Currently, all types that we handle during inference need to be resolved as far
as possible at the time. It's maybe too brittle of an invariant; I need to think
how we can do this better. This should fix#484 though, I hope (if
it's the same case as I managed to reproduce).