The problem appears more likely to be that $^E [a.k.a. Win32::GetLastError()] is being reset more quickly in your environment than in most. A better fix is to have Win32::TieRegistry::_ErrNum() stop using Win32::GetLastError() and start using Win32API::Registry::_regLastError() [and either just require a modern Win32API::Registry or make that change conditional on that function being available].

Please report the bug over at rt://Win32::TieRegistry for the world to see. Would you mind also trying the change I suggested above in your environment (with your fix removed, of course) so I'll have verification that my guess is correct?

I really need to release some module updates. I wish I didn't need a "day job". :)

Update: Oh, and nice work on tracking down details about the bug.