Two corrections, both from the box.
1. The previous commit made the between-session clean unconditional on the
grounds that /restart-device is idempotent and costs 0.07 s. It is not free:
after ~6 restarts in one afternoon .173 put the devnode into RESTART PENDING,
and pnputil then refuses every further attempt -- 'a system restart is
pending for this device to complete a previous operation' -- until an actual
reboot. Restarting speculatively on every capture-mode connect would burn the
only lever we have. So gate it on an outstanding declare again.
2. That failure was invisible. The script discarded pnputil's output and
reported a bare exit code, so three separate runs looked like a wiring bug
(the clean 'not firing') when in fact it ran every time and the restart
failed. Keep the text, and surface a restart_pending field so the one
failure that no retry can fix is named in the log.
Also corrects the previous commit's message: CURSOR_DECLARED was never the
problem. Both host processes run as SYSTEM and serve/handshake/capture share
one process, so the flag was set and read correctly all along.
enricobuehler
df74dd5aee