Hard-wiring of `fx` to Vercel AI Gateway, preventing host-configured custom AI model providers for `libfx` and WASM embeddings.

## Problem

Production fx is hard-wired to [Vercel AI Gateway](vercel.com/docs/ai-gateway `fx login`, `fx setup`, and `AI_GATEWAY_API_KEY` all talk to `ai-gateway.vercel.sh` using the Gateway language-model protocol, not OpenAI Chat Completions.

`FX_GATEWAY_CHAT_URL` exists, but it only accepts the canonical Gateway URL or **loopback HTTP**. That is enough for a local shim during CLI development. It is not enough for:

- a **custom / third-party provider** (OpenAI-compatible proxy, self-hosted vLLM/Ollama, company gateway)
- **`libfx` / `fx-term.wasm` embedders**, especially browser and extension hosts, which cannot realistically run a loopback Gateway adapter next to the WASM module

I am embedding `createFxTerminal()` from `libfx/browser` in a Chrome extension. The only credential I can pass today is `env.AI_GATEWAY_API_KEY`. There is no host-trusted way to say “use this provider, this base URL, this key, this model.” Intercepting `fetch` and translating Gateway LM-spec v4 into another protocol is possible but fragile and unofficial.

## Request

Add a first-class, **host-trusted** custom provider — not repository config.

Suggested shape (names flexible):

```js
await createFxTerminal({
env: {
FX_PROVIDER: 'openai-compatible',
FX_MODEL: 'deepseek-chat',
OPENAI_BASE_URL: 'api.example.com/v1&
OPENAI_API_KEY: '…',
},
});
```

Native CLI equivalent could be the `fx setup openai-compatible` / `FX_PROVIDER` surface already sketched...