Thanks! This makes the feature clearer. It also makes it clearer how good an ides the managed dict is, though I'd still prefer if it was communicated earlier/wider, to catch more edge cases.

One thing I still don't understand: why should tp_offset be set to -size - 3*sizeof(PyObject*)?. If it doesn't point to a dict, how is this useful?

I'm skipping news as this doesn't add any features or actually fix anything.

It does change tp_dictoffset for a bunch of classes, doesn't it?

But most importantly HeapCTypeWithManagedDict as not a good example of how to write a class with Py_TPFLAGS_MANAGED_DICT. It'll cause segfaults when subclassed, as in:

    def test_heaptype_with_managed_dict_subclass(self):
        class C(_testcapi.HeapCTypeWithManagedDict):
            pass
        inst = C()
        inst.foo = inst
        # (run GC)

If it's meant as a bad example, could you name it more ominously?

HeapCTypeWithManagedDict isn't supposed to be flawed.
Subclassing is problematic (#95589), but I'll try to make it more robust.

Given that #92678 has been closed, I'm closing this.
I'll add the ability to handle third-party code setting the Py_TPFLAGS_MANAGED_DICT in a new PR for 3.12.