chore: merge master → dev/video-fer (SSE transcribe-stream)

fix: detect Git LFS pointer in FER model, auto-download real binary

feat: per-second FER timeline for live emotion panel

fix: correct FER model path and STT error propagation

fix: add required language= arg to apply_transcription_request()

debug: add /debug-inference endpoint to smoke-test local model

fix: use bfloat16 on CPU to halve memory usage (~6 GB vs ~12 GB)

refactor: restructure repo into api/ proxy/ web/ training/ docs/