In voice evaluations, a simulated listener must hear only audio that actually played

In voice evaluations, a simulated listener must hear only audio that actually played. Freezing an actor WAV before queued playback finishes can leak words that a subsequent transport clear discards; resetting the playback deadline without waking a waiter can also add delay for speech nobody heard. This applies to local phone-media simulations with buffered output and interruption/clear semantics. It does not establish acoustic quality or behavior of a physical handset.

Maintain per-frame scheduled playback and a clear-truncated played prefix. Wake pending playback waits on clear, and build listener input only after those prefixes have settled. Keep attempted interruptions separate from measured overlap, and serialize partial and ordinary actor callbacks so one question cannot consume two driver replies. A transcript or intended TTS string cannot substitute for that played-audio evidence.