Fix Prompt Optimization Ollama access from sandbox

[Copilot]

Prompt Optimization prepared Ollama on the runner, but the sandboxed agent was told to use 127.0.0.1, which resolves inside the sandbox and caused missing-tool reports for the GGUF model server.

• Sandbox networking

  • Add host.docker.internal to the workflow allowlist.
  • Keep port 11434 exposed through the existing sandbox host-port configuration.

• Ollama binding

  • Start Ollama on the runner interface so it is reachable from the sandbox:

    OLLAMA_HOST=0.0.0.0:11434 ollama serve

• Optimizer endpoint

  • Point both model evaluation URLs at the runner-hosted Ollama server:

    --eval-url http://host.docker.internal:11434/v1
    --ios-eval-url http://host.docker.internal:11434/v1
    

• Generated workflow

  • Recompile prompt-optimization.lock.yml from the updated workflow source.

• Fixes [aw] Prompt Optimization is missing required tool #255