I would bet on the long-context half. A model that degrades with length and a model that degrades with the number of things it has to hold consistent look identical from outside, and only one of those gets fixed by a bigger window. For the report's reading to hold, degradation has to keep tracking input length after everything load-bearing has been moved to the front. I have not seen that separated out. So the fixes and the diagnosis are allowed to be right for the wrong reason.