Function gpu_for_test 

pub fn gpu_for_test(label: &str) -> GpuTestGate

Resolve a runtime for a GPU-conditional test.

Panics when the device is faulted (Err) or when the process policy is GpuPolicy::Required and no device is present. Returns GpuTestGate::AbsentDevice only for a genuinely device-free host.

Two questions, two sources

Is there a device? is asked of the driver with an explicit GpuPolicy::Auto, never with global_policy. Must an absent device be fatal? is the only question global_policy answers here.

Reading the process policy for the availability question would be wrong, and not theoretically: crate::configure_global_policy is a first-writer-wins OnceLock, and GpuRuntime::resolve(Off) short-circuits to Ok(None) before probing any device. Eleven test files set the policy to GpuPolicy::Off, and one of them — backend_status_and_policy_dispatch_are_consistent — shares the tests/arrow_gpu binary with the gated tests here. Had this asked resolve(global_policy()), then on a host WITH a CUDA device, whenever that sibling won the race, every gated test in the binary would record “no CUDA device” and be counted as an absent-device skip. The count this module exists to make trustworthy would then be measuring test execution order, and the eventual “any no-cuda skip on a GPU runner is a failure” gate would fire on a machine that has a GPU.

resolve(Auto) returns Ok(None) for genuine absence only — never for a policy reason — which is what makes an AbsentDevice here mean what it says.