The cost is the ffmpeg read (LUKS-decrypt + WAV parse), not the FFT or pipe
(astats-only emit was no faster). Decode the L channel only — stereo orbit
stems share the activity envelope, and that single full-SR decode now serves
both the 2s-bin envelope and the loudest-window spectral profile. Orbits decode
concurrently in a bounded thread pool (ffmpeg releases the GIL). 78-min SET take:
96s->57s; further gains hit the disk I/O wall (~190MB/s), not CPU. Low-SR was
rejected (anti-alias lowpass kills hat brilliance + mislocates the loud window).