Scaling Coding-Agent RL to 32x H100s. 160% Improvement on Stanford's TBench