perf: add ability to write downcasted indices

[ilan-gold]

TODO:

• gather feedback in Downcast indices for sparse matrices if possible on-disk #2153
• tests for + run on GPU CI We don't support io directly into GPU anyway!

Checks

• Closes Downcast indices for sparse matrices if possible on-disk #2153
• Tests added
• Release note added (or unnecessary)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.65%. Comparing base (5212db8) to head (1361d36).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2159      +/-   ##
==========================================
- Coverage   84.77%   84.65%   -0.12%     
==========================================
  Files          46       46              
  Lines        7132     7142      +10     
==========================================
  Hits         6046     6046              
- Misses       1086     1096      +10     

Files with missing lines	Coverage Δ
src/anndata/_io/specs/methods.py	90.78% <100.00%> (+0.14%)	⬆️
src/anndata/_settings.py	91.86% <100.00%> (+0.04%)	⬆️

... and 2 files with indirect coverage changes

[flying-sheep]

looks great!

When reading this in, will the dtype be preserved? And if so, how do certain APIs deal with tiny int dtypes? Like anndata’s concatenation and scanpy’s algorithms?

[ilan-gold]

When reading this in, will the dtype be preserved? And if so, how do certain APIs deal with tiny int dtypes? Like anndata’s concatenation and scanpy’s algorithms?

scipy will convert for us to match indptr from what I can tell from the issues and the behavior of the tests passing (i.e., the data is read back in and matches the input). So this isn't really a scanpy problem. In the future when supporting different sparse matrices, (like finch tensor), I think we will hopefully be able to preserve types. I don't see anything in https://graphblas.org/binsparse-specification/ that would indicate that the data types have to match

[scverse-benchmark]

Benchmark changes

Change	Before [5212db8]	After [1361d36]	Ratio	Benchmark (Parameter)
-	21.1±3ms	15.6±0.4ms	0.74	backed_hdf5.BackedHDF5Indexing.time_slice_obs_to_memory('sparse')
+	1.99±0.01ms	2.36±0.01ms	1.19	dataset2d.Dataset2D.time_getitem_slice('h5ad', (-1,), 'cat')
+	301±3μs	351±6μs	1.17	sparse_dataset.SparseCSRContiguousSlice.time_getitem_adata('alternating', True)

Comparison: https://github.com/scverse/anndata/compare/5212db80485432719021445084d93407c0ce11b2..1361d36a4dd2e4bc20dfe3f89f03369f131be637
Last changed: Mon, 3 Nov 2025 13:57:48 +0000

More details: https://github.com/scverse/anndata/pull/2159/checks?check_run_id=54361625499