Add frozendict support for 3.15+ - #1052

Great stuff @sobolevn! Couple of questions :)

Just some thoughts on the docs, the rest looks good :)

🔥

Took a careful pass on this (including building the branch against 3.15.0a7) and found two issues to fix before merge. Nice feature otherwise, the type-system integration is clean.

1. Reference leak in every frozendict construction path.

PyFrozenDict_New(dict) does not steal its argument: for a plain dict it copies (via PyTuple_Pack + frozendict_new) and leaves the caller's reference intact (CPython Objects/dictobject.c, which is exactly why you proposed PyFrozenDict_FromDictSteal upstream). The wrappers do return PyFrozenDict_New(dict); without releasing the intermediate dict, so it leaks on every call: mpack_decode_frozendict, json_decode_frozendict, to_builtins_frozendict, convert_dict_to_frozendict, and the frozendict branch in convert_other.

Reproduced on 3.15.0a7, counting live dict objects after 5000 ops:

decode -> dict         : +0
decode -> frozendict   : +5000
convert -> frozendict  : +5000
to_builtins frozendict : +5000

Fix at each site:

PyObject *res = PyFrozenDict_New(dict);
Py_DECREF(dict);
return res;

(this also handles the case where New returns NULL). Worth a refcount regression test, since the functional tests pass straight through the leak.

2. msgspec.yaml.encode raises on frozendict.

to_builtins yields a frozendict, and PyYAML's SafeDumper has no representer for it:

msgspec.yaml.encode(frozendict({"a": 1}))
# RepresenterError: ('cannot represent an object', frozendict({'a': 1}))

Same for a struct or dict containing a frozendict. (stdlib json.dumps(to_builtins(...)) happens to work, so it is specific to consumers that do not special-case frozendict, but yaml.encode is public API.) The yaml path likely needs to emit a plain dict for frozendict. A yaml roundtrip test would catch it.

Both reproduced on a local 3.15.0a7 build of this branch.

Thanks for working on this! I wish we had better benchmarks (or added back the old profiling test dependency) so that we could assess the impact on performance.