Accurate surgical phase recognition in laparoscopic pulmonary lobectomy is essential for workflow optimization, intraoperative guidance, and surgical training. However, high procedural variability and visual similarity between key phases (vessel dissection and bronchus dissection) pose significant challenges. To address this, we target a subset of fine-grained steps selected for their clinical relevance and distinct visual features. We propose a unified framework that integrates global temporal modeling and local feature refinement, capturing long-range procedural dependencies and fine-grained spatial dynamics. We also construct a clinically validated dataset covering diverse lobectomy subtypes and intraoperative variations with standardized annotations. By modeling temporal continuity and refining local anatomical features, our method enables robust recognition across complex surgical scenes. Experiments show that it achieves 92.15% accuracy, surpassing prior state-of-the-art methods across several major metrics. The proposed framework shows potential for supporting future research on real-time surgical workflow analysis and remote surgery assistance.