harden speech transcription against unreadable audio and empty transcripts

add local speech transcription boundary

tune character behavior

try to align chracters better

implement trust gate architecture

updated zerogpu llama version

first commit, testing mvp