eval: LLM-as-judge qrels + real ablation numbers (nDCG@10 0.89)

feat: initial LectureLens — hybrid RAG learning copilot (phases 0-7)