io.BufferedReader.read(n) is documented to return "up to" n bytes, and
a subprocess pipe will return a short read whenever the producer closes
the pipe mid-frame — even on a blocking pipe. The existing assignment
frame_buffer[:] = ffmpeg_process.stdout.read(frame_size)
then raises ValueError on the slice assignment, and the surrounding
``except`` only breaks out of the capture loop when ``poll()`` reports
the process has already exited. If ffmpeg is still alive (e.g. an RTSP
source that dropped a chunk and recovered), the loop hits ``continue``
and the next read picks up data from the middle of the next frame,
putting capture out of sync with frame boundaries until the watchdog
later forces a restart.
Read into the pre-allocated buffer in a loop and treat any short read
as the end of the stream, so the watchdog can respawn ffmpeg from a
clean state rather than continuing at the wrong offset. As a side
effect, ``readinto`` writes directly into the shared-memory frame
buffer, avoiding the per-frame Python ``bytes`` allocation and copy
that the previous slice assignment performed.