Dump the trace of any rollout that scores zero: one turn and a zero is not the same as solving it wrong

Add the task screener: measure base vs trained instead of guessing