fix: groupby-sum fast path reduces multi-dim DataArray grouper over all its dims

[FabianHofmann]

Closes #823.

This aligns the grouping mechanism across fallback and non-fallback groupby operations.

Note

The following content was generated by AI.

Changes proposed in this Pull Request

LinearExpression.groupby(...).sum() returned wrong dimensions when grouping by a multi-dimensional DataArray grouper on its default (fast) path: the reduction happened over only one of the grouper's dimensions and the others survived in the result. Only use_fallback=True gave the correct result.

Root cause: the fast path did group.to_pandas(), which turns a 2-D grouper into a DataFrame that is then routed through the multikey encoding path — reducing over only the row axis. A genuinely N-D grouper (whose values define the groups jointly over all its dims) was never handled. This is a long-standing bug present since v0.8.0, not a regression from #802.

The fix:

• Detect a genuine N-D grouper — a DataArray with ndim > 1 whose dims are all data dims — and stack the data over those dims into a single key, so the fast path reduces over every grouper dim at once. This preserves the fast path's memory benefit.
• Grouper DataArrays whose extra axis is not a data dim (i.e. multikey frames built from a DataFrame) keep flowing through the existing DataFrame/multikey handling.
• _grouped_sum now accepts the (stacked) dataset to operate on.

Tests

• test_linear_expression_groupby_ndim now exercises the fast path too (use_fallback=[True, False]) and asserts it equals the fallback; the stale TODO was removed.
• New test_linear_expression_groupby_multidim_preserves_extra_dim covers the issue's reproducer (2-D grouper over (i, j), preserved k).

Checklist

• AI-generated content is marked (see AGENTS.md).
• Code changes are sufficiently documented; i.e. new functions contain docstrings and further explanations may be given in doc.
• Unit tests for new features were added (if applicable).
• A note for the release notes doc/release_notes.rst of the upcoming release is included.
• I consent to the release of this PR's code under the MIT license.

[codspeed-hq]

Merging this PR will not alter performance

✅ 173 untouched benchmarks
⏩ 173 skipped benchmarks¹

---

_{Comparing fix-823-multidim-grouper (562c518) with master (a6ca0be)}

Footnotes

1. 173 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[FBumann]

@FabianHofmann Great that you found that bug right away. Im really sorry about the regression!

EDIT: Nevermind, seems like it wasnt my PR...

[FBumann]

@FabianHofmann I dived a bit deeper in and found some other issues (positional alignment when using pd.Series or xr.DataArray). This is arguably a bug, but could also be part of v1. Ill create issues for both. We should decide if we treat this as a bugfix or a semantics thing!

[FBumann]

See #827

[FBumann]

@FabianHofmann This should be merged no matter what happens in #827