The rise in online information has increased the spread of misinformation and disinformation, which makes claim detection important for the reliability of digital content. Claim detection refers to the problem of determining whether a factual statement is real or fake. While there has been great advancement in high-resource languages, Arabic is particularly challenging due to its special characteristics and the low availability of annotated resources. In this paper, we make use of the Arabic News Stance (ANS) corpus and develop a multi-stage framework to help tackle these problems. Our method effectively models the structural interplay between claims and words with a carefully heterogeneous bipartite graph and the use of advanced Graph Neural Networks (GNNs). In parallel, we used two Small-scale Language Models (SLMs) in this domain, namely LLaMA-3.2-1B and Qwen2.5-3B, and demonstrated that together with QLoRA-based fine-tuning, these lightweight models can be as effective as or even better than heavier baselines. To this point, we suggested a novel neuro-symbolic fusion paradigm that unifies structural reasoning with linguistic prior information. By integrating the outputs from a Graph Convolutional Network (GCN) with the fine-tuned Qwen2.5-3B logits, we achieved the state-of-the-art performance on the ANS dataset with an accuracy of 0.732 and an F1-score of 0.717. This framework shows that the combined results of structural graph reasoning and SLMs lead to accurate and computationally efficient claim detection. This work not only enhances the detection of Arabic misinformation but also provides a scalable solution for low-resource languages.