Whisper / alternative STT models for much better speech recognition

[grajat90]

Summary

Add higher-accuracy speech-to-text backends behind the existing core.Transcriber interface (internal/core/interfaces.go), so users can opt into much better recognition than the default.

Context

• STT today is native-first: SpeechAnalyzer/SpeechTranscriber in the Swift helper (macos-helper/Sources/Helper/Capture/LiveSpeechPipeline.swift), adapted to Go via internal/stt.
• DESIGN.md/CLAUDE.md already anticipate whisper.cpp as an optional post-v1 fallback behind the same interface.

Scope

• Integrate whisper.cpp (and leave room for other models) as an alternative transcriber, selectable via internal/config like the distiller/answerer providers.
• Keep the hard invariant: audio never crosses the IPC boundary; any whisper backend that needs audio must run inside the helper (macos-helper/), emitting only Utterance text to the core.
• Model management: download/select model size, document accuracy vs latency/resource tradeoffs.
• Benchmark accuracy against the native transcriber on a fixed fixture (test/fixtures/speech.wav) so the win is measurable.

Open questions

• whisper.cpp in-helper (Swift bridging) vs a separate on-device process; both must keep audio local.
• Streaming vs chunked transcription and its effect on segmentation latency.

Related: multilingual support (separate issue) largely rides on model choice.