Some questions;

• Do we know if this impacts older versions as well? (if so, we should add cherry-pick labels for those probably)
• I'm guessing "it's complicated", but do we have any tests on Windows that restart the daemon? (I know testing on windows is more complicated, because we can't spin up daemons "per test", so not sure if we can easily test, but if there's an existing test that does a restart, perhaps we should have one for this as well).
• Are there other custom options besides "Labels" and "Options" for Windows networks? (mostly considering; would we have some way to catch regressions if there would ever be other options?)

• Do we know if this impacts older versions as well? (if so, we should add cherry-pick labels for those probably)

Yes, all older versions I think ... I added a cherry-pick label for 27.x, I'll add 25.0 and 26.1 too - although I'm not sure we'll want to update those.

• I'm guessing "it's complicated", but do we have any tests on Windows that restart the daemon? (I know testing on windows is more complicated, because we can't spin up daemons "per test", so not sure if we can easily test, but if there's an existing test that does a restart, perhaps we should have one for this as well).

Yes, it's a wider problem ... the test suites start their own daemon, and it's not currently possible to start another on the same host as we do for Linux. Perhaps we should have a Windows test suite that doesn't start its own daemon, so that tests can manage their own - but that's out of scope here.

• Are there other custom options besides "Labels" and "Options" for Windows networks? (mostly considering; would we have some way to catch regressions if there would ever be other options?)

Nothing jumped out at me but yes, there may be other things that get lost when these networks are re-created. If we could write an integration test for it, comparing before/after inspect outputs would be good - although it'd only catch missing attributes populated by the test. (I'm not sure how we'd write a test that could catch dropped options that don't exist yet.)

I'll go ahead and merge this ... it fixes the reported problem, and we should look at improvements to regression testing infrastructure separately.

Thanks! Yes, definitely fine to go otherwise. I also can't come up with an easy way to address the Windows testing situation; not having DIND available as option, and starting multiple daemons being less than trivial for sure makes some things more complicated.