opaque auto trait leakage in coherence

[lcnr]

we should add tests that we don't leak auto traits of opaques in coherence and update the impl to make sure that we don't

[compiler-errors]

This can never happen currently because we always hit ambiguity when normalizing a TAIT during candidate assembly. We probably should protect against it regardless, but I don't think we have any tests that are unsound currently.

I couldn't think of a more creative test than this tho:

#![feature(type_alias_impl_trait)]

mod foo {
    type Tait = impl Sized;

    pub struct Wrapper(Tait);

    fn test() -> Tait { std::ptr::null_mut::<()>() }
}

trait Foo {}

impl<T: Send> Foo for T {}

impl Foo for foo::Wrapper {}
//~^ ERROR conflict implementations of trait `Foo` for type `Wrapper`

[lcnr]

closing due to rust-lang/rust#116689