SCALE's cub wrapper header breaks thrust against CUDA 13.1 CCCL

[diskdog]

SCALE 1.7.1 and the 2026.07.02 nightly (clang 20.1.8), NVIDIA target, CUDA 13.1 toolkit, sm_75, compile-only.

Any file that includes a thrust container header fails under SCALE, with the error inside SCALE's own compatibility shim:

#include <thrust/device_vector.h>
int main() { thrust::device_vector<int> v(4); return 0; }

• NVIDIA nvcc -arch=sm_75 -std=c++17 -c repro.cu: ok
• SCALE llvm/bin/nvcc -arch=sm_75 -std=c++17 -c repro.cu (1.7.1 and nightly, same error):

llvm/lib/clang/20/include/cuda_wrappers/cub/util_device.cuh:7:26: error: explicit
instantiation of 'EmptyKernel' does not refer to a function template, ...

Cause: the wrapper does

CUB_NAMESPACE_BEGIN
template __global__ void EmptyKernel<void>();
CUB_NAMESPACE_END

but in the CCCL that ships with CUDA 13.1, EmptyKernel lives in cub::detail (include/cccl/cub/util_device.cuh:83), not at cub:: scope, so the explicit instantiation names a template that no longer exists there. The wrapper predates the CCCL move; detail::EmptyKernel<void> is what it wants now.

Footprint in the validation sweep: every repo that touches thrust via CUDA 13.1 dies on this (8 of the first 122 swept repos: Alenka, cs344, multem, NKSR, nvParse, PyBindToGPUs, SIGA, and a Multigrid solver). In larger translation units the error is often followed by the issue #43 assertion crash, so this wrapper error is also the gateway to a chunk of the #43 noise.