impls with higher ranked projections are now usable

[lcnr]

trait Trait {}

trait OtherTrait {
    type Assoc<'a>;
}

impl<T: OtherTrait> Trait for (T, for<'a> fn(<T as OtherTrait>::Assoc<'a>)) {}

impl OtherTrait for u32 {
    type Assoc<'a> = &'a u32;
}

fn impls<T: Trait>() {}

fn main() {
    impls::<(u32, for<'a> fn(&'a u32))>();
}

This does not work with the old solver but will compile with the new solver. In the old solver we are unable to eagerly normalize <T as OtherTrait>::Assoc<'a> and also can't replace it with an inference variable (as it contains a bound variable), so then after constraining T the projection is still treated as rigid, even though it is now concrete.

This means that rust-lang/rust#105787 will end up being exploitable with the new solver which is not great 🤷

[lcnr]

This will also break the following example with -Ztrait-solver=next-coherence: https://rust.godbolt.org/z/zW16n87oq

trait Trait {}

trait OtherTrait {
    type Assoc<'a>;
}

impl<T: OtherTrait> Trait for (T, for<'a> fn(<T as OtherTrait>::Assoc<'a>)) {}

impl OtherTrait for u32 {
    type Assoc<'a> = &'a u32;
}

impl<'a> Trait for (u32, fn(&'a u32)) {}

The overlap is unsound, but currently not exploitable.