Why functions shouldn't assume that their return type is WF

fn.rs

struct OnlySized<T: Sized>(std::marker::PhantomData<T>);
 
impl<T: Sized> OnlySized<T> {
    fn new() -> Self {
        println!("{}", std::mem::size_of::<T>());
        OnlySized(std::marker::PhantomData)
    }
}

trait Bar { }

fn foo() -> OnlySized<dyn Bar> {
    OnlySized<dyn Bar>::new()
}

fn bar<T>(f: fn() -> T) {
    let _ = f();
}

fn main() {
    bar(foo);
}

// Comparison:
// defining a function `fn foo() -> OnlySized<dyn Bar>` is like defining a new type and then
// ```
// impl Fn<()> for foo {
//     // here we should be proving that the assoc type value is WF, which is not the case
//     type Output = OnlySized<dyn Bar>;
// }
// ```