feat(voice): surface FlushSentinel boundaries in transcription_node

[longcw]

Summary

A FlushSentinel yielded from llm_node was stripped before the transcription node, so a single reply was always one transcript segment with no way to detect the internal boundaries.

FlushSentinel now takes an optional id and is surfaced into transcription_node as an empty-string FlushSentinelText carrying that id. A custom transcription_node can detect the boundary and read the tag; it is dropped before reaching the TextOutput, so default consumers are unaffected. The published TranscriptionSegment.id is unchanged.

async def llm_node(self, chat_ctx, tools, model_settings):
    yield "Yes, let me look that up. "
    yield FlushSentinel(id="ack")        # flush to TTS, mark a boundary
    ...
    yield "It is 18 degrees and sunny."

async def transcription_node(self, text, model_settings):
    async for chunk in text:
        if isinstance(chunk, FlushSentinelText):
            # boundary reached; chunk.id == the id from the FlushSentinel
            ...
        yield chunk

Notes

This flows through the non-aligned transcription path. With use_tts_aligned_transcript, the transcription input is TTS-derived timed text and carries no FlushSentinel, so the marker does not appear there.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.