Deceiving Semantic Shortcuts on Reasoning Chains: How Far Can Models Go without Hallucination?