[BUG] cute.slice_ SIGABRTs on a static swizzled ComposedLayout

Describe the bug

cute.slice_ native-aborts on a fully static swizzled ComposedLayout. No Python exception, no MLIR diagnostic. Compiler crash, not a kernel failure.

Not the same as #3255 (that one is tensor[…] with dynamic ? strides).

Steps/Code to reproduce bug

import cutlass.cute as cute

@cute.jit
def repro():
    layout = cute.make_composed_layout(
        cute.make_swizzle(3, 4, 3),
        0,
        cute.make_layout((10, 2), stride=(2, 1)),
    )
    print(layout, flush=True)
    cute.slice_(layout, (None, 0))

if __name__ == "__main__":
    cute.compile(repro)

S<3,4,3> o 0 o (10,2):(2,1)
Aborted (core dumped)

Same slice succeeds if the first mode is 8 instead of 10. cute.select(layout, [0]) on this layout also succeeds.

Expected behavior

Return a sliced layout, or raise a Python/MLIR error. Should not abort().

Environment details

• Bare-metal, NVIDIA H20 (sm_90), driver 535.183.06, Linux x86_64
• Python 3.13.14, CUDA 13.0
• nvidia-cutlass-dsl == 4.6.1 and 4.7.0 (both abort on this repro)

Additional context

Python stack ends in SliceOp.__init__ (cute.slice MLIR op). _cutlass_ir*.so is stripped, so there is no usable C++ backtrace.

Start with the provided @cute.jit repro and the Python stack ending in SliceOp.init for the cute.slice MLIR operation. Compare the failing static swizzled layout with the working mode-8 case and the successful cute.select call. Done means the repro no longer aborts and instead returns a sliced layout or raises a Python/MLIR diagnostic.