fix(rag): handle indexed empty row in Ragged.to_numpy

[d-laub]

Summary

Fixes #67. Ragged.to_numpy() raised ValueError: cannot reshape array of size 0 into shape (0) on an empty row (length 0) obtained by indexing a multi-dimensional Ragged:

rag = Ragged.from_lengths(np.array([5, 7], np.int32), np.array([[[2], [0]]]))
empty = rag[0, 1, 0]   # lengths [0], is_base=False
empty.to_numpy()       # ValueError before this fix

A directly-constructed empty row converted fine (it took the is_base fast path); only the indexed, non-base empty row failed.

Root cause

In to_numpy (python/seqpro/rag/_core.py), when a Ragged is fully indexed down to a single ragged axis, leading (the leading dense dims) is empty, so the final reshape used (leading or (-1,)) → data.reshape(-1, 0). numpy cannot infer -1 against a 0 dimension on a size-0 array, so it raised.

Fix

Compute an explicit row count n_rows in both the validate=True branch (lengths.size) and the trust-the-caller branch (already present), and use (leading or (n_rows,)) instead of -1. Empty rows now reshape cleanly to (n_rows, 0); non-empty single-row results are unchanged (n_rows == 1).

Tests

Added test_to_numpy_indexed_empty_row (parametrized over validate=True/False) asserting the indexed empty row converts to shape (1, 0). Fails before the fix, passes after. Full ragged suite (263 tests) passes; ruff/pyrefly clean.

No SKILL.md update required — bugfix with no public signature change.

Context

Unblocks the genoray cleanup noted in the issue: removing the INT64_MAX offset sentinel workaround that exists specifically to detect-and-skip empty rows before to_numpy().

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will improve performance by 68.96%

⚠️ 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 improved benchmarks
✅ 11 untouched benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_bench_ragged_cres	2,574.2 µs	955.2 µs	×2.7
⚡	test_bench_baseline_ragged_short_alleles	2.9 ms	2.1 ms	+38.22%
⚡	test_bench_dense_batch	5 ms	3.8 ms	+29.5%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing fix/to-numpy-indexed-empty-row (e66e259) with main (cbbdabf)¹}

Footnotes

1. No successful run was found on main (e66e259) during the generation of this report, so cbbdabf was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩