Run evals and A/B test prompts

Manage prompts as versioned objects, split real traffic across versions with weighted experiments, and run evals on the resulting traces. getPrompt picks a variant, the served version rides along on every trace, and you compare quality, latency, and cost side by side without a redeploy.

• ▸Weighted variants — ship a new wording to 10% of traffic before 100%.
• ▸selectedVariant tells you exactly which version served each request.
• ▸Run LLM-as-judge evals over the exact production traces each version generated.
• ▸Link promptVersion onto generations so the split shows up in traces, cost, and eval results.
• ▸Promote or roll back instantly by moving a label — no deploy.