A1: Vectorize _codes_to_group_indices

[FBumann]

Problem

xarray/core/groupby.py:91 _codes_to_group_indices is a Python-level loop that builds per-group index lists via list.append. Profile (200k rows, 1k groups) shows it consumes 89% of groupby() setup time:

89.00% 1739 _codes_to_group_indices xarray/core/groupby.py
 3.48% 68 factorize_array pandas/core/algorithms.py

200,001 list.append calls visible in the profile.

Proposed fix

Replace the loop with vectorized NumPy:

• order = np.argsort(codes, kind="stable")
• split_points = np.searchsorted(codes[order], np.arange(n_groups))
• groups = np.split(order, split_points[1:])

Or use np.unique(codes, return_inverse=True) + a single bincount + argpartition pass.

Expected outcome

~10× faster groupby construction for moderate-to-large group counts. Most user-visible in time_init benchmarks and as setup cost before .mean()/.sum() agg calls.

Benchmark

Existing `asv_bench/benchmarks/groupby.py::GroupBy.time_init` covers low-group-count case. Add a high-cardinality variant (200k rows, 1k groups) as the realistic regression target.

Acceptance

• All existing groupby tests pass
• New asv benchmark shows the speedup
• No behavioral change for empty groups, NaN codes, or unordered codes

---

[This is Claude Code on behalf of Felix Bumann]