So, the problem in #54940 is that the
AstConv code (which converts from HIR types to the internal types) is doing normalization as it works. This sometimes introduces region variables which then get converted to
'?0 (because it's an inference variable). I'm not sure the best fix yet. The first thing I thought is to stop normalizing in
AstConv and push the normalizations so they occur later (I think that should be fine...?) but I'm a bit afraid of subtle interactions. It's always hard to ensure everything is normalized.
Another thought I had is to leave this unfixed and push harder on lazy normalization (which we are also working on over in #wg-traits)