Self-hosted LLM gateway: one OpenAI-compatible endpoint, sovereign French models by default, controlled escalation.