hybrid. Local-first LLM router: answer the easy majority on a small local model, escalate only the hard queries to any OpenAI-compatible frontier endpoint. ~160 lines, stdlib-only.