test(cuda.core): fold the pairwise cross-type object-protocol tests into one equality test - #3044

Description

No linked issue; test-suite maintenance. Refs #2745.

test_hash_distinct_cross_type and test_no_cross_type_equality in cuda_core/tests/test_object_protocols.py were parametrized over itertools.combinations of the 25 sample types: 600 of the module's 898 items per CI job, each paying the per-test CUDA context setup and teardown to assert one comparison. On the L4 push row the two tests took about 40 s of a 288 s Run cuda.core tests step (step timestamps of a recent push run on main). A separate single measurement put the 600 items at 43.7 s with the per-test teardown and 5.1 s without it, so the cost is the item count, not the comparisons.

Changes:

• test_no_cross_type_equality is now one unparametrized test over a new eq_samples fixture. The fixture resolves every EQ_TYPES sample at setup on the main thread (pytest-run-parallel workers only read), and the test compares all pairs; a failure lists every offending pair by fixture name.
• A sample whose fixture skips (for example the graph memory nodes on devices without mempool support) is dropped from the comparison and recorded in the test's JUnit skipped_samples property; the remaining types are still compared.
• test_hash_distinct_cross_type is removed. Python does not promise distinct hashes for objects of different types (the test's own comment said so). test_hash_consistency, test_hash_distinct_same_type, and test_equality_same_handle remain.
• HASH_TYPES and EQ_TYPES are unchanged; a new type is still one table row.

Collected items in the module: 898 -> 299 per job. The 600 removed items covered no cuda.core line that another test does not also cover (per-test coverage contexts, 2026-09-10 coverage report).

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

🤖 Generated with Claude Code