[TEST] Cover Gluon core simulation ops

[Jokeren]

Summary

Adds focused CPU regression coverage for Gluon core simulation ops and fixes the simulator layout propagation needed by those kernels. The new tests cover scalar range loops, program_id, arange, masked load/store, 2D broadcasted offsets, SliceLayout, convert_layout, and elementwise add without requiring a CUDA device.

Refactors the Gluon simulator builder so inherited Triton interpreter operations can attach Gluon layouts through a shared _WRAPPED_LAYOUT_OPS path while keeping special layout/signature cases as explicit builder methods. create_broadcast, create_expand_dims, and create_histogram now delegate to the parent interpreter implementation for data behavior and attach the Gluon layout locally. Tensor-producing paths use _tensor_result so layout metadata stays attached to returned handles.

Validation

• PYTHONPATH=/mnt/keren/triton-viz pytest tests/end_to_end/test_gluon_core_ops.py tests/end_to_end/test_gluon.py tests/end_to_end/test_core.py tests/unit/test_adapters.py -q
• pre-commit run --files triton_viz/core/simulation/gluon.py

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.096s	0.097s	+1.1%	20 / 20
gemm_oob	0.108s	0.108s	+0.5%	20 / 20
indirect_load	0.019s	0.019s	+0.5%	20 / 20
nested_loop	0.201s	0.202s	+0.4%	20 / 20
block_pointer_loop_advance	0.195s	0.196s	+0.7%	20 / 20
liger_jsd	0.133s	0.133s	-0.2%	20 / 20
flaggems_layernorm	0.352s	0.353s	+0.2%	20 / 20
swiglu	0.164s	0.164s	+0.4%	20 / 20
cross_entropy	0.920s	0.914s	-0.6%	20 / 20
fused_linear_jsd	0.202s	0.203s	+0.7%	20 / 20
Total	2.390s	2.390s	+0.0%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8928cab66b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve the data layout for FP4 upcasts

When an FP4 value is upcast with a separate scale tensor, this tags the unpacked result with the scale tensor's layout rather than the packed data/output layout. In MXFP4-style kernels the scale tile often has a different layout/shape from the values it scales, so the next Gluon operation infers the wrong distributed layout and can reject otherwise valid combinations with data-layout tensors as a layout mismatch. The FP8 path just above keeps the source layout; the FP4 path should not derive the result layout from scale.

Useful? React with 👍 / 👎.