Unsure if bug: static_property_type needs dynamic heap?

[makslevental]

Note, I'm on 2.10.4 on CPython 3.11.4.

Recently I tried to do this

pure_subclass &def_property_readonly_static(const char *name, Func &&f) {
  py::cpp_function cf(std::forward<Func>(f), py::name(name));
  auto *staticPropertyType = py::detail::get_internals().static_property_type;
  auto builtinProperty =
      py::reinterpret_borrow<py::object>((PyObject *)staticPropertyType);
  thisClass.attr(name) = builtinProperty(cf);
  return *this;
}

Roughly, this is adding a static property to a pure python class, where the getter is implemented in C++. Ignoring why one might want to do this¹, I was met with a curious error:

AttributeError: 'pybind11_static_property' object attribute '__doc__' is read-only

Digging into the implementation of make_static_property_type, I found a pretty relevant/useful comment:

#    if PY_VERSION_HEX >= 0x030C0000
    // Since Python-3.12 property-derived types are required to
    // have dynamic attributes (to set `__doc__`)
    enable_dynamic_attributes(heap_type);
#    endif

As a kind of "by hook or by crook" approach, I moved the enable_dynamic_attributes out of the #if and the AttributeError was resolved. Great! But the confusing thing is I'm not testing/using 3.12² and thus the comment suggests/implies that __doc__ doesn't need to be writeable? I tried searching around for when/where the requirement (for dynamic attributes of property-derived types in CPython) would be required but couldn't find much (but I'm not a CPython contributor so I don't really know where to look for that kind of info, other than release notes). I also couldn't quite figure out how to

My intuition is that I'm doing something wrong because conventional users of pybind and def_property_readonly_static aren't seeing this error (I assume...).

Any advice would be appreciated.

P.S.

On writing this up and going back to double check some things I've found there's some history/context here:

1. @sizmailov submitted a patch to CPython bpo-41287: Handle doc argument of property.__init__ in subclasses python/cpython#23205 and correspondingly added that very comment/#if in make_static_property_type
2. The patch was broke something and the original behavior was restored? property assumes that its subclasses have __dict__ python/cpython#98963.

Still digesting but posting anyway for posterity/reference.

cc @martin-luecke

Footnotes

1. It's our way of enabling downstream users to sign up for an object hierarchy in C++ and have that reflected in python... ↩

2. And if I were well then the #if would catch and __doc__ would be writeable. ↩