fix: eliminate API leaks in training loop, set 30k steps for A100

perf: force ST to CUDA, disable LLM tasks, reduce to 100k steps

SpindleFlow RL — periodic push + log persistence