Confidently Wrong, Silently So: Auditing Undetectable Failures of a Deployed On-Device Language Model | ResearchPod