feat: keyword-based message routing for pre-turn model selection

[scottgl9]

Add zero-cost pre-turn model routing via keyword matching in the message before any LLM call.

Config:

messageRouting: {
  rules: [
    { match: ["code review", "PR"], model: "claude-sonnet" },
    { match: ["research", "search"], model: "gemini-flash" },
  ],
  default: "qwen35",
}

First-match-wins, case-insensitive, zero LLM cost.

Prototype: scottgl9/nanoclaw feature/message-routing-20260406