Add Linux STT support + harden LLM normalization output

[F0urO4]

Summary

The STT path was macOS-only. listInputDevices() used system_profiler and startRecording() spawned sox with -t coreaudio. On Linux, /stt-mic returned "No input devices found" and recording produced an empty WAV. This PR adds Linux support and fixes two LLM normalization issues I hit while testing on Pop!_OS 24.04 (PipeWire).

Changes

1. Linux STT support (feat)

• listInputDevices(): branch on process.platform. On Linux, use pactl list short sources and filter out .monitor sinks. Works with both PulseAudio and PipeWire (PipeWire exposes PulseAudio-compatible sources).
• startRecording(): use -t pulseaudio on Linux instead of -t coreaudio.
• Drop the sox silence 1 0.1 1% effect on Linux: it gates recording start on input exceeding 1% and on PulseAudio often yields an empty WAV because the gate never triggers before the user speaks. Push-to-talk (ctrl+r toggle) does not need auto-start gating, so the effect is kept only on macOS where it was originally tuned.
• README: document Linux install (apt/dnf + whisper.cpp build from source) and add a CPU performance note pointing CPU-only users at base.en instead of large-v3-turbo (large-v3-turbo transcribes a 3s clip in ~36s on CPU vs ~2s for base.en).

2. LLM normalization output hardening (fix)

Small local LLMs (e.g. llama3.2:3b via Ollama) often wrap their output in preambles like "Here is the cleaned-up transcription:" and surround the text in quotes, despite the system prompt instructing output-only. The user would then see the wrapper text appear in their OpenCode prompt instead of just the cleaned transcription.

• cleanLLMOutput(): strips common preambles, surrounding quotes, and trailing notes from LLM output. Exported and unit-tested.
• normalizeTranscription(): filter out auto-generated session titles ("New session - <timestamp>") before passing them to the LLM as context. Small LLMs sometimes fold these literally into the output ("This is a new session: 2026-08-19T18:58:56.307Z") instead of using them to resolve ambiguous references. Only meaningful titles are now passed, and the prompt is clearer about not including them in output.
• Strengthen the user prompt: "Output ONLY the cleaned text, no quotes, no preamble".

Testing

• npm run check — 0 lint errors, formatting clean (oxlint + oxfmt)
• npm test — 12/12 tests pass (10 existing + 2 new for cleanLLMOutput)
• Manual end-to-end test on Pop!_OS 24.04 (PipeWire) with a JLab USB microphone:
  • /stt-mic correctly lists both built-in and USB microphones
  • ctrl+r → speak → ctrl+r records, transcribes with base.en, normalizes via Ollama, and appends clean text to the OpenCode prompt
  • LLM output no longer contains preambles or quoted wrappers

Notes

• No macOS behavior changes — all Linux code paths are behind process.platform === "darwin" checks.
• No new dependencies.
• The silence effect change is the only behavioral difference on macOS: none — it's still applied on Darwin exactly as before.