feat(gateway): echo voice transcripts back to user before agent reply

[dgsratiu]

Summary

• When STT transcription succeeds, immediately sends the transcribed text back to the user as a 🎤 "..." message before the agent processes it
• Lets users confirm what was heard without waiting for the full agent response
• Matches the echoTranscript: true behaviour familiar from other agent frameworks

Implementation

Stores successful transcripts on _pending_voice_transcripts during _enrich_message_with_transcription, then flushes them via the platform adapter at the call site before the agent turn begins. Failures are silently swallowed so a send error never blocks the agent response.

Test plan

• Send a voice message via Telegram — should see 🎤 "transcribed text" echoed back immediately, followed by the agent reply
• Send a voice message when STT is misconfigured — echo should not fire, existing error message still sent
• Send a text message — no change in behaviour

🤖 Generated with Claude Code

[alt-glitch]

Duplicate of #13727 (and #22123). Third independent implementation of the STT transcript echo feature. #13727 is the original with stt.send_transcription config key.

[teknium1]

This is an automated hermes-sweeper review. The transcript-echo behavior requested here is now implemented on main.

Evidence:

• gateway/run.py:7609 calls _enrich_message_with_transcription(...), receives _successful_transcripts, and sends each transcript back through the platform adapter before the agent turn continues.
• gateway/run.py:11541 defines _enrich_message_with_transcription(...) -> tuple[str, List[str]]; successful STT results are collected as raw transcript strings for caller-side echo.
• gateway/run.py:11682, gateway/run.py:14691, and gateway/run.py:15059 also echo transcripts in the active-run interrupt/dequeue paths, so queued voice messages get the same behavior.
• The implementing commit is d55304c39f204dd8b9e23f3bb681cc2f3472dc56 (fix(gateway): transcribe voice messages during active agent runs), whose message explicitly notes that it echoes raw transcripts back to the user.

The duplicate note linking #13727/#22123 was useful context; this narrower PR's requested behavior is covered by the implementation now present on main.