Given dict.setdefault, a parallel method like set.add_unique, which
returns the existing element or None, might be the better approach.

Although it breaks the convention of these methods returning the item, I'm not sure returning the element is a good idea.
"set.add_unique(None)"  will return None.

The original example of "if dict.setdefault()" is problematic as well for the exact same reason. You can't tell if the default value was already there in the first place.

The set.add is already threadsafe as GIL is not checked while running the function, but it shouldn't be relied upon, if only because of ABC not guaranteeing other implementations to be threadsafe, let alone other python implementations as well.

De-facto I've never seen TOCTOU issues with these things as they don't contain external resource access and are usually surrounded by locks for plenty of other reasons, but theoretically it can be useful to prevent the need of locks.

Either way, modifying the existing method is probably a no-no - it will break backwards compatibility in both Python and C-API.

Regarding a new method, what are the actual arguments in favor? Apart from saving one line of code ofc. Besides, I don't believe trying to 1-line it makes the code any clearer, but that's personal taste.