Stale data in large language model workflows can silently produce incorrect outputs without triggering system error messages