How Sharp and Bias-Robust is a Model? Dual Evaluation Perspectives on Knowledge Graph Completion