[Tarzelf]

The v0.3 local verifier trusts its primary observation packet. This change adds a separately processed RGB camera path and PiPER drive-channel checks, then measures where those extra signals catch or miss false completion readings.

The fresh native benchmark covers 272 cases across eight layouts: 96 native scenes, 80 primary-packet forgeries, 64 auxiliary-channel faults and 32 cross-channel controls. The gate prevents 69 of 112 baseline false completions, while 43 remain through orientation, visual identity, support, temporal aliasing and replay blind spots. It also abstains on 73 of 80 baseline true completions, including 64 injected evidence faults and nine unobservable native positives. All denominators and every case are retained.

The complete run is in benchmarks/sensor-goal-verifier-v04/evidence/. Its source-frozen benchmark used 90e84701ce7d5bc330ef41b02793e2ce037d702d. Median/p95 processed-window latency is 22.1/31.5 ms; acquisition of the 200 ms evidence window, native rendering and actuation are excluded. These are shared-simulator synthetic channels, not authenticated physical sensors or hardware qualification.

Validation: 107 core tests and 27 Hugging Face publication tests pass; 11 native test cases requiring an asset environment variable are skipped in this generic unittest invocation. The pinned 8-layout benchmark itself generated all 96 scenes with zero invalid cases and was independently replayed before publication. OpenCV's CPU kernel was used for the recorded timing run.