“Large language models exhibit 'context anxiety,' where they tend to give up on complex problems or take overly simplistic approaches when their context window exceeds 100k-200k tokens, even if the advertised limit is much higher.”