Hi Petr,

yes I have that what generated the wrong thing, but it is inside
a specific PySide repository in a big project. Before I try to extract
and simulate that, let me just show it.

All types which had been no heaptypes before were already quite complicated
types which had their own deallocators before, or a 0 if they did not
need one, or better "should not have one".

When switching to the new interface, those deallocators which were 0 were
turned into the default deallocator, which crashed most of the time.

By replacing it by a dummy function fixed the problem. Example:

static PyType_Slot PySideSignalMetaType_slots[] = {
    {Py_tp_methods, (void *)Signal_methods},
    {Py_tp_base, (void *)&PyType_Type},
    {Py_tp_free, (void *)PyObject_GC_Del},
    {Py_tp_dealloc, (void *)SbkDummyDealloc},
    {0, 0}
};
static PyType_Spec PySideSignalMetaType_spec = {
    "PySide2.QtCore.MetaSignal",
    0,
    // sizeof(PyHeapTypeObject) is filled in by PyType_FromSpecWithBases
    // which calls PyType_Ready which calls inherit_special.
    0,
    Py_TPFLAGS_DEFAULT,
    PySideSignalMetaType_slots,
};

You can find the checkins here. This branch contains everything relevant in small steps.

https://github.com/pyside/pyside2-setup/commit/2f0658be36107097810985f2190fe0f2acfba178

The first working transformation was in the commit
"PEP 384-8-HT-4: Successful Restart using PyType_FromSpec".

I have yet no idea how to extract a minimum example that produces this problem.

Summary: 
In a way, without being heaptypes, the old types were dynamic, too, and inserting
something that was not there before was a bad idea.

(sorry, there is more...)