The proliferation of misinformation, exacerbated by Generative Artificial Intelligence and Large Language Models, necessitates a paradigm shift from pattern-matching to reasoning-based detection systems. While Transformer-based models and Graph Neural Networks have advanced the field, they often lack interpretability and robustness against adversarial attacks. This survey presents a systematic review of Reasoning-Centric Fake News Detection methodologies for the last five years. Unlike conventional surveys that categorize works by architecture, we propose a novel four-phase evolutionary framework that traces the transition from surface-level semantics to multi-view reasoning paradigms. We introduce a multi-axis taxonomy that unifies content semantics, dynamic propagation topologies, and cross-modal consistency. Furthermore, we critically analyze the “performance saturation” phenomenon (reaching 99% on static benchmarks such as FakeNewsNet), contrasting it with the degradation observed in low-resource languages and in the face of evolving events. Finally, we propose an Integrated Multi-View Reasoning Framework, a five-layer pipeline designed to harmonize automated detection with human-in-the-loop verification, and outline a future roadmap focused on neuro-symbolic reasoning and adversarial resilience in the Artificial Intelligence vs Artificial Intelligence era.