Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective