`dyn for<'a>` lifetime inference does not work within typedefs, but works elsewhere

I tried this code:

pub trait Foo<'a> { type Ty: 'a; }

pub fn i_work<'x>(x: &'x [u8]) {
  // can take an outside lifetime parameter just fine
  let _: <dyn for<'a> Foo<'a, Ty = &'x [u8]> as Foo<'_>>::Ty = x;
  // can be infered
  let _: <dyn for<'a> Foo<'a, Ty = &[u8]> as Foo<'_>>::Ty = x;
  // can be explicit
  let _: <dyn for<'a> Foo<'a, Ty = &'a [u8]> as Foo<'_>>::Ty = x;
  // infers a lifetime just fine (which is probably 'static)
  let _: <dyn for<'a> Foo<'a, Ty = &[u8]> as Foo<'_>>::Ty = &[1, 2, 3];
}

// Fails:
// type IDontWork = Vec<<dyn for<'a> Foo<'a, Ty = &[u8]> as Foo<'static>>::Ty>;
pub type IWork = Vec<<dyn for<'a> Foo<'a, Ty = &'a [u8]> as Foo<'static>>::Ty>;

I expected to see this happen: Code compiles successfully because the lifetime parameter for &[u8] should be inferred as there is only one lifetime in scope, and that lifetime comes from a binder rather than the typedef.

Instead, this happened: Code does not compile. rustc complains about needing an explicit lifetime parameter.

Meta

rustc --version --verbose: 1.94.0 stable (tested in playground)

Start by compiling the minimal reproducer in the issue, comparing the working expression forms with the failing typedef IDontWork and the explicit-lifetime IWork alias. Trace the compiler's lifetime inference for associated types under the dyn for<'a> binder. Done means the typedef form accepts the inferred lifetime and a regression test covers both the failing and working cases.