petbox-dev / tafra

98%

Build Type

push

github

Committed by dsfulf

Commit Message

code-correctness: revert the signed-zero tie-break; narrow the StringDType cast

- Revert the C min/max tie-break. numpy's sign of a zero min or max is not
  defined beyond about 4 values (it follows SIMD lane order: nanmax of
  [0,0,0,0,-0.0] is -0.0), so the tie-break matched numpy only for small
  groups, and it cost up to 3.5x per kernel call on tie-heavy data (+36-50%
  per group_by, 1k groups, MSVC /O2). The guide now states the sign is
  unspecified; the test checks values, including 5-row groups.
- Cast a StringDType join key only when its na_object is not None:
  StringDType() and coerce=False columns are no longer copied (9-72 ms per
  side per 1M-row join).
- Test the non-default na_object on the left side as well.
- joins.md NULL table names the StringDType na_object; variance note says
  the difference grows with the square of the offset.
- docs/todo: NaN-as-text evidence on the non-equi item; new item for
  union/concat with mixed StringDType na_object.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Coverage Stats

1 of 1 new or added line in 1 file covered. (100.0%)

11 existing lines in 2 files now uncovered.

1724 of 1756 relevant lines covered (98.18%)

0.98 hits per line