OpenClAadEngine

OpenCL replay engine: the recorded tape becomes a fused forward+adjoint OpenCL C kernel, built once with clBuildProgram and replayed one scenario per work-item.

Priority sits below SIMD, so .fastest() never lands here: the OpenCL device chosen by default is usually the vendor GPU runtime, which on the integrated parts this repo runs on shares the display scheduler and carries the same wedge risk as the ROCm path. Ask for it deliberately with -Dnablatensor.engine=opencl / .on("opencl"), and pick a specific ICD with -Dnablatensor.opencl.platform=<substring> (e.g. rusticl, portable, pocl) when more than one is installed.

Holds no reference to libOpenCL beyond isAvailable(), which probes through the OpenClCompute facade, so merely enumerating engines cannot fail where OpenCL is absent.