fix: hash-based encoding for object-dtype join/groupby keys (v2.2.5)

[dsfulf]

Summary

• Fix: left_join / inner_join / group_by on object-dtype key columns containing heterogeneous Python types (e.g., str and int across sides, or mixed within one side) no longer raise TypeError: '<' not supported between instances of ....
• Root cause: the pure-Python encoding fallback in _encode_columns and _encode_columns_paired calls np.unique(..., return_inverse=True), which internally argsorts the values. Object arrays with mixed Python types fail cross-type < comparison. The C accelerator (_c_encode_strings) handles this correctly via hash equality but only triggers at >= 50k rows; _validate_dtypes lets both sides through because _dtypes='object' on both sides is an equality match.
• Fix: added GroupSet._hash_encode_object — a dict-based codebook used for object-dtype arrays in the pure-Python path. Semantics match _c_encode_strings (hash equality, not ordering).
• Also fixes a pre-existing mypy return-value error in Tafra.to_pandas via cast(DataFrame, ...) — the protocol is intentionally a "fake class to satisfy typing of a pandas.DataFrame without a dependency" and can't structurally match pandas's actual signatures.
• Documents the three version-bump files (pyproject.toml, recipe/meta.yaml, docs/changelog.md) in CLAUDE.md.

Test plan

• Three new regression tests in TestStringDtype:
  • test_join_on_object_dtype_mixed_python_types_across_sides
  • test_join_on_object_dtype_mixed_within_side
  • test_group_by_object_dtype_mixed_python_types
• Full suite: 204 passed
• ruff check tafra — clean
• mypy tafra — clean (was 1 pre-existing error, now resolved)

Generated with Claude Code

[Copilot]

Pull request overview

Fixes TypeError raised when left_join / inner_join / group_by operate on object-dtype key columns containing mixed Python types (e.g., str and int). The pure-Python encoding fallback used np.unique(..., return_inverse=True), which internally argsorts and fails on cross-type <. A new dict-based hash codebook (GroupSet._hash_encode_object) is added with semantics matching the C accelerator (_c_encode_strings), and used for object arrays in both single-column (_encode_columns) and paired (_encode_columns_paired) paths.

Changes:

• Add GroupSet._hash_encode_object() and route object-dtype arrays through it in _encode_columns and _encode_columns_paired.
• Wrap Tafra.to_pandas() return in cast(DataFrame, ...) to silence a pre-existing mypy return-value error against the lightweight DataFrame Protocol.
• Bump version to 2.2.5 across pyproject.toml, recipe/meta.yaml, docs/changelog.md; document the three bump locations in CLAUDE.md.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tafra/group.py	New hash-based object encoder; used in single-column and paired encoding paths below the C threshold
tafra/base.py	cast(DataFrame, ...) around pd.DataFrame(...) return in to_pandas
test/test_tafra.py	Three regression tests for mixed-type object-dtype joins and group_by
docs/changelog.md	2.2.5 release notes
pyproject.toml	Version bump 2.2.4 → 2.2.5
recipe/meta.yaml	Conda recipe version bump 2.2.4 → 2.2.5
CLAUDE.md	Documents the three version-bump locations

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.