udev: Keep the systemd tag across change events

[fildunsky]

Fixes #163.

The udev rule only matches ACTION=="add". udev rebuilds a device's properties from the rules on every event, so a later change event on the hidraw node drops both TAG+="systemd" and SYSTEMD_WANTS. The device unit goes inactive and iptsd is stopped while /dev/hidrawN is still there. Nothing starts it again until the device is re-added.

This PR matches every action except remove instead, so both are re-applied on change. Real removal behaves as before.

How it shows up

Surface Pro 8 (ITHC, 045E:0C37), Ubuntu 26.04, iptsd 3.1.0. On Ubuntu, snapd reloads udev rules and triggers change on every device whenever a snap is installed. Three times in the journal, installing a snap (obs-studio, kf6-core24, inkscape) stopped iptsd within 1–2 seconds:

systemd-udevd[368]: Configuration file ... (rules reload)
boltd: udev: device changed: authorized -> authorized
systemd[1]: Stopping iptsd@dev-hidraw0.service - Intel Precise Touch & Stylus Daemon...
systemd[1]: iptsd@dev-hidraw0.service: Deactivated successfully.

Afterwards the node is present and ITHC is still bound, but:

$ udevadm info /dev/hidraw0 | grep -E 'CURRENT_TAGS|SYSTEMD_WANTS'
E: CURRENT_TAGS=:seat:
$ systemctl show dev-hidraw0.device -p ActiveState -p Wants
Wants=
ActiveState=inactive

Once, touch and pen stayed dead for over a day until a reboot.

Tested

With this rule as a local override in /etc/udev/rules.d/50-iptsd.rules:

$ sudo udevadm trigger --action=add /dev/hidraw0       # brings the stopped daemon back
$ sudo udevadm trigger --action=change /dev/hidraw0    # what snapd does
$ systemctl is-active iptsd@dev-hidraw0.service
active
$ udevadm info /dev/hidraw0 | grep -E 'CURRENT_TAGS|SYSTEMD_WANTS'
E: SYSTEMD_WANTS=iptsd@dev-hidraw0.service
E: CURRENT_TAGS=:systemd:seat:

The daemon keeps the same PID across the change event; nothing is restarted. iptsd-check-device does run again on each change, which is a read-only descriptor check.

I only reproduced the failure through snap installs, not with a bare udevadm trigger on the unfixed rule. The journal shows the same change event either way, though.

Relation to open PRs

• iptsd@.service stops itself ~6s after start on cold boot due to StopWhenUnneeded=yes #219 removes StopWhenUnneeded=yes, for a stop about 6 s after a cold boot with no udev event involved. That looks like a separate trigger. Here the device unit itself goes inactive after change, so I'd expect BindsTo= to stop the service even with iptsd@.service stops itself ~6s after start on cold boot due to StopWhenUnneeded=yes #219 applied, but I have not tested that combination. The two PRs touch different files.
• udev: hide raw ithc input nodes duplicated by iptsd #221 appends a new rule to the same file; this one only changes the first line, so they don't overlap.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KLPZcjA8vscVEpwGEMZAmK