Struct containing only associated type of generic may be too constraining variance wise

The following code is currently rejected by the compiler, because sub-typing for S is taken on the whole T and not only its associated type. Maybe this is something that could be improved?

trait TraitAssoc {
    type Assoc;
    fn assoc(self) -> Self::Assoc;
}

impl<T> TraitAssoc for &T {
    type Assoc = ();
    
    fn assoc(self) -> Self::Assoc {
        ()
    }
    
}

struct S<T: TraitAssoc> (
     T::Assoc
);

impl<T: TraitAssoc> S<T> {
    fn new(x : T) -> Self {
        Self(x.assoc())
    }
}

fn maybe_could_compile() {
    let v : Vec<u8> = vec![];
    let s = S::new(&v);
    drop(v);
    drop(s);
}

This also happens with nex-gen trait solver AFAICT.

Start by compiling the Rust reproducer in the issue and examining how the compiler applies subtyping and variance to S when its field uses only T::Assoc. Compare the behavior with the stated associated-type expectation; done means the intended behavior is decided and the example has a clear compiler outcome, with coverage added if the project identifies an appropriate test location.