feat(runtime): first-class router-tools executor backend (off-box tool use)

[drewstone]

What

createExecutor gains backend: 'router-tools' — a real agentic loop over the Tangle router's tool-calling, off-box (no sandbox). Each turn passes tools; the model's tool_calls run via the seam's executeToolCall on this host and fold back as tool messages, repeating until the model answers or maxTurns (one turn = one completion = the equal-compute unit vs random@k).

This is the multi-turn capability the single-shot routerInlineExecutor couldn't express, and it routes around the sandbox→router egress block (#984). executeToolCall(name, args, task) receives the task so per-task tool surfaces (e.g. a gym keyed by task) dispatch correctly. RouterToolsSeam + ToolSpec exported via /loops.

createExecutor({
  backend: 'router-tools',
  routerBaseUrl, routerKey, model: 'gpt-4o-mini',
  tools: [weatherTool], maxTurns: 4,
  executeToolCall: async (name, args, task) => runMyTool(name, args),
})

Why

The chat-only router backend can't do agentic work (it was the wrong worker for any tool-using domain — e.g. EOPS). This makes the tool-using off-box worker first-class, symmetric with the other createExecutor backends, so any adapter with a tool surface gets it for free.

Verified live

createExecutor({backend:'router-tools', tools:[get_weather], …}).execute("weather in Paris?") → called get_weather({city:"Paris"}), ran the host tool, answered "The weather in Paris is clear with a temperature of 18°C." in 2 turns.

Test

typecheck + build clean (router-tools exported in dist/loops.d.ts); lint clean; live end-to-end proof above.