How Causal Reasoning Addresses the Limitations of LLMs in Observability