GH-50161: [C++][IPC] Validate CSF sparse index buffer counts

[metsw24-max]

Rationale for this change

ReadSparseCSFIndex in cpp/src/arrow/ipc/reader.cc sizes indptr_data/indices_data from the tensor shape (ndim - 1/ndim) but fills them by looping over the flatbuffer-supplied buffer counts, which are never checked against ndim. A crafted CSF sparse tensor message with mismatched counts writes past the end of those vectors, and ndim == 0 builds a SIZE_MAX-sized vector. The payload path is already guarded by CheckSparseTensorBodyBufferCount; the file/stream path was not, so this is reachable from ReadSparseTensor(io::InputStream*) on untrusted bytes. GetSparseCSFIndexMetadata in metadata_internal.cc has the same issue as an out-of-bounds read (axisOrder() vs indicesBuffers() lengths).

What changes are included in this PR?

Callee-side count validation in both functions: reject CSF indices whose indptrBuffers/indicesBuffers counts don't match ndim - 1/ndim (and ndim < 1), and whose axisOrder/indicesBuffers lengths differ, with Status::Invalid.

Are these changes tested?

Covered by the existing sparse tensor IPC round-trip tests; the rejected inputs are only producible from hand-crafted flatbuffers.

Are there any user-facing changes?

No, only invalid inputs are rejected.

This PR contains a "Critical Fix". A crafted IPC sparse tensor message could trigger a heap out-of-bounds write (shared_ptr<Buffer> constructions past the vector end) or an out-of-bounds flatbuffer read from the public ReadSparseTensor API.

• GitHub Issue: [C++][IPC] ReadSparseCSFIndex does not validate CSF index buffer counts against ndim #50161

[pitrou]

@metsw24-max Bug fixes or security fixes are not minor changes, can you open a corresponding GH issue for this PR? Thank you!

[metsw24-max]

Makes sense. Opened #50161 and retitled the PR, and filled in the template body since this qualifies as a critical fix.

[github-actions]

⚠️ GitHub issue #50161 has been automatically assigned in GitHub to PR creator.

[rok]

This looks good. Could you please add some regression tests? I'm mostly interested in ndim=1 (it should be invalid CSF index if I remember correctly).

[metsw24-max]

Added regression tests in tensor_test.cc that feed hand-crafted CSF messages through ReadSparseTensor.

You're right about ndim=1: it has no indptr buffers, so it slipped past the old ndim < 1 guard and reached SparseCSFIndex's constructor, which dereferences indptr_.front() on an empty vector. I've tightened the reader to require ndim >= 2. The new test covers that case plus too-many indices buffers and an axisOrder/indicesBuffers length mismatch, and the existing round-trip tests still pass.

[metsw24-max]

Pushed a clang-format fix for the lint failure. The other two reds look unrelated to this change: the C++ AVX2 job fails on a Gandiva timezone test (TestTime.TestCastTimestampWithTZ) and the Windows Python job on a pyarrow flight import, neither of which this patch touches. The IPC tests built and passed on the ASAN/UBSAN and macOS C++ jobs.

[rok]

This is a good idea to check.

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit e92d5d7.

There were 3 benchmark results indicating a performance regression:

• Commit Run on amd64-m5-4xlarge-linux at 2026-06-30 18:47:40Z
  • dataset-selectivity (Python) with dataset=chi_traffic_2020_Q1, selectivity=100%
  • dataset-selectivity (Python) with dataset=chi_traffic_2020_Q1, selectivity=10%
• and 1 more (see the report linked below)

The full Conbench report has more details.