Multiple Kernel Clustering (MKC) aims to improve clustering performance by integrating complementary information from candidate kernels. Among existing MKC methods, late fusion MKC (LFMKC) offers superior scalability by aggregating clustering partitions rather than full kernel matrices. However, its reliance on fixed base partitions often leads to suboptimal representations and degraded clustering performance. To address this, we propose Late Fusion MKC Refined via Optimal Linear Graph Filtering (LFMKC-OLF), which enhances partition quality while preserving linear computational complexity. Specifically, bipartite graphs are constructed for each base partition, upon which high-order low-pass filters based on heat kernel diffusion and complementary high-pass filters are designed to capture both global consistency and fine-grained structural details. A consensus filtering mechanism is introduced by optimally combining view-specific filters to refine multi-scale representations. Furthermore, a joint clustering objective integrates both smoothed and detail-enhanced partitions to effectively mitigate over-smoothing. Extensive experiments on eight benchmark datasets demonstrate that LFMKC-OLF consistently outperforms 12 state-of-the-art LFMKC methods, while maintaining high computational efficiency for large-scale clustering tasks. Our code is publicly available at: https://github.com/sxuHui/LFMKC-OLF .