Internet memes have become important carriers of online expression, and pun memes, as a typical multimodal category, pose significant challenges to machine understanding due to their reliance on textual ambiguity and cross-modal interaction with visual content. Existing multimodal models for pun meme understanding suffer from classification bias, weak sentiment perception, and performance gaps compared to human comprehension, mainly due to neglecting fine-grained correlation cues for semantic reasoning. To address this, we propose a fusion framework enhanced by multi-source auxiliary cues, including visual salient perception, textual deep semantics, and abnormal pun linguistic features, which are embedded into the multimodal reasoning pipeline. Experiments show that our framework stably optimizes vision-language models, improving accuracy in pun meme detection and implicit sentiment fine-grained analysis.