Migrate the summariser behind a regression suite

The procedure is the lesson's: run the old prompt on both models first, then compare case by case, then make one change and run again. The decision that matters most is that the gate reads tiers and trials, not the average: control and safety cases must pass every trial on the new model, no regression case that passed before may fail, and capability cases are reported but don't gate. With that rule, a model that is better on average can still be blocked, and here it is.