Runs rpicam-vid under MuonTrap, behind an Erlang port.
Everything BB.Camera.FFmpeg.Backend.MuonTrap says about why MuonTrap is
here holds for rpicam-vid too: it cannot outlive its owner or the VM, no
pid is ever signalled from outside, and its output is flow-controlled. Only
open/3 differs, and acknowledging, stopping and closing are delegated to
it unchanged.
Why not BB.Camera.FFmpeg.Backend.MuonTrap itself
That backend refuses any argument vector without -nostdin, because ffmpeg
reads its stdin for keyboard commands and would steal MuonTrap's
acknowledgements. rpicam-vid has no such flag - it would refuse it as
unknown - and has no need of one: it reads stdin only when started with
--keypress, which this driver never passes. Measured on a Pi 5 through a
16 KiB :stdio_window, 170 KB of frames arrived, which is only possible if
every acknowledgement reached MuonTrap.
libcamera's logging
libcamera logs at INFO to stderr by default: half a dozen lines on every
capture start. rpicam-vid --verbose 0 quietens rpicam-vid but not
libcamera, so this sets LIBCAMERA_LOG_LEVELS to :log_levels, by default
"*:WARN". stderr itself is left attached to the VM's, as with ffmpeg: a
busy camera or a failed buffer allocation is explained there, and nowhere
else.