test: add a property-based kernel test with CPU and GPU backends

[TaiSakuma]

Summary

As discussed in the awkward–uproot meeting, this is a pilot for property-based testing of the compiled kernels. Instead of the frozen samples in kernel-test-data.json, inputs are Hypothesis-generated and the kernel is checked against its pure-Python reference implementation in kernel-specification.yml (the single source of truth).

This starts with one kernel, awkward_BitMaskedArray_to_ByteMaskedArray, plus the CPU/GPU infrastructure. If the approach works well, it will be extended to the remaining kernels.

What's here

• tests/properties/kernels/test_BitMaskedArray_to_ByteMaskedArray.py — parametrized over backends:

  • cpu: compiled kernel via awkward_cpp.cpu_kernels (ctypes), always run.
  • cuda: compiled kernel via the CuPy backend, added only when a GPU is present and marked @pytest.mark.cuda.

  The reference implementation is loaded from kernel-specification.yml, so the test does not depend on the generated awkward-cpp/tests-spec/kernels.py.

• pyproject.toml — registers the cuda marker (--strict-markers).

• reusable-test.yml — new GPU step pytest tests/properties -m cuda, gated on run-gpu-kernel-tests.

• reusable-change-detection.yml — tests/properties/** added to the GPU change filter, so property-test changes trigger the GPU job.

• requirements — hypothesis-awkward + PyYAML in requirements-test-gpu.txt; PyYAML in requirements-test-full.txt.

Dependency on #4128

This test intentionally carries no @settings (no max_examples, no deadline). The example budget and per-example deadline are expected to come from the central Hypothesis profile in #4128. Until that lands, the GPU run uses Hypothesis defaults (200 ms deadline), which GPU launch latency can exceed; #4128 is the intended source of a GPU-appropriate profile.

Test plan

• pytest tests/properties/kernels — cpu param passes.
• pytest tests/properties/kernels -m cuda deselects cleanly with no GPU; cuda marker registered.
• pre-commit on changed files: ruff, codespell, workflow validation, zizmor.
• GPU job on this PR (self-hosted runner) runs the CuPy kernel against the spec's reference implementation. (Validated in CI; see ci: run property-based tests nightly with a large max_examples #4128 note re: deadline.)

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.42%. Comparing base (95fcb30) to head (6fc4e22).

Additional details and impacted files

see 1 file with indirect coverage changes

[TaiSakuma]

After #4128 was merged, all tests in this PR passed.

[ianna]

@TaiSakuma -,this is great! Thank you. Please go ahead and merge it if you are done with it. Thanks