Amusingly, this is because of an old hack to make directly calling somedict.__getitem__ fast: https://github.com/python/cpython/commit/8f5cdaa784f555149adf5e94fd2e989f99d6b1db

Reverting the dictobject.c changes of the commit, i.e., the following patch, fixes your example.

diff --git a/Objects/dictobject.c b/Objects/dictobject.c
index 413557d667..d85834977d 100644
--- a/Objects/dictobject.c
+++ b/Objects/dictobject.c
@@ -3032,8 +3032,6 @@ dict_sizeof(PyDictObject *mp, PyObject *Py_UNUSED(ignored))
     return PyLong_FromSsize_t(_PyDict_SizeOf(mp));
 }
 
-PyDoc_STRVAR(getitem__doc__, "x.__getitem__(y) <==> x[y]");
-
 PyDoc_STRVAR(sizeof__doc__,
 "D.__sizeof__() -> size of D in memory, in bytes");
 
@@ -3071,8 +3069,6 @@ PyDoc_STRVAR(values__doc__,
 
 static PyMethodDef mapp_methods[] = {
     DICT___CONTAINS___METHODDEF
-    {"__getitem__", (PyCFunction)dict_subscript,        METH_O | METH_COEXIST,
-     getitem__doc__},
     {"__sizeof__",      (PyCFunction)dict_sizeof,       METH_NOARGS,
      sizeof__doc__},
     DICT_GET_METHODDEF