fix(gateway/weixin): route voice messages through Hermes STT instead of trusting Tencent Cloud's text

[Kewe63]

Summary

When WeChat (Weixin) returned voice_item.text (Tencent Cloud's STT output),
Hermes trusted that text as the user-visible message body and skipped
downloading the raw audio. For non-Chinese audio that text is garbage — the
original report was a Russian voice message that came back as English phonemes
— and the user sees nonsense as their own message. International users on the
WeChat gateway effectively can't use voice.

Two short-circuits in gateway/platforms/weixin.py caused this:

• _download_voice() returned None whenever voice_item.text was set, so
  the raw SILK/Opus audio was never fetched.
• _extract_text() returned voice_item.text verbatim as the body, so even
  if the audio had been downloaded, the central STT pipeline in gateway/run.py
  never had a chance to replace it.

Fix both: always download the raw audio (the central pipeline picks it up via
_collect_media()), and skip voice items in _extract_text() so the body
comes from Hermes' own mlx-whisper / whisper.cpp / faster-whisper transcription
instead of Tencent's. Behavior is unchanged when voice_item.text is absent.

---

Changes

• gateway/platforms/weixin.py:
  • _download_voice(): drop the if voice_item.text: return None
    short-circuit so the raw audio is always downloaded
  • _extract_text(): voice item no longer returns Tencent's text;
    the central STT pipeline in gateway/run.py now produces the body
• tests/gateway/test_weixin.py:
  • New TestWeixinVoiceAlwaysDownloaded class (5 tests): primary repro
    for both functions, _collect_media integration path, and a regression
    guard for the text-item path

---

How to Test

# Scoped
python3 -m pytest \
  tests/gateway/test_weixin.py::TestWeixinVoiceAlwaysDownloaded -v
# ✅ 5/5 passed

# Full file
python3 -m pytest tests/gateway/test_weixin.py
# ✅ 74/74 passed

---

Checklist

• Tests pass — 74/74 in file, 5/5 new
• ruff check — PASS, 0 warnings
• Follows Conventional Commits
• Changes scoped to this fix only (2 files, 151+/5-)

---

Risk & Impact

Low. Two single-function changes in one platform adapter with explanatory
comments at each edit site. Public API surface is unchanged; _download_voice
and _extract_text return shapes are the same. Behavior only changes for the
previously-broken case (voice_item.text set); the fall-through path is
exercised by the existing happy path (voice_item.text absent).

Type: 🐛 Bug fix
Closes: #27300

[teknium1]

Thanks for the focused fix. Current main still has both reported guards: gateway/platforms/weixin.py:963-966 returns Tencent's voice_item.text, and gateway/platforms/weixin.py:1646-1650 skips the raw-audio download when that field is present. The existing runner path will transcribe the resulting audio/silk voice attachment (gateway/platforms/weixin.py:1587-1591, gateway/run.py:10454-10458), so the production change matches the stated goal.

Suggested changes

• Add one integration-level regression test covering the final routing contract: an inbound Weixin voice item with Tencent text should construct a VOICE event with audio/silk media and reach _enrich_message_with_transcription. The new tests cover the adapter prerequisites, but not that gateway-runner handoff.

Automated hermes-sweeper review.

[teknium1]

Merged into main via consolidated salvage PR #73515 (merge c911a5f10f). Your Weixin fix routing voice through Hermes STT instead of trusting Tencent's voice_item.text was cherry-picked with your authorship.

Your contribution is credited to you in git history. Thank you! Closing this PR as merged-via-salvage.