Confusing "no field found" diagnostic

struct S {
    field: (),
}
fn f(_: impl FnOnce(S)) {}
fn main() {
    let cb = |s| _ = s.field;
    f(cb)
}

Here we report a "no field field on type S" diagnostic on the field access. Now this diagnostic is correct (rustc will fail here as well), the issue is that by the time we try to resolve the field we don't actually yet know what type s has, we only learn about this fact later on when the closure is used and we resolve all outstanding type variables once at the end of inference. This is where we then resolve the unknown type to S within the diagnostic which causes this confusing output. We should prevent the last part from happening.

Start by reproducing the supplied closure example and tracing field-access diagnostic handling through closure inference. Focus on where outstanding type variables are resolved for the diagnostic; done means the diagnostic no longer substitutes the later-resolved S type while the field access remains correctly rejected.