In this test all numbers are big (~1e9), so it can be precision errors/overflow. You should use NTT with big modulo to avoid doubles