Rerun AI

One layer from collection to training. Stream dataset mixes directly to GPUs. No preprocessing jobs, no stale copies, no idle GPUs.