Knowledge Graph Completion (KGC) aims to infer implicit new facts from existing facts in the Knowledge Graphs (KGs). Prevailing path-based reasoning has become a dominant paradigm for its interpretability and multi-hop reasoning ability. However, this approach often overlooks neighborhood interactions and cross-query correlations, resulting in biased subgraph pruning that hinders overall reasoning performance. To address these limitations, we propose MN-Cascade, a multihop neighborhood-aware cascaded reasoning framework designed to enhance semantic propagation in relational KGs. Specifically, MN-Cascade first adopts a multi-layer cascaded reasoning architecture, where each layer consists of three core modules: cascaded propagation, attention fusion, and path pruning. For cascaded propagation, we first perform monodirectional propagation centered on the query entities to obtain preliminary semantic representations of the one-hop entities, and then refine local dependencies through predecessor and successor neighborhoodaware modeling across cascaded views. Subsequently, the attention fusion module integrates cascaded-view representations, followed by a path pruning module that filters out irrelevant entities while preserving promising targets. Finally, a multi-layer fusion optimization module is employed to consolidate hierarchical semantics into a unified embedding for entity prediction. Experimental results demonstrate the superior performance of MN-Cascade compared to prominent baseline methods on transductive and inductive benchmarks.