With the fact that the existence of "Python without closures" predates Python 2.2, this now reads like a straight up compiler bug to me.

Compare the behaviour with no local assignment in the class body:

>>> def f():
...     n = 1
...     class C:
...         print(n)
... 
>>> f()
1

To the behaviour once the local is assigned:

>>> def f():
...     n = 1
...     class C:
...         n = n
...         print(n)
... 
>>> f()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 3, in f
  File "<stdin>", line 4, in C
NameError: name 'n' is not defined

The issue is that the latter will still emit a LOAD_NAME/STORE_NAME pair, while the former emits LOAD_CLASSDEREF (and has emitted LOAD_DEREF for as long as I can recall hacking on the compiler).

Off the top of my head, I'm not sure our current symbol table analysis pass can actually cope with this idea though - it would require separating "just a class body local, which may or may not first be retrieved as a global or builtin" from "a class body local which is first retrieved from a closure reference".