🐛 fix: match the frozendict backport to the Python 3.15 built-in

[nstarman]

Brings xmmutablemap._frozendict.frozendict (the backport used on Python < 3.15) in line with the PEP 814 built-in.

How this was checked

I wrote a differential probe that loads the backport from its file on CPython 3.15.0 and runs ~145 operations against both it and builtins.frozendict, including subclasses. It covers construction, ==, |, |=, hashing, repr (including recursive), views, get/in/[] with bad keys, pickling at every protocol, copy/deepcopy, match, generic aliases, and introspection. Before this PR: 25 differences. After it, the only differences left are the ones listed under "Not changed" below.

Behavioural fixes

	Built-in	Backport before
d.__init__(b=2)	no-op	replaced the contents (an immutability hole)
Subclass __init__(self, x)	__new__ builds from the args first	__init__ never set _data
frozendict(d) is d (exact type)	True	False
frozendict(a=1) | Plain(...) (Mapping without |)	TypeError	merged it
Plain(...) | frozendict(a=1) message	names 'frozendict'	named 'dict'
copy.copy(d) is d (exact type)	True (3.15's copy treats it as atomic)	False
Pickle protocol 0/1	TypeError: cannot pickle 'frozendict' object	TypeError with a __slots__ message
__getnewargs__ / __reduce_ex__	present / (__newobj__, (cls, data), state, None, None)	missing / slot state
hash(frozendict(a=[])) message	unhashable type: 'list'	on 3.15, cannot use 'tuple' as a set element (...)
frozendict(1, 2) message	frozendict expected at most 1 argument…	dict expected…

The probe also showed that the built-in's hash is hash(frozenset(items)), so the docstring no longer claims the hash values differ.

Compatibility: pickles written by the current release still load (checked with protocol 4 for both frozendict and ImmutableMap). ImmutableMap behaves the same apart from the fixes above.

Not changed (documented in the module docstring)

• The Mapping base class stays, so type checkers on 3.10–3.14 still see a Mapping. As a result the MRO, dir(), and the ABCMeta metaclass differ from the built-in, which is only registered as a Mapping.
• .keys().mapping wraps the backing dict. Matching it would need custom view types, which would no longer be dict_keys.
• Some messages name dict where the built-in says frozendict, e.g. for unhashable keys. The exception types match.
• Cosmetic: __module__ and inspect.signature differ.
• Intentional shortcut: the private _data slot can still be rebound, and a subclass's extra __slots__ aren't pickled. Both are marked ponytail:.

Tests

I added parity tests to TestFrozenDict. That class runs against the backport on 3.10–3.14 and against the built-in on 3.15+, so every expected value comes from the built-in. They cover __init__ immutability, the identity constructor, subclass __init__, | rejecting a plain Mapping, the hash value and its error message, copy/deepcopy on the exact type and a subclass, __getnewargs__, and the refusal of protocols 0/1.

• Python 3.10 and 3.13 (backport): 135 passed, 100% coverage.
• Python 3.15.0 with jax 0.11.2 (built-in): 135 passed.
• Hooks: prek run --all-files passes.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@27442c0). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff            @@
##             main       #82   +/-   ##
========================================
  Coverage        ?   100.00%           
========================================
  Files           ?         3           
  Lines           ?        98           
  Branches        ?         0           
========================================
  Hits            ?        98           
  Misses          ?         0           
  Partials        ?         0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.