what we measured

we compared output audio, not whether the transcript looked responsive. the protocol defines acoustic windows and non-inferiority gates before the scored baseline. the arms differ in model, voice, transport and delegation. this compares products, not a single changed component.

round 2 still failed

the round 2 scoreboard keeps the earlier omp control rather than rerunning it. direct-turn end-of-speech to first-audible:

arm p50 ms p95 ms
omp-live 1,112 1,539
z0live-bidi 1,265 1,960
z0live-realtime 1,815 2,978

both candidates were FAIL. realtime missed the direct-response latency gates. bidi also missed interruption gates. all arms got the repository answer correct 5/5 in this round; that did not make the latency failures disappear.

barge-in to silence p50 / p95 was 974 / 1,220 ms for omp, 1,190 / 1,392 ms for bidi and 237 / 239 ms for realtime. the control itself misses the absolute interruption-tail gate. it is not ground truth for acceptable experience.

the narrow fix was narrow

the later bidi subset covers only barge-in and rapid follow-up. missing audible replies went to 0/15, and bidi barge p50 / p95 was 885 / 1,167 ms. there was still 1 stale resume. the scoreboard remains FAIL, with delegation gates undetermined. its absent repository scenarios are not evidence of incorrect answers; they were not run.

we can credit the buffered-tail fix without promoting the whole voice stack.

what this does not prove

the protocol uses a headless client and synthetic speech, not a physical phone microphone and speaker. omp is metered before speaker buffering; chromium after its jitter buffer. 19 omp turns were excluded for harness stalls in the round 2 scoreboard; the included-stall alternative is also reported there. later blocks are not a fresh matched control, and the host was loaded. no acoustic-naturalness or mobile-background claim follows. raw audio and transcripts stay out of this draft.