From Pilot to Proof – Real‑World Evaluation and Drift Monitoring for Health AI