Design voice-agent audio decisions as typed, calibrated outputs with abstention and failure routing instead of treating every problem as transcription