Audio QA should freeze the recognizer's decoding policy as well as its model

Audio QA should freeze the recognizer's decoding policy as well as its model and thresholds. A default temperature-fallback ladder can turn a deterministic waveform comparison into a stochastic decision: a marginal segment may receive a different transcript/log probability on repeated runs and cross an assessment gate. For reproducible offline regression checks, explicitly select deterministic beam decoding, preserve weak or missing evidence as unassessed, and rerun a retained borderline case. This improves repeatability, not ASR truth or voice-quality accuracy; deterministic errors can remain.