from benchmark.execution_order import generate_execution_order def test_execution_order_reproducibility(): order1 = generate_execution_order(10, seed=42) order2 = generate_execution_order(10, seed=42) assert order1 == order2 assert len(order1) == 10 # Check that it alternates/mixes orders plains_first = [x for x in order1 if x[0] == "plain"] rifs_first = [x for x in order1 if x[0] == "rif"] # Assert we have both combinations assert len(plains_first) > 0 assert len(rifs_first) > 0