Voice Message Handling

Read audio messages from their spoken content. A Slack voice note arrives as
an empty-text message carrying one hosted audio/m4a file. Route in this
order:

1. Native audio input (preferred)

A conversation model that declares audio input hears the clip directly:
wording and prosody reach the reasoning turn in one request. Backend lineups
change, so take the deployment's declared modality over memory.

2. On-box decode

scripts/decode_audio.py turns any audio file into
the transcript on stdout: sherpa-onnx Qwen3-ASR 0.6B int8 + Silero VAD on
CPU, models cached under <charliebot_home>/models with pinned sha256 (GPU
hosts select voice.engine=qwen3_hf). Run
uv run --with av python3 <skill-dir>/scripts/decode_audio.py <file>;
8.8 s of speech decodes in ~2.7 s on the reference CPU.

3. Outside the checkout

A host without the repo decodes with uvx --with faster-whisper python3
(small, CPU). On Chinese clips the on-box Qwen3-ASR reads English
loanwords better ("Skill" versus "Scale") and stays the reference decode
where the checkout is present.

Slack specifics

Fetch the hosted file through url_private_download with the app bearer;
the on-box decode returns in seconds while Slack's files.info
transcription lags by minutes. Voice content is PII: redact before quoting
outside the thread.