fix(weixin): convert Silk voice messages to WAV via pilk for STT

[annguyenNous]

Summary

Fixes #42084

WeChat voice messages arrive in Silk audio format, which is not supported by any STT backend (Whisper, faster-whisper). Previously the gateway saved .silk files directly, causing silent STT failures.

Changes

gateway/platforms/weixin.py

1. Remove premature return: Don't skip voice download when voice_item has pre-transcribed text — the WeChat API text is unreliable, always download and transcribe via STT.

2. Silk → WAV conversion: After downloading, convert the .silk file to .wav using pilk (same approach as the QQBot adapter). Return the WAV path for STT compatibility.

3. Fallback: If pilk is not installed or conversion fails, return the .silk file as before.

4. Media type: Update caller to use audio/wav when the conversion succeeds, audio/silk as fallback.

Dependency

pip install pilk (optional — graceful fallback if not installed)

Testing

• Verified voice messages download and convert to WAV
• Verified STT (Whisper) can transcribe the converted WAV files
• Verified fallback to .silk when pilk is not installed

[alt-glitch]

Same fix as the open #31487 (decode WeChat Silk voice -> WAV so STT can transcribe) and overlaps #11593 (generic .silk preprocessing via pilk). Fixes #42084. Worth consolidating onto the shared _silk.py helper extracted from the QQ adapter in #31487 rather than inlining a second pilk path in weixin.py.

[tonydwb]

Code Review Summary

Verdict: Approved

Looks Good

• Fixes WeChat voice message STT: converts Silk format to WAV using pilk.silk_to_wav() (same approach as QQBot), enabling Whisper/faster-whisper to decode the audio.
• Graceful fallback to .silk if pilk is not installed or conversion fails.
• Removes the previous incorrect skip logic that skipped download when voice_item had pre-transcribed text.
• No security concerns, no debug artifacts.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for tackling the real Weixin STT gap. Current main still caches inbound voice as .silk (gateway/platforms/weixin.py:1660), while transcribe_audio rejects .silk (tools/transcription_tools.py:103,1032-1036), so converting before gateway STT is the right direction.

Problems

• The removal of the text-present branch changes documented behavior: _extract_text() already uses voice_item.text (gateway/platforms/weixin.py:963-966), and the Weixin docs say text is used when available (website/docs/user-guide/messaging/weixin.md:198). Please establish and test the intended behavior before always adding an STT attachment.
• The new inline pilk path duplicates QQBot's existing conversion logic (gateway/platforms/qqbot/adapter.py:2039-2091). This also matches the member comment recommending consolidation with the shared-helper work in #31487.
• Please add inbound conversion/fallback/MIME tests; the current Weixin SILK tests are outbound-only (tests/gateway/test_weixin.py:769-907).

Suggested changes

• Salvage the conversion through shared SILK decoding logic and add mocked inbound regression coverage for WAV success and SILK fallback.
• Preserve or explicitly revise the supplied-transcript contract, including its documentation.

Automated hermes-sweeper review.

[teknium1]

This removes the current transcript-first behavior: _extract_text() already returns voice_item.text on main (gateway/platforms/weixin.py:963-966), and the docs describe that contract (website/docs/user-guide/messaging/weixin.md:198). Please add a regression test and explicitly decide whether an upstream transcript should be retained, replaced, or accompanied by STT output before changing this branch.

[teknium1]

QQBot already has pilk conversion with output validation and retry behavior in gateway/platforms/qqbot/adapter.py:2039-2091. Please consolidate this through shared SILK decoding logic, as the member comment on this PR suggests, instead of maintaining a second inline implementation.

[andrexibiza]

Verification (Vox Lockin lane 08 — platform adapter voice class): this PR is superseded by main and now conflicts with it; its adapter-level pilk conversion is redundant.

Status vs origin/main @ 70db671:

• git merge-tree shows a conflict in gateway/platforms/weixin.py — main has since restructured _download_voice (the WeChat voice messages use Tencent Cloud STT which garbles non-Chinese languages — should route through Hermes' own STT pipeline #27300 fix, cb1798b43c) so it always downloads raw audio and surfaces audio/silk media.
• Main's central STT pipeline already decodes .silk → WAV: f50a7c307a fix(stt): preprocess .silk voice notes before transcription landed in tools/transcription_tools.py::_prepare_audio_for_transcription (lazy pilk==0.2.4 via stt.silk dep, graceful error if pilk missing). This covers the same conversion for every adapter (weixin AND qqbot) at the correct layer — the transcription boundary — instead of duplicating it per-adapter.
• The PR's second intent (don't trust voice_item.text) is also already on main via cb1798b43c + regression cf6ff8cd30 (TestWeixinVoiceGatewayHandoff).

Evidence: tests/gateway/test_weixin.py -k Voice 7 passed; tests/tools/test_transcription_tools.py -k silk 1 passed, on main pin.

Recommendation: close this PR as superseded (credit to @annguyenNous for identifying the silk problem — the approach was adopted centrally). No duplicate PR will be opened.

[andrexibiza]

Coordination: same-defect sibling

This PR fixes #42084 — fix(weixin): convert Silk voice messages to WAV via pilk for STT. The same defect was also reported as #32196 (Weixin(WeChat) voice messages in Silk format skip STT — russian transcription broken), which is not in this PR's close set.

Please add Closes #32196 to the PR body so the duplicate closes with the fix. The interlock keeps both reports visible and the fix attached to one surface.