[u-simha]

What:

• Fix the CoreAI op-order check in tests/export/test_joint_sparsity_palettization_e2e.py so it passes with any Core AI build, not only the one it was written against

Why:

• The check read the lowered MLIR text for coreai.build_sparse_with_bitmask and coreai.sparse_with_bitmask_to_dense. Builds of Core AI that lower the sparse ops under other names make it fail with assert 0 == 6, even though the conversion and the output comparison pass. That fails the six CoreAI test_mnist cases and both test_mnist_vector_palettization cases.

How:

• For CoreAI, check the order on the exported torch program instead: lut_to_dense -> sparse_to_dense -> constexpr_blockwise_shift_scale (the last one only when a per-channel scale or a quantized LUT is used). This is the layer coreai-opt produces, and the test already uses it to count the ops, so the check no longer depends on how a Core AI version lowers or prints them.
• The program is traced on a copy of the finalized model before it is converted, since converting may modify the model.
• The CoreML check on the MIL program is unchanged.
• The order check now lives in _export(..., check_order=True), so the tests no longer call a separate _assert_op_order.

Testing:

• The non-slow tests of the file pass locally with the public coreai-core and coreai-torch releases.
• Inserting an extra op between lut_to_dense and sparse_to_dense makes the same eight CoreAI cases fail, so the check still guards the order.
• This is a test-only change. I could not run it against the other Core AI build that showed the failure, so a green CI run on this PR is the confirmation.

[dengqiaoyu]

Thanks for the fix!