Treat STT audio-window offsets as required when they slice audio

In a hybrid setup where a streaming turn detector decides end of turn and a batch transcriber does the text, cutting "everything buffered" at end of turn leaks the start of the next turn, because the end-of-turn event can arrive after new audio is buffered. The detector's sample-indexed window offsets are the only valid boundary, so treat them as required, and do not fall back to receipt time. In the same path the provider once sent a word start of about minus ten nanoseconds. A strict non-negative check raised a protocol error, and the coordinator tore down both streams, so no turns came through. Clamp tiny negative float jitter (under about a microsecond) to zero, still reject clearly negative or non-finite values, and use the exact live payload as a regression test.

Applies when: STT or turn-detection metadata decides which audio samples reach another model. Keywords: STT optional offsets audio window turn detection Deepgram Flux timestamps float jitter protocol error hybrid transcription