On the Brittle Foundations of ReAct Prompting for Agentic Large Language Models