fix: dedupe identical NaN objects in append_unique, matching set semantics - #54

Fixes #53

Summary

ExtendedSet bucket membership compares with == only, but CPython's set membership short-circuits on identity first (PyObject_RichCompareBool uses x is y or x == y). The original set(base)-based append_unique relied on that fast path; the ExtendedSet rewrite silently dropped it, so the same float('nan') object is now appended again instead of being recognized as a duplicate:

>>> nan = float('nan')
>>> nan in {nan}                    # True
True
>>> nan in ExtendedSet([nan])        # False
False
>>> merge([nan], [nan])             # was [nan] before ExtendedSet
[nan, nan]

_insert and __contains__ now compare identity-first via a small _is_or_equal helper that mirrors PyObject_RichCompareBool, restoring plain-set semantics for identical objects whose __eq__ is not reflexive. Two different NaN objects are still both kept, exactly as a plain set does.

Validation

• New regression tests in tests/strategy/test_list.py:
  • test_strategy_append_unique_identical_nan — fails on un-fixed HEAD.
  • test_extended_set_membership_identity_semantics — direct membership pin, fails on un-fixed HEAD.
  • test_strategy_append_unique_distinct_nan_kept — pins the unchanged distinct-NaN behavior (passes both sides).
• Full suite: 48 passed on system Python 3.13 (45 baseline + 3 new).
• Integration battery (standalone): append_unique unchanged for plain values, equal dicts, unhashable lists/tuples, and the Deduplicate incoming values in append_unique #46 repeated-incoming-values cases; same-object unhashables ([1], {"a": 1}) deduped as before.
• make ready-pr not run on this host (Black/validate-pyproject unavailable); CI to cover.

Notes

This is complementary to (and intentionally does not overlap) #52, which addresses the uninitialized set storage inherited by the class. No shared hunks; branches are independent cuts off master.