generalizing projections is sus

[lcnr]

when generalizing a projection we have to be careful of 2 things, using ?ident.universe for inference variables in a given universe.

?x sub <() as Trait<?y>>::Assoc

must not generalize ?x to <() as Trait<?y>>::Assoc as ?x may get inferred to a subtype of the normalized projection. Otherwise this would be incomplete. Should emit AliasRelate in this case instead.

?x.0 eq <() as Trait<?y.1>>::Assoc

This must not generalize ?x.0 to <() as Trait<?z.1>>::Assoc because that should result in a universe error but it must also not relate it to <() as Trait<?z.0>>::Assoc because the projection may not actually normalize to a type mentioning z so if we then equate the substs in AliasRelate we end up pulling down the universe of ?y to 0 as well which would be incomplete. Should instead emit AliasRelate(?x.0, <() as Trait<?y.1>>::Assoc, Equate).

[lcnr]

closed in favor of rust-lang/trait-system-refactor-initiative#8