External LLM

Serves the `llm` capability socket from either of two credential sources — a bound connection's MINTED credential when one is bound (`-C llm=<connection>`), or pasted vars otherwise — with no bundled model server. The pasted path is not a legacy fallback: a local Ollama, a LiteLLM/vLLM gateway, an OpenRouter/Groq/Together key and any OpenAI-compatible proxy have no connection type at all, so for them it is the only path. ONE generic socket serves every provider because every provider in the set speaks the same OpenAI-compatible `/v1/chat/completions` wire — the same reason the `s3` connection type serves R2, B2, Wasabi, MinIO and AWS with no per-backend code. Publishes exactly five fields (provider / baseUrl / apiKey / model / region). Presets: OpenAI, Anthropic, Azure OpenAI, Bedrock, Vertex AI, Cloudflare Workers AI, Scaleway, Ollama, and a generic OpenAI-compatible slot.