CI-style truth oracle for AI coding agents — mechanically re-runs an agent's stated claims (tests pass, build green, only these files changed, lint clean) in the actual repo and emits a deterministic pass/fail receipt. Verifies claims, not code quality; U