our folding strategy in check_type_bound is insufficient

[compiler-errors]

rust-lang/rust#125786 folds associated type bounds to replace associated types w/ their concrete types. However, it only does it for identity-substituted associated types, and only for the associated type that we're checking.

This is known to be insufficient, see rust-lang/rust#135246. Fixing this likely relies on proving item-bounds during normalization and also affects the old solver. We will probably only fix this issue after stabilization.

---

edit: old description

This seems to be sufficient, but I'm somewhat suspicious about things like:

type Assoc<T>: Bound<Self::Assoc<u32>>;

type Assoc: Bound<Self::Other>;
type Other: ...;

We should revisit this strategy and convince ourselves it's sound before stabilization, or else make it more general (which doesn't seem too difficult).

Previous unsoundnesses:

• associated_type_defaults are unsound in the new solver #46
• Associated type bounds can be used to prove themselves with trivial where-clauses #62

[lcnr]

no :3 rust-lang/rust#135011 (comment)

[lcnr]

structurally folding instead of using ordinary normalization also avoids proving the trait assoc-type where-clauses when checking that the impl definition is wf, causing tests/ui/generic-associated-types/issue-91883.rs to pass: