Structure-Aware Corpus Construction and User-Perception-Aligned Metrics for Large-Language-Model Code Completion