GPU

GPU is an optional expert path and must not block a first task. Begin ordinary finite work with submit_auto(lambda); enter this topic only when an operation has independent CPU/GPU implementations, its data size can cover transfer and launch cost, and deployment needs a diagnosable GPU backend. When CUDA/OpenCL runtime is unavailable or no device exists, the basic CPU path still works.

1. Diagnose backend and fall back safely: verify registration failure and CPU fallback even on a machine without GPU.
2. Register and submit GPU work: register, submit, and query status on a real device.
3. CPU/GPU automatic selection: express dual paths with cpu_gpu_task(), configure fallback, and understand the legacy overload boundary.

Build options, device query, and examples are in docs/BUILD.md, examples/gpu_device_query.cpp, and docs/design/gpu_executor.md. Any performance conclusion records device, driver, data size, and build type.