Why: The first r0.3 code review found three unsafe inference paths despite green mechanism tests. Evidence: Reviewer's read-only replay, Backend RED→GREEN counterexamples, isolated PostgreSQL and task-owned HTTP browser receipts are identified separately. Impact: Keep the old review failure and new fix candidate traceable; real-model quality stays NOT_RUN. Refs: Dev task feature/fast-slow-u2-u3-plan; Documents PR #95