candidate selection for normalization and trait goals disagree

#![feature(discriminant_kind)]
use std::marker::DiscriminantKind;

fn trait_bound<T: DiscriminantKind>() {}
fn normalize<T: DiscriminantKind<Discriminant = u8>>() {}

fn foo<'a, 'b>()
where
    &'b (): DiscriminantKind<Discriminant = u8>,    
{
    trait_bound::<&'a ()>();
}

fn bar<'a, 'b>()
where
    &'b (): DiscriminantKind<Discriminant = u8>,    
{
    normalize::<&'a ()>();
}

foo compiles, bar does not:

error: lifetime may not live long enough
  --> src/lib.rs:18:5
   |
14 | fn bar<'a, 'b>()
   |        --  -- lifetime `'b` defined here
   |        |
   |        lifetime `'a` defined here
...
18 |     normalize::<&'a ()>();
   |     ^^^^^^^^^^^^^^^^^^^ requires that `'b` must outlive `'a`
   |
   = help: consider adding the following bound: `'b: 'a`

error: lifetime may not live long enough
  --> src/lib.rs:18:5
   |
14 | fn bar<'a, 'b>()
   |        --  -- lifetime `'b` defined here
   |        |
   |        lifetime `'a` defined here
...
18 |     normalize::<&'a ()>();
   |     ^^^^^^^^^^^^^^^^^^^ requires that `'a` must outlive `'b`
   |
   = help: consider adding the following bound: `'a: 'b`

help: `'b` and `'a` must be the same: replace one with the other

Candidate selection for the trait goal prefers the trivial builtin impl. Normalization instead prefers the where-bound. This is inconsistent and means that whether we use the associated items impacts whether a trait bound holds.

It impacts all trivial builtin traits with associated types, I don't think this effects stable rn as either the trait is unstable or the builtin impls only exist for unnameable types. Nominating for t-types vibeck

Start by compiling the Rust example in the issue and compare the successful foo call with the failing bar call. Investigate candidate selection for the trait goal and associated-type normalization; done means both paths handle the trivial builtin impl consistently, with a regression test covering the discrepancy.