Review-3 · FULL) Lens: **is the BUILT thing correct, complete vs the contract, and safe — proven, not vibed?** ## Step 0 — gate Run `compass.sh gate .claude/builds/<slug> build`. **Non-zero → STOP** (build incomplete), offer the right earlier stage. Read `contract.md` + `plan.md`. Set `progress.md` = `in-review (R3)`. Check the product against the contract feature-by-feature. ## Engine - **Ledger** (create if absent): same columns. - **Material** = new Critical/Major. **Clean round** = zero new material AND the regression suite RE-RUNS green. **Proof-of-work footer:** `> Round N (R3): suite=\`<cmd>\` exit=0 passed=k/k; reconcile→PASS; new Crit/Maj=0. Clean? yes` — no command line = not clean. **Converged = two consecutive clean rounds.** Cap **5**. - **Fan-out economy:** round 1 spawns all groups; **rounds 2+ spawn the groups the last round's fixes touched** — the full regression suite still re-runs every round. Closure = Validation command **re-run with fresh output**; agent agreement is not evidence. - **A FIX IS NEW CODE — re-attack it (the rule that catches self-introduced defects).** A fix can introduce a *new* defect, including the exact class it fixed (a pagination fix that opens an IDOR; a redaction fix that misses a field). So: **any round that applied a fix is NOT clean by definition.** The next round MUST adversarially re-attack the fix surface, and the **independent agents [D] Security/RBAC, [E] Secret-leak, and [F] Verification-audit A