...i suppose we'd have to schedule the entire context in execution_graph, rather than only scheduling portions that we missed for