Efficient expert merging for MoE-based graph foundation models approaching ensemble-level performance.