sum() mishandles a float subclass when accumulating into a complex total

[eendebakpt]

Bug report

Bug description:

The complex-total fast path in builtin_sum_impl (Python/bltinmodule.c:3049) guards with PyFloat_Check(item) (accepts subclasses) but frees the item with _Py_DECREF_SPECIALIZED(item, _PyFloat_ExactDealloc)`, which assumes an exact float.

Reproducer:

class F(float):
    def __del__(self):
        print("__del__ ran")

sum((F(x) for x in (2.0, 3.0)), 1j)   # generator so refcount hits 0

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-150868: Fix sum() deallocating a float subclass as an exact float #151063