thaw-vllm

The fork primitive for LLM inference. Snapshot a running session — weights + KV cache + scheduler state — and hydrate it into N divergent children that skip prefill. For RL rollouts, parallel coding agents, agent branching. Supports vLLM and SGLang.