fix: compare by equality, not just hash, in ExtendedSet

[binggao1230]

Summary

The append_unique list strategy silently drops genuinely-distinct elements when they collide on a hash. ExtendedSet.__contains__ decides membership purely from the element's hash and never checks equality, and _hash_element builds a lossy "key:value" string for dicts:

from deepmerge import Merger
m = Merger([(list, ["append_unique"])], ["override"], ["override"])

m.merge([{"a": "1", "b": "2"}], [{"a": "1,b:2"}])
# -> [{'a': '1', 'b': '2'}]            # second dict dropped; expected both kept

Three independent collision classes, all silently wrong:

case	hash collision	result (before)	expected
{"a": "1", "b": "2"} vs {"a": "1,b:2"}	both → "a:1,b:2"	one dropped	both kept
{"k": 1} vs {"k": "1"}	both → "k:1"	one dropped	both kept
-1 vs -2	hash(-1) == hash(-2) == -2	-2 in ExtendedSet([-1]) is True	False

Cause

ExtendedSet stored a single {hash: element} mapping and __contains__ returned self._hash_element(obj) in self._values_by_hash — a hash-only comparison. Equality was never consulted, so any hash collision (including the lossy dict-string hash) aliased distinct values together.

Fix

Group elements into hash-keyed buckets (dict[int, list]) and verify true == against the bucket members on both insert and lookup. The hash is only a bucket key; equality is the arbiter. No public API change.

This preserves the existing intended behavior — equal dicts still dedupe regardless of key order (test_strategy_append_similar_dict) and distinct dicts are still both kept (test_strategy_append_unique_nested_dict).

Verification

• Added test_strategy_append_unique_keeps_hash_colliding_dicts and test_strategy_append_unique_hashable_hash_collision, covering all three collision classes. They fail before the fix and pass after.
• Full suite: 19 passed. black --check, mypy, and the existing append_unique/append_similar tests all pass.

This pull request was prepared with the assistance of AI, under my direction and review.

[toumorokoshi]

Thanks for the change, and great catch! if you can fix the lint errors, LGTM.

[binggao1230]

Thanks. I pushed a Black formatting fix for deepmerge/strategy/fallback.py.

Local checks run:

uvx --python 3.12 black --check deepmerge
.venv/bin/validate-pyproject pyproject.toml
.venv/bin/mypy deepmerge
.venv/bin/pytest deepmerge

The new GitHub Actions run is currently waiting for maintainer approval (action_required) because this is a fork PR.

[toumorokoshi]

great, thank you!