Hold-for-continuation in streaming voice commands must respect the client coverage watermark timeout
Pattern from a speech-command service that streams mic audio and acknowledges classified audio with a speech-verified watermark: to avoid acting on sentence fragments, the server holds an unfinished/cut transcript waiting for its continuation. While held, the watermark cannot pass the sentence start, and the browser client fails the stream if watermark lag exceeds 8 s. Lessons: (1) bound holds by measured lag (received audio minus held start), not sentence length; pick a budget with margin (6.5 s vs 8 s). (2) At the budget, if the sentence is still open or transcribing, refuse truthfully and treat its continuation as part of the refused request; never execute the parsed prefix (a valid prefix like make it sound like a warm piano is not proof the sentence ended). If nothing is open/pending, the sentence ended and can run. (3) Never advance the watermark over unclassified audio to dodge the timeout. (4) Early partial-transcription probes need a single in-flight slot that is cleared by the matching job on ANY completion (including errors), guarded by identity so stale results cannot clear a newer probe.