Doing the initialization later is not really an option as there can be warnings in the parser so it has to be resolved very early on.

I guess some form could be a possibility, but that's also tricky as subclass matching also still needs to work. I think we would need to see a patch on the proposed solution before considering its acceptance.

Another option would to parse it twice: once early on for Python internals and then again later once site-packages and various other things have been set up. Once again I would need to look at a patch before making a decision on whether it's an acceptable solution.