Как организовать multi-model routing и fallback в production: LiteLLM, OpenRouter, Portkey?

LiteLLM = Python router с unified API + fallback / retry / rate-limit (self-host или embedded). OpenRouter = HTTP gateway, openrouter/auto + models: [...] для cross-provider fallback. Portkey = config-driven routing + semantic cache + observability. Всегда: timeout, adapter для разных response shapes, classifier-based semantic routing, context_window mismatch check при fallback.