Function plan_epistemic_gpu_execution 

pub fn plan_epistemic_gpu_execution(
    program: &Program,
) -> Result<EpistemicGpuPlan>

Build the production-facing GPU execution contract for an epistemic program.

This does not launch kernels. It proves that the semantic boundary can be represented as a GPU-native execution plan with explicit hot-path phases, required device buffers, WCOJ planning obligations, and zero CPU fallback counters.