Auto-optimizing Jev: half the errors, 1/7 the cost

We pointed our autoresearch loop at Jev, a classifier that can't be fine-tuned. It ended up making half as many mistakes as a reasoning model, for a seventh of the cost and a thirtieth of the latency.