feat: LLM post-processing for transcriptions

[charlesdubt]

Summary

• Adds an optional LLM post-processing step that runs after transcription to fix typos, remove filler words, correct grammar, and reformat text
• Uses any OpenAI-compatible API — defaults to Ollama (localhost:11434) for fully local, offline processing with no API key
• New settings section with toggle, endpoint/model config, and a user-editable system prompt
• Graceful fallback: if the LLM call fails, the original transcription is used unchanged

How it works

1. User dictates as normal → Whisper/Parakeet transcribes
2. If post-processing is enabled, the transcription is sent to the configured LLM with the user's system prompt
3. The LLM-cleaned text is pasted into the active app

Setup (for local use with Ollama)

1. Install Ollama from ollama.com
2. Run ollama pull qwen2.5:0.5b (290 MB, fast)
3. Enable post-processing in Settings → LLM Post-Processing

Also works with OpenAI, LM Studio, or any OpenAI-compatible provider by changing the base URL and adding an API key.

Test plan

• Toggle post-processing on/off — dictation works in both states
• With Ollama running locally, dictation produces cleaned-up text
• With Ollama stopped, dictation falls back to raw transcription gracefully
• Custom system prompt changes are persisted and applied
• API key field works for cloud providers (OpenAI, etc.)
• Settings UI renders correctly in light and dark mode

🤖 Generated with Claude Code