Fix Groq hallucinations around long silence

[NathanSkene]

Summary

• replace the failed feat: add VAD-based silence compression for efficient transcription #504 approach with a Groq-specific long-silence fix
• gate full-clip silence in the recorder before cloud transcription
• trim long leading/trailing silence before Groq upload
• suppress custom dictionary prompt bias for long-leading-silence Groq clips
• use Groq verbose_json segment metadata to reject likely silence hallucinations

Why

Manual validation showed #504 was not usable: long silence still produced hallucinated text. This replacement was tuned against real failing clips and validated in the installed app.

Validated behavior

• pure silence for ~10s -> blank, no hallucination
• ~10s silence then short phrase -> only the spoken phrase
• short phrase then ~10s silence -> only the spoken phrase
• ambient tapping / room noise without speech -> blank
• custom-dictionary term after silence still transcribes
• normal dictation remains intact

Notes

• this supersedes the direction in closed feat: add VAD-based silence compression for efficient transcription #504
• discussed on Whisper hallucinations ("Thank you for watching!") during silence — needs smarter VAD for cloud backends #462