fix(transcription): suppress whisper non-speech tokens

[maubrowncow]

Summary

Tell the Whisper decoder not to emit bracketed sound labels like [cough], [Music], [applause], [♪♪♪], [inaudible] in the first place — by flipping suppress_blank and suppress_nst on whisper.cpp params. Add a backstop cleanup prompt rule for the WhisperKit path on Apple Silicon, which doesn't expose the same single flag.

Motivation

Whisper sometimes emits environmental sound labels into transcripts. The whisper.cpp library has built-in decoder flags (suppress_blank, suppress_nst) designed for exactly this — they bias the decoder against the specific token IDs used to denote non-speech events, without affecting regular spoken words like "music" or "cough".

Changes

File	Change
native-core/windows/src/model_whisper_cpp.rs	Set set_suppress_blank(true) and set_suppress_nst(true) on whisper-rs FullParams
native-core/macos/Sources/WhisperCppBridge/WhisperCppBridge.swift	Set params.suppress_blank = true and params.suppress_nst = true on whisper.cpp whisper_full_params
native-core/shared/prompts/cleanup-default-instructions.md	Add rule 4: drop bracketed non-speech annotations during cleanup (backstop for the WhisperKit path)

Coverage

Backend	Used on	Fix
whisper-rs	Windows	Decoder flag ✅
whisper.cpp bridge	macOS Intel / x86_64	Decoder flag ✅
WhisperKit	macOS Apple Silicon	Prompt-based cleanup (decoder flag not exposed as a single boolean)

WhisperKit's DecodingOptions.supressTokens takes a [Int] of token IDs and the library has a // TODO: implement these as default comment for the non-speech list. Reimplementing OpenAI's non_speech_tokens() against WhisperKit's tokenizer is ~80 lines accessing internal APIs and is fragile across WhisperKit upgrades, so the prompt backstop is the better trade until WhisperKit ships the default.

Caveats

• Apple Silicon users with cleanup disabled may still see [cough] / [Music]. Cleanup is enabled by default.
• suppress_nst is a probability bias, not a hard ban — extremely rare leakage is still possible.

Test plan

• Manual on Apple Silicon: dictate with cleanup enabled near a noise source, confirm [cough]/[Music] no longer appear.
• Manual on Windows or Intel macOS: same, with cleanup disabled, confirm decoder-level suppression works.
• Regression: dictate text legitimately containing the words "music" / "applause" / "cough" and confirm they still transcribe normally.

Privacy / Security

No new network calls, no telemetry. Changes are local decoder configuration plus a prompt string.

🤖 Generated with Claude Code