Thinking about it. I feel like it should be fine just dropping the udevadm control --reload from the genrator binary. There shouldn't be any race conditions. Any .rules or .link files generated are written at systemd's generator stage. The systemd-udevd.service starts only afterwards and reads all available rules, before listening for kernel uevens (or triggering pending uevents). So we shouldn't miss any rule exection on first-boot/reboot.

At runtime, we can do whatever we need, through the Python CLI.

I'm pondering about adopting the autostart autopkgtest, which does a reboot test, adding some udev settings (like offloading flags) to the physical eth0 en* management interface of the test runner (ugh, we shouldn't fiddle with that, but it's the only physical interface we have).

We've used this in the Ubuntu devel series for a while now and didn't observe any fallout.
I feel relatively confident this works as expected and think it should be merged.

@daniloegea WDYT?

It seems to be working fine.

The only thing I've found so far, that might not be related to this change, is that the integration test TestNetworkManager.test_remove_virtual_interfaces from the "scenarios" is consistently failing on ppc64el in my backport of "1.0.1-1ubuntu2" to Noble, which includes this change. I'm still testing it to be sure this change is not causing problems.