Filter Whisper hallucinations (closes #5)

[kenodressel]

Summary

Closes #5.

Whisper produces canned phrases ("Thank you.", "Untertitel im Auftrag des ZDF") and consecutive-token/n-gram repetition on noise, silence, or short audio. This change drops those after transcription and shows a brief ⚠️ "Likely noise" badge instead of pasting garbage.

• Post-Whisper, two detectors in sabbel/hallucinations.py: token/n-gram repetition heuristic + normalized exact-match against a 22-phrase default phantom set (English + German + bracketed music/applause + dots/notes).
• Users can extend the phantom set via a new [hallucinations] section in ~/.config/sabbel/dictionary.toml — hot-reloaded between recordings.
• Filter runs between the empty-text check and apply_replacements so phantom matches aren't broken by dictionary substitutions.

Spec: docs/superpowers/specs/2026-05-13-hallucination-filtering-design.md

Test plan

• Unit suite: uv run pytest — 113 passing (39 new tests across test_hallucinations.py + test_dictionary.py)
• Manual on hardware (after make install-app):
  • Hold hotkey in a silent room → ⚠️ "Likely noise" badge, no paste
  • Speak a normal sentence → normal dictation, no filter trigger
  • Add a custom phrase to ~/.config/sabbel/dictionary.toml under [hallucinations] → picked up on next recording without restart
  • tail /tmp/sabbel-runtime.log after a noise rejection → shows Filtered hallucination: ...

Known limitations

• Emphatic "nein nein nein nein" or "ja ja ja ja" (4+ consecutive identical tokens) is filtered. Accepted false positive — emphatic 4-repeat speech is rare in dictation.
• Single-word "you" and "Bye." are dropped (Whisper noise-bias phantoms). Acceptable for dictation; users dictating one-word messages should add a custom replacement if it bites.
• 5+-gram repetition (e.g., "a b c d e a b c d e a b c d e") is not flagged. Spec caps n at 4 — extremely unlikely in practice.

🤖 Generated with Claude Code