Deterministic, falsifiable evaluation of LLM agents: multi-agent reasoning, theory of mind, game theory and causal spatial reasoning (iXentBench)