feat: add /add-telegram-voice skill for local whisper.cpp transcription

[vweaver]

Summary

Adds a new skill (/add-telegram-voice) that upgrades the Telegram channel with local voice message transcription using whisper.cpp. Follows the project's "skills over features" philosophy — no changes to core code.

• Channel-agnostic transcription module — transcribeAudio(Buffer): Promise<string | null>, usable by any channel
• Local whisper.cpp — no cloud API, no API key, no cost. Uses ffmpeg + whisper-cli on-device
• Depends on telegram skill, conflicts with voice-transcription (incompatible src/transcription.ts API — Baileys-coupled vs channel-agnostic)
• Voice notes arrive as [Voice: <transcript>] instead of [Voice message] placeholders

Skill contents

File	Purpose
SKILL.md	Setup: install deps, download model, apply, verify
manifest.yaml	Metadata, deps, conflicts
add/src/transcription.ts	Channel-agnostic whisper.cpp module
modify/src/channels/telegram.ts	Async voice handler with download + transcribe
modify/src/channels/telegram.test.ts	5 new voice tests + fetch/transcription mocks
tests/telegram-voice.test.ts	15 skill package validation tests

Built and tested on Linux (Ubuntu) with whisper.cpp built from source (-DBUILD_SHARED_LIBS=OFF). SKILL.md includes Linux-specific build instructions and the static linking gotcha.

Test plan

• 15 skill package tests pass (manifest, file presence, intent files, API shape, no Baileys deps)
• 54 Telegram channel tests pass (including 5 new voice transcription tests)
• Full suite passes (374 tests)
• Verified end-to-end: voice note in Telegram -> [Voice: Can you hear me now?]

🤖 Generated with Claude Code

[TomGranot]

There's already a use-local-whisper skill that landed in #702. How does this differ? If it's the same thing, we should close this one.

[vweaver]

They solve different problems — use-local-whisper swaps the OpenAI backend for whisper.cpp on WhatsApp, while this PR adds voice transcription to Telegram which has no voice support today.

That said, use-local-whisper already notes that Telegram just needs audio download logic added. Rather than a separate skill, I could rework this as an update to use-local-whisper that:

1. Adds a channel-agnostic transcribeAudio(Buffer) export to src/transcription.ts
2. Keeps the existing transcribeAudioMessage(WAMessage, WASocket) as a wrapper so WhatsApp isn't broken
3. Adds the Telegram voice handler and tests

Would you prefer that approach, or is there a different way you'd like to see it structured?

[vweaver]

Closing in favor of a new PR that adds Telegram support to the existing use-local-whisper skill instead of creating a separate skill.