fix(gateway): guard empty/inaudible voice transcripts in enrichment

[deacon-botdoctor]

Summary

• speech-to-text can return success=True with an empty/whitespace transcript on silence, cut-off, or inaudible audio
• previously this emitted Here's what they said: "", sending empty content to the model (which can hallucinate a reply and loop)
• extract a small pure helper that emits a clear sentinel note for the empty case instead of empty quotes

Validation

• python3 -m py_compile gateway/run.py tests/gateway/test_voice_transcription_note.py
• .venv/bin/python -m pytest tests/gateway/test_voice_transcription_note.py (4 passed)
• git diff --check

Scope check

• rebuilt from fresh NousResearch/hermes-agent main
• scanned staged diff for obvious secrets/private identifiers and downstream overlay markers

[liuhao1024]

✅ Verified — empty voice transcript guard is correct

Reviewed the diff for gateway/run.py and tests/gateway/test_voice_transcription_note.py.

• Empty/whitespace handling: _voice_transcription_note() correctly checks transcript.strip() — empty string, whitespace-only, and missing key all return the sentinel note asking the user to resend. Real transcripts get quoted as before.
• Agent-loop prevention: The sentinel note explicitly says "Do not guess at the content; ask the user to resend or type it out." — clear instruction that prevents the agent from replying to silence.
• Test coverage: 4 test cases cover non-empty, empty, whitespace-only, and missing transcript key. Edge cases are well covered.

The fix is correct and the extraction into a helper function is clean. No issues found.

[teknium1]

Thanks for identifying the blank-transcript path. The premise remains valid on current main: gateway/run.py:15247-15255 still turns a successful empty transcript into "".

Problems

• The submitted helper restores the old voice-message wrapper and adds a Do not guess directive. Current main deliberately replaced that format with a plain quoted transcript at gateway/run.py:15250-15255; commit 5582b51a680b16b399a82e9c04b3317c0fd06255 removed STT-mode prompt chatter because it persisted into later turns.
• Current code appends every successful transcript before formatting it (gateway/run.py:15247-15249), while queued paths echo those values at gateway/run.py:15327-15339. A blank guard must avoid that append too, or users can still receive an empty 🎙️ "" echo.

Suggested changes

• Apply the blank check in the current success branch, preserving the existing neutral [voice message could not be transcribed] contract for blank results and excluding them from successful_transcripts.
• Add an async regression test through _enrich_message_with_transcription() that verifies blank and whitespace-only successful results produce no quoted empty prompt and no echo candidates.

This is an automated hermes-sweeper review.