Stream live recording audio and transcripts

Upgrade to WebSocket, then send an auth message containing the recording capability within ten seconds. Never put the token in the URL. Send TaskCreationStreamMessage audio packets separately for each registered audio track: base64 Float32 little-endian PCM, sequence, sampleOffset, offsetMs, sampleRate, and channels. offsetMs is the source's actual capture start on the recording timeline. Receives TaskCreationStreamResponse ready, ack, transcript, and error messages. A transcript may include segment {id, text, startMs, endMs, final}; trackId remains on the outer message. Replace the text for an existing segment id. Segment timestamps are recording-relative milliseconds from provider audio timing plus capture and reconnect offsets, not message arrival time. final becomes true when a 1.5-second speech pause completes the utterance; intermediate provider chunk updates remain partial. Live transcripts are provisional; upload source chunks separately for durable capture and final word alignment.