fix: route Weixin voice replies to real file attachments

[imink]

Summary

• force Weixin send_voice() to use the reliable file-attachment fallback instead of native outbound voice bubbles
• make .silk use native voice metadata only when the caller explicitly wants a voice payload
• add regression tests covering audio attachments, forced .silk attachment fallback, and Weixin voice payload metadata

Testing

• source /home/yuewang1/.hermes/hermes-agent/venv/bin/activate && python -m pytest tests/gateway/test_weixin.py -q

[teknium1]

Merged as part of Batch-3 salvage: #11571

All three of your commits were cherry-picked onto current main with authorship preserved (Patrick Wang). The voice-fallback strategy resolves the TTS-audio delivery gap reported in #9971 — users now receive playable audio as a file attachment instead of a broken voice bubble. SILK encoding metadata preserved for the direct _send_file() path. Thanks for the thorough 3-commit trajectory showing both the SILK approach and the pragmatic fallback.

Commit SHAs on main: 678b69e, 649f383, 4ed6e4c