Argumentative reasoning over long or distributed discourse remains challenging because the number of possible relations between textual units grows quadratically, while true argumentative links remain sparse. As a result, exhaustive relation classification becomes computationally impractical for long documents, multi-document collections, and fragmented discourse. We address this problem with a modular graph-based architecture designed to aggressively reduce the candidate relation space prior to classification. The framework encodes graph-structured discourse representations and applies a two-stage unsupervised filtering pipeline to eliminate unlikely argumentative links before downstream inference. Across experimental settings, the proposed approach removes more than 99% of candidate relations while substantially improving classification performance (up to 6.5× MCC over an unfiltered baseline). These results show that lightweight modular filtering can make large-scale argumentative reasoning possible without sacrificing predictive quality, enabling practical deployment in long-context and distributed-document environments.