feat(conductor): add opt-in voice STT to Telegram bridge

[Abeansits]

Summary

• Opt-in voice STT: Telegram voice messages are transcribed locally using parakeet-mlx via a subprocess worker (stt_worker.py), crash-isolated from the bot event loop. No cloud API needed.
• Disabled by default: Voice transcription only activates when BRIDGE_STT_ENABLED=true env var is set. Without it, voice messages are silently ignored.
• No hardcoded paths: stt_worker.py uses shutil.which('parakeet-mlx') to find the CLI on PATH, with PARAKEET_CLI_PATH env var as an explicit override.
• File logging: Bridge now logs to ~/.agent-deck/conductor/bridge.log in addition to stdout.

Changes

• conductor/bridge.py: transcribe_voice() downloads voice files and calls stt_worker via async subprocess (60s timeout). handle_message() now handles message.voice when BRIDGE_STT_ENABLED=true.
• conductor/stt_worker.py: New standalone STT worker that finds and invokes the parakeet-mlx CLI, reads output text files, and prints transcription to stdout.

Configuration

# Enable voice transcription
export BRIDGE_STT_ENABLED=true

# Optional: explicit path to parakeet-mlx CLI
export PARAKEET_CLI_PATH=/path/to/parakeet-mlx

Test plan

• With BRIDGE_STT_ENABLED=true, send a voice message via Telegram and verify transcription
• Verify voice messages produce a Transcribing... status then the transcribed text is forwarded to the conductor
• Verify failed transcription returns [Could not transcribe voice message.]
• With STT disabled (default), verify voice messages are silently ignored
• Verify normal text message handling is unaffected

[asheshgoplani]

Interesting voice transcription feature! A few issues to address:

1. Hardcoded ffmpeg path: The code uses a hardcoded path for ffmpeg instead of using shutil.which('ffmpeg') as described in the PR description. Please use dynamic path resolution so it works across different systems and installation methods (e.g., Homebrew on macOS puts it in /opt/homebrew/bin/, Linux distros vary).

2. Error handling: The voice transcription path needs proper error handling for common failure cases: ffmpeg not installed, microphone not available, transcription API timeout, invalid audio format. Users should get clear error messages rather than cryptic tracebacks.

3. Needs rebase: This PR is based on the stale v0.27.0 base. Main was rolled back to v0.26.4 after the Go 1.25 incident, so this needs a rebase onto current main.

Please fix the hardcoded path issue and add error handling, then rebase onto current main.