multimodel-reviewlisted

# Multi-model review (GLM + M3) Reusable recipe for the cross-model review step. Use on a FOCUSED diff/snippet only. ## Steps 1. Stage focused text (`git diff -- <paths>` or a snippet). No `.env`/secrets/whole-repo (the CLIs also hard-block, but check). - **Give reviewers enough CONTEXT, not just the raw diff (v2.5).** A diff-only view is the #1 source of false positives — a model sees a call to `confirmClock()` in the diff, can't see its definition, and reports "undefined function." When a finding could hinge on code outside the hunk (a helper/type/import defined elsewhere), attach the FULL relevant file(s) alongside the diff — still T0–T2 only, still no secrets/whole-repo. Cost is a few tokens; it removes a whole class of phantom P0/P1s. 2. GLM red-team (full-ZDR): `git diff -- src | <glm-bin>/glm-redteam "Find concrete P0/P1 bugs. Be concise."` 3. M3 second auditor — L2/L3 only; M3 is a reasoning model so raise the budget: `git diff -- src | M3_MAX_TOKENS=12000 <glm-bin>/m3-audit "Find concrete bugs/risks. Be concise."` On L0/L1, skip M3 (one reviewer + tests is enough). 4. Reconcile (cross-model agreement): CONVERGENCE (both flag) = high-confidence must-fix; DIVERGENCE (only one) = verify against the code, that is where the 2nd model earns its keep — don't discard it. 5. Every finding is a hypothesis — verify against actual code before acting. 6. Confirm the audit trail: `tail -n 4 ~/.config/openrouter/egress.log` (hash only, privacy-routed). 7. **Log the re