This leads e.g. to the following issue:

Python 3.10.6 (main, Aug  2 2022, 00:00:00) [GCC 12.1.1 20220507 (Red Hat 12.1.1-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from rapidfuzz.distance import Editop
>>> Editop.__name__ = 'test'
Segmentation fault (core dumped)

I will add a test for this later today.

@da-woods apparently we do rely on the mutability, so I am unsure how we want to fix this.

edit: as far as I can see this is a new test which was introduced as part of: #4955. Is this postinit usage something we actually support, or was this just added this way, since it did not fail?

Ah I probably didn't mean to rely on the mutability. I basically just enabled all the CPython dataclasses tests that seemed to pass. However, Cython dataclasses aren't Python dataclasses so they're allowed to differ in places. I'll revert the specific tests that rely on it.

I wasn't sure if immutability should be considered a feature or a side effect (i.e. whether there was a harm in leave things mutable). However, if you can cause segfaults with it so easily then we should definitely keep it.

The dataclass test is easily disabled. The second failing test is to do with setting Py_TPFLAGS_SEQUENCE on the memoryview wrapper class. I think that's well worth doing since it makes it behave like a sequence for pattern matching. It needs a bit of thought about how we do it with immutable classes though since it can't be done after the fact.

I think a @cython.tp_flags decorator might be useful (taking either "sequence" or "mapping") - I could imagine users might want to register their own classes like this so it shouldn't just be a utility code hack for memoryview.

I think proceed with this PR because we will want to make these immutable again, but fixing up the cases where we(/I) accidentally used mutability won't be instant.

Link 1 #5026

#5030 solves the remaining blocker for this. It is adding a feature though so I'll be patient and let it be reviewed properly.

@scoder I'd quite like to get this in the next release. It's been a bug for a while but it seems like something to fix before people start relying on the wrong behaviour.

It could do with a test - I'll have a go at writing one later today.

It's blocked by #5030 (but there's other options if you don't like that - in the short-term just to disable the tests)

Closing/reopening to rerun tests

Thanks @maxbachmann