Fix MATX_ASSERT_COMPATIBLE_OP_SIZES macro checks of first N-1 dims

[tbensonatl]

The MATX_ASSERT_COMPATIBLE_OP_SIZES macro checks for compatible sizes between the dimensions of two operators. It previously replace the compatible flag when computing it for each dimension and thus would only throw an error when operators differed in the last dimension.

This PR fixes the check such that compatible is only true if true for all dimensions. It also adds new unit tests that would fail without the fix.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

This PR fixes four distinct but related bugs around dimension/size checking in MatX operators, and adds regression tests to cover each case.

• include/matx/core/error.h — Core fix: compatible = → compatible &= in MATX_ASSERT_COMPATIBLE_OP_SIZES. Without this change the loop overwrote the flag on every iteration, so only the last dimension was ever validated.
• include/matx/operators/cov.h — CovOp constructor now correctly assigns the last two output dimensions as {M, M} (where M = a.Size(Rank()-1)) instead of naively copying all input dimensions; previously the second-to-last output dimension was set to the number of observations rather than the number of variables.
• include/matx/transforms/cov.h — Batch-dimension equality check in matxCovHandle_t now starts from index 0 instead of 2, ensuring the first two batch dimensions are also validated.
• include/matx/operators/conv.h — Conv2DOp constructor now computes per-tensor dimension indices (a_idx, b_idx) when OpA::Rank() != OpB::Rank(), preventing incorrect or out-of-bounds Size() calls on the lower-rank tensor.
• New unit tests in complex_cast_exceptions_test.cu and Cov.cu exercise all fixed code paths and would have caught the original bugs.

Confidence Score: 5/5

• All changes are targeted, well-reasoned bug fixes with comprehensive new tests; safe to merge.
• Each fix is minimal and directly addresses a documented bug. The one-character change in error.h (= → &=) is the simplest possible correct fix. The cov.h and transforms/cov.h changes are self-consistent and validated by new typed tests. The conv.h index-offset fix is logically sound for all rank combinations ≥ 2, which is the only valid input for a 2D convolution operator. New regression tests cover the previously failing scenarios for every touched code path.
• No files require special attention.

_{Last reviewed commit: ff6cd3d}

[tbensonatl]

/build

[tbensonatl]

/build

[tbensonatl]

/build