Voice attestation AI sign-off shipped (flag-off): speaker verification with phrase-challenge anti-replay

Voice joins face as a biometric modality for AI order sign-off. New kit adapters add speaker verification on a kaldi log-Mel front end, with WeSpeaker and TitaNet model presets selectable via an env var — swapping models is config, not code.

A phrase-challenge liveness step guards against replay, and the attestation is bound to the order's target reference so a captured voiceprint can't be reused for a different action.

The feature ships flag-off and reuses the existing biometric migration. Note: VoxCeleb weights are non-commercial, so production deployments ship NVIDIA TitaNet or a self-trained model.