Class creation doesn't call __set_name__ when creating class members in another __set_name__

[AJMansfield]

Bug report

Bug description:

When inserting a class member during some other class member's __set_name__ call, the __set_name__ method of the newly-inserted member doesn't get called as expected.

As an example, take this snippet:

class I_Am_Altering_The_Deal:
    def __set_name__(self, owner, name):
        print(f"__set_name__({self.__class__.__name__}, {owner.__name__}, {name!r}")
        owner.han_solo = Pray_I_Dont_Alter_It_Any_Further()

class Pray_I_Dont_Alter_It_Any_Further:
    def __set_name__(self, owner, name):
        print(f"__set_name__({self.__class__.__name__}, {owner.__name__}, {name!r}")

class Cloud_City:
    luke_skywalker = I_Am_Altering_The_Deal()

I would expect both __set_name__ functions to be called, resulting in the output:

__set_name__(I_Am_Altering_The_Deal, Cloud_City, 'luke_skywalker')
__set_name__(Pray_I_Dont_Alter_It_Any_Further, Cloud_City, 'han_solo')

But what is actually output, is:

__set_name__(I_Am_Altering_The_Deal, Cloud_City, 'luke_skywalker')

Discovered while testing a PEP-487 implementation for Micropython: micropython/micropython#15503.

Maybe be related to #72983.

CPython versions tested on:

3.8, 3.10, 3.12

Operating systems tested on:

Linux

[serhiy-storchaka]

__set_name__() is only called for values set in the class body. Direct setting the class attribute does not trigger calling __set_name__(). If you need to call it, do it explicitly.

[blhsing]

@serhiy-storchaka The documentation says that __set_name__ is "Automatically called at the time the owning class owner is created", and says nothing about the attribute having to be set in the class body. And from a user's perspective the attribute han_solo is set while Cloud_City is still being created, so it may be reasonable for the OP to expect owner.han_solo.__set_name__ to be called.

I guess we may have to either clarify the existing behavior in the documentation, or implement type_new_set_names with a while loop that looks for any new names in __dict__ after calls to __set_name__s, perhaps by just comparing the size of __dict__ before and after the calls.

[serhiy-storchaka]

I think we should clarify the documentation. Cc @ncoghlan

[AJMansfield]

perhaps by just comparing the size of __dict__ before and after the calls.

I don't think the size of __dict__ is specific enough, since a __set_name__ also has the potential to delete members from a class. But, type_new_set_names could set some sort of "class creation set-names phase" flag, and then rely on type_new_set_attrs to append newly added names to a list or even invoke __set_name__ directly. (Invoking it directly would also have the benefit of appropriately failing on code that reaches a recursion depth limit, instead of needing explicit code to fail out on an infinite loop in the list version.)

I think we should clarify the documentation. Cc @ncoghlan

Documenting and adding tests that that it's specifically not called on members inserted this would be a perfectly acceptable resolution for me -- to be honest, my main interest here is just to determining what the official correct behavior is, so I can implement this feature in Micropython with the same semantics as CPython.