feat: support Gemini Embedding 2 as native multimodal embedding provider

[lost9999]

Problem

GBrain v0.37 has a useful multimodal embedding path, but the native Google recipe only exposes gemini-embedding-001 as a text embedding model. embedMultimodal() currently rejects the Google recipe because supports_multimodal is not set, and the error message points operators to Voyage.

Google has shipped Gemini Embedding 2 / gemini-embedding-2-preview, a native multimodal embedding model that maps text, images, video, audio, and documents into a single embedding space.

Evidence

Google docs: https://ai.google.dev/gemini-api/docs/embeddings

• gemini-embedding-2 / gemini-embedding-2-preview supports multimodal input: text, images, video, audio, PDF.
• Output dimensions support 128-3072, recommended 768/1536/3072.
• Current GBrain Google recipe only lists gemini-embedding-001 under embedding.
• Current embedMultimodal() error says: Today: voyage:voyage-multimodal-3.

Local validation

LiteLLM proxy in front of Gemini can expose gemini-embedding-2-preview and text embeddings work:

POST /v1/embeddings model=gemini-embedding-2-preview input="hello world"
→ OK, 3072 dims

But the OpenAI-compatible multimodal content-array shape that GBrain sends through embedMultimodalOpenAICompat() fails via LiteLLM:

POST /v1/embeddings
input=[{type:"image_url", image_url:{url:"data:image/png;base64,..."}}]
→ 400 INVALID_ARGUMENT
Invalid value at 'requests[0].content.parts[0]' (text), Starting an object on a scalar field

So the LiteLLM path is not currently sufficient as a drop-in bridge for Gemini multimodal embeddings.

Desired behavior

Allow configs like:

gbrain config set embedding_multimodal true
gbrain config set embedding_multimodal_model google:gemini-embedding-2-preview
# or google:gemini-embedding-2 when stable

Then image ingestion / gbrain reindex --multimodal / cross-modal query paths should use Google native multimodal embeddings instead of requiring Voyage.

Why this matters

• Single-vendor deployment for users already on Gemini.
• Native all-modal support: text, images, video, audio, documents.
• Avoids Voyage free-tier 429s during vault image ingestion.
• Aligns with GBrain's provider-agnostic embedding architecture.

Implementation sketch

• Extend src/core/ai/recipes/google.ts with Gemini Embedding 2 model(s).
• Add Google native embedMultimodal request serialization for models/*:embedContent with parts containing text and inline image data.
• Handle output dimensions consistently with GBrain schema / embedding_image and embedding_multimodal columns.
• Add tests for text-only, image-only, text+image request payloads and model validation.