feat: support provider_base_urls for native-openai path

[leoge007]

Problem

When provider_base_urls.openai is configured in ~/.gbrain/config.json (e.g. pointing to SiliconFlow, Azure OpenAI, vLLM, or any OpenAI-compatible endpoint), gbrain still sends requests directly to OpenAI's API for the native-openai recipe. The base_urls config is only respected by the openai-compat recipe tier.

This means users who want to use OpenAI-compatible embedding/chat/expansion providers through the native-openai path cannot do so, even though provider_base_urls is a documented config field in GBrainConfig.

Changes

1. gateway.ts — Pass baseURL to createOpenAI() in embedding path

The expansion and chat native-openai paths already read cfg.base_urls?.[recipe.id] and pass it as baseURL. The embedding path was missing this. Now all three paths consistently route through the configured base URL.

Also passes hasCustomBaseUrl to assertTouchpoint() and dimsProviderOptions() so downstream logic can adapt.

2. model-resolver.ts — Relax model whitelist when custom base URL is set

assertTouchpoint() now accepts a hasCustomBaseUrl parameter. When true, it skips the model-name whitelist check for native-tier recipes. This is safe because:

• A custom base URL means the user is explicitly routing to a non-OpenAI endpoint
• Such endpoints may offer models not in the OpenAI recipe list (e.g. Qwen/Qwen3-Embedding-8B)
• The user has already opted into this by configuring provider_base_urls

3. dims.ts — Pass dimensions when using custom base URL

dimsProviderOptions() now accepts a hasCustomBaseUrl parameter. When true for the native-openai path, it passes the dimensions parameter unconditionally (not just for text-embedding-3-*). This is needed because:

• OpenAI-compatible endpoints (SiliconFlow, Azure OpenAI, etc.) typically support the dimensions parameter
• Embedding models like Qwen3-Embedding support Matryoshka dimension reduction
• Without this, dimension mismatches occur (e.g. model returns 4096 dims but schema expects 1536)

Usage Example

// ~/.gbrain/config.json
{
  "engine": "pglite",
  "provider_base_urls": {
    "openai": "https://api.siliconflow.cn/v1/"
  },
  "embedding_model": "openai:Qwen/Qwen3-Embedding-8B",
  "embedding_dimensions": 1536
}

Then set OPENAI_API_KEY to your SiliconFlow API key, and gbrain will route all OpenAI requests through SiliconFlow with correct base URL, model name, and dimension handling.

Testing

• bun run verify passes (tsc, privacy checks, admin build, all linting)
• Verified end-to-end: gbrain embed successfully embeds pages using SiliconFlow + Qwen3-Embedding-8B → 1536 dimensions

Scope

This PR only changes behavior when provider_base_urls is explicitly configured. Default behavior (routing directly to OpenAI) is completely unchanged.

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews