I think it would be better to talk about not appearing in "a generic type constraint" rather than in "a where clause", as the 'a lifetime parameters in both fn<'a, T: Trait<'a>>(t: T) and fn<'a>(t: impl Trait<'a>) are early bound without there being any where clause.

Good point!

Is this what you mean? https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=c30a35a698a82969827f8260139c0f9d
It is possible in the sense that it just generates a warning, but warns about it being an error in the future.

Apologies, only just saw your comment (the trials and tribulations of GitHub mobile!). Yes, _bar_fn2 in your playground is what I meant—a number of the tests output the error message "cannot specify lifetime arguments explicitly if late bound lifetime parameters are present", but I didn't appreciate they were only warnings (which compiletest denies/upgrades to errors); I think it'd still be helpful to make some reference to it?

That _bar_fn3 in your playground compiles is completely consistent with your existing explanation: no attempt was made to (erroneously) specify concrete lifetimes for late-bound parameters in a function reference, nor was the function reference cast to a pointer that would require explicit universal quantification; lifetime parameter 'b is early-bound to the lifetime parameter of the second argument, whatever that ends up happening to be—it is existentially, not universally, quantified.