Add prime-rl training cookbook

[AnandK27]

Summary

• Adds a new cookbook (cookbook/prime-rl-training/) that bridges SimLab trajectory collection with Prime Intellect's prime-rl for RL training of agent models
• Full pipeline: collect tool-use trajectories from SimLab environments → convert to SFT datasets → build and push a verifiers environment → run hosted RL training via prime rl run
• Includes trajectory converter, verifiers environment wrapper, CLI collector, SFT/RL configs, human guide, and agent SKILL.md
• Ships with 3 example customer support tasks (billing disputes, API escalation, SLA issues)
• Verifiers environment pushed and validated on Prime Intellect hub as collinear-simlab/simlab-tasks
• End-to-end tested: RL training run completed 50 steps on Qwen/Qwen3.5-9B (run p8zaxiqbge02roi3g89o7wdg)

Test plan

• uv sync installs all dependencies from simlab repo
• simlab env init + simlab tasks-gen generate environment and tasks
• python -m prime_rl_training.collect sft converts artifacts to SFT dataset
• Verifiers environment loads correctly (load_environment() returns SingleTurnEnv)
• Environment pushed to Prime Intellect hub (prime env push)
• RL training run completed on Prime Intellect platform (prime rl run)
• Checkpoints produced (5 checkpoints at steps 25-45)

🤖 Generated with Claude Code

[aditj]

• Is there a reason to inclyde the 3 sample generated tasks?
• How should we document the experiment you did for 30 steps? any reward curves we can post on X