> I think a simple iscoroutinefunction check will be pretty close to 100% matching what users expect.

Yes, "pretty close", but not reliable. :)  E.g. I can easily design a decorator that when applied first would break the proposed iscoroutinefunction logic.  While decorators like that aren't popular, I'd be wary about introducing a solution that can lead to hours of debugging in some weird (and maybe stupid) cases.

> Or, if we don't change the semantics, then we can still be 100% confident that if iscoroutinefunction returns true, then the user has made a mistake. (I.e., if we make this issue a warning, then it's possible we'll miss print a warning in some complicated cases, but we can be confident that all the warnings we do print are correct.)

+1 to implement a warning the way you suggest.