Cache verified model files across browser visits

Preserve full transcripts across overlapping speech chunks

Clean private browser preview