Function get_kernel 

pub fn get_kernel(
    registry: &'static AtlasRegistry,
    module: &str,
    func: &str,
) -> RawCudaFunc

Resolve module::func for a bench.

Each bench function calls this ONCE, before its timed group — so the OnceLock the benches used to declare at file scope memoized a lookup that already happened exactly once. It is a local here, which keeps raw_function_cached’s signature satisfied without a process global per kernel per bench binary.