feat: add LiteLLM provider with /v1/models discovery

[prodmanpd]

Summary

Adds LiteLLM as a first class provider (provider: "litellm"), so one endpoint reaches 100+ models (OpenAI, Anthropic, Gemini, Bedrock, Vertex, Azure, Groq, and more) through a self hosted LiteLLM proxy.

A LiteLLM proxy speaks the OpenAI compatible wire format, so, per this repo's own design ("adding a wire format is one new adapter"), the adapter reuses OpenAIChatModel's translation and only changes the defaults. The value over provider: "openai" plus a manual baseUrl is a self describing provider:

• a sensible default proxy address (http://localhost:4000), and
• /v1/models auto discovery (listLiteLLMModels) so callers can enumerate whatever models the proxy is configured to serve instead of hardcoding a model id.

Changes

• src/ai/llm/litellm.ts (new): LiteLLMChatModel extends OpenAIChatModel (default proxy base URL and litellm:<model> label) and listLiteLLMModels() which calls GET /v1/models and returns the available model ids.
• src/ai/llm/factory.ts: ProviderKind gains "litellm"; createChatModel builds a LiteLLMChatModel.
• src/ai/providers.ts: default model for the litellm provider.
• test/litellm-adapter.test.ts

Tests

Types: npx tsc -p tsconfig.json --noEmit is clean.

Tests: npm test (repo runner): 840 passed, 0 failed (4 new litellm tests plus no regressions).

test/litellm-adapter.test.ts
  ok litellm: reuses OpenAI-compatible translation and labels as litellm
  ok litellm: /v1/models auto-discovery returns the proxy's model ids
  ok litellm: factory builds a LiteLLMChatModel for provider 'litellm'
  ok litellm: exposes a default proxy base URL

The new tests assert: the adapter forwards the provider prefixed model and temperature and reuses the OpenAI translation; listLiteLLMModels maps /v1/models to ids; and the factory wires provider: "litellm".

Usage

import { createChatModel } from "./ai/llm/factory.js";
import { listLiteLLMModels } from "./ai/llm/litellm.js";

// Chat through a LiteLLM proxy (defaults to http://localhost:4000)
const model = createChatModel({
  provider: "litellm",
  apiKey: process.env.LITELLM_API_KEY ?? "",
  model: "anthropic/claude-3-5-sonnet", // any model the proxy serves
});

// Discover what the proxy offers
const models = await listLiteLLMModels({ apiKey: process.env.LITELLM_API_KEY ?? "" });

Or via env: GRAFT_PROVIDER=litellm, GRAFT_MODEL=anthropic/claude-3-5-sonnet, GRAFT_BASE_URL=http://localhost:4000.