Although automatic fake news detection has achieved significant progress, unseen fake news detection remains a major challenge. First, they lack generalizable representations of deception patterns, resulting in poor performance on unseen fake news. Second, the scarcity of comment data during early or breaking events weakens the available supervision signals. Third, sentiment cues, though robust and largely domain-independent, have long been underestimated. To address these challenges, we propose a unified detection framework that fully leverages the generative and reasoning capabilities of large language models (LLMs). The framework consists of two complementary paths. The first performs cross-domain variant generation, synthesizing multi-domain samples centered on the same deceptive logic to guide the model in learning consistent deception patterns, thereby enhancing its ability to detect unseen misinformation. The second performs controllable comment generation, producing comments with adjustable stance and sentiment to provide sentiment-aware supervision when comment data are scarce. These two paths work together under a shared training objective: cross-domain variants offer a stable semantic context for comment generation, while sentiment supervision further reinforces the alignment of generalizable deception cues. Experimental results show that our framework outperforms existing methods for unseen fake news detection.