This PR hasn't gotten much conversation yet, so I guess I'll stop waiting for someone else to ask smart questions and instead start asking stupid questions.

Why do we want this? Is it just to make the implementation of #4235 easier or simpler, or does it fix an independent issue?

What exactly is happening here? The changelog indicates that the inferred type of Nothing goes from forall a1. Maybe a1 to forall a. Maybe a, which from a naive reading sure looks like an alpha conversion. I suppose at the level of types in the compiler instead of types in the source language, I'm meant to know that a1 is a skolem and a is a type variable. Even so, how are those types handled differently?

Are there downsides or is there an argument that this is unambiguously what we should have been doing all along?

The issue in the VTA context was briefly mentioned here:
#4235 (comment)

The issue is that the re-generalization would discard the VTA information, even though the constructors supported VTA by definition.

Sure, but would we still want this change in the (probably counterfactual) world where we decide not to implement VTA? If so, why; if not, what consequences does this have for things that are not related to VTA?

Outside of #4235, this effectively just removes some historic baggage introduced by #835 and reworks a part of #2625 to accommodate for constructors. I don't think it ever came up in discussion back then but it's also just nice having constructors infer w/o the numeric suffixes 😄