Multimodal fake news detection is critical for maintaining the integrity of social media. Current models often rely on shallow cross-modal correlations and rigid domain labels, which limits robustness to logical inconsistency and domain shift. We propose the Open-Vocabulary and Logic-Aware Fake News Detection (OVLAFND) framework to address these issues and relax traditional categorization constraints. To model fine-grained logical conflicts rather than surface relevance, our framework introduces Logical Adversarial Data Augmentation (LADA), a multimodal LLM-driven pipeline that constructs counterfactual text-image pairs as hard negatives. Then, we propose a Logic Consistency Module (LCM) that uses these adversarial signals to learn a dedicated logic-consistency subspace for assessing whether textual claims are supported by visual evidence. To improve topic flexibility, we transform the rigid domain labels into an Open-Vocabulary Topic Manifold (OVTM) to achieve highly interpretable and label-free topic reasoning. Across Weibo and Weibo-21, OVLAFND achieves competitive fixed-domain performance and better leave-one-domain-out generalization. Our code is available at https://github.com/lllfxy/OVLAFND.