Backend: - RuntimeAIConfig singleton: mutable in-memory overrides for provider/model - PATCH /admin/api/ai/config: change active provider/model on the fly - GET /admin/api/ai/config: read current runtime state - ai_vision.py: uses runtime_ai_config before falling back to settings - health/status endpoints reflect runtime overrides Frontend: - 'Définir comme actif' button in AI config section - Live feedback message on save