fix(rag)!: nested fixed axes, tuple-key axes, and record to_ak depth

[d-laub]

Fixes three Ragged bugs: indexing that silently returned wrong data, and one to_ak() inconsistency. Each fix is its own commit.

#73: fixed axis before two ragged axes

(b, p, ~v, ~w) arrays (gvl 0.43 _FlatWindow.to_ragged()) ignored p in indexing and to_ak().

• R=2 arrays with more than one leading axis now use the multidim segment gather. Outer offsets are gathered and inner offsets stay global.
• If every leading axis gets an int key, the result is peeled, so r[i, j] == r[i][j].
• Numeric R=2 to_ak() wraps fixed leading dims in RegularArray.

#74: tuple keys applied to axis 0 (breaking)

A key after a non-int key went to axis 0 of the intermediate result. For example, rag[:, 0] on (n, ~) returned row 0.

• Each key now goes to the next output axis, as in NumPy and awkward.
• Ragged axes accept an int (one element per row) or a step-1 slice (Python semantics, including negative bounds).
• Trailing axes index the data buffer.
• Unsupported combinations raise NotImplementedError: stepped slices, fancy keys on a ragged axis (except the first of two ragged axes), and trailing keys on records.

#75: record to_ak() depth (breaking)

Records are now always at the leaf, whatever rag_dim is, so rec.to_ak()[k] == rec[k].to_ak(). This also fixes trailing fixed dims being wrapped twice.

Breaking changes

• rag[:, k] and similar tuple keys follow NumPy semantics. To get row k, use rag[k].
• Record to_ak() with one leading axis returns n * var * {a, b} instead of one record of lists per row.

gvl's values[:, t] only indexes leading axes, so it is not affected. No downstream repo calls to_ak() on a record.

Testing

• tests/test_ragged_nested_fixed_axis.py and tests/test_ragged_tuple_index.py check results against awkward indexing on the same data. They cover packed and gathered offsets, strings, and records.
• 778 tests pass. Hooks pass (ruff, ruff format, pyrefly, commitizen).
• Tuple-key overhead is about 0.5 µs (~4%). Single-key paths are unchanged.
• skills/seqpro/SKILL.md is updated.

Closes #73
Closes #74
Closes #75

🤖 Generated with Claude Code

[codspeed]

Merging this PR will degrade performance by 23.03%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

❌ 3 regressed benchmarks
✅ 11 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	test_bench_baseline_ragged_short_alleles	2.8 ms	4.4 ms	-36.01%
❌	test_bench_dense_batch	2.7 ms	3.3 ms	-18.32%
❌	test_bench_baseline_ragged_cres	6.3 ms	7.3 ms	-12.73%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing fix-73 (ff1f21a) with main (63a8439)}