Make embedding provider, model, and dimensions configurable

[100yenadmin]

Summary

GBrain currently hardcodes OpenAI text-embedding-3-large at 1536 dimensions in the shared embedding service.

Why this matters

• Users may want to run a different provider, like Voyage AI, without forking the code.
• Different embedding models have different native or preferred output dimensions.
• Making provider/model/dimensions config-driven keeps the default path intact while enabling cheaper, higher-quality, or infrastructure-specific setups.
• This also helps future engine work because chunk metadata can track the chosen embedding model cleanly.

Proposed direction

• Keep the default behavior exactly as-is: OpenAI text-embedding-3-large, 1536 dims.
• Add explicit env-driven overrides such as EMBEDDING_PROVIDER, EMBEDDING_MODEL, and EMBEDDING_DIMENSIONS.
• Support Voyage via VOYAGE_API_KEY.
• Keep docs minimal and upstream-friendly.

Notes

This is intentionally a small compatibility-preserving change, not a broader embedding subsystem redesign.

[nbzy1995]

PR #172 implements this exact request — pluggable EmbeddingProvider interface with OpenAIProvider (default, identical behavior) and OllamaProvider (free, local). Schema dim is templated by the resolved provider so gbrain init --provider=ollama --model=nomic-embed-text produces a vector(768) brain. All defaults preserved; 883 tests pass.

Tested end-to-end on a real personal corpus (447 pages, 925 chunks, 742 embeddings via local Ollama) — vector queries return 1.0 cosine on topic matches.

Designed so adding Voyage / Gemini / Cohere = one new file in src/core/embedding/providers/. Cc #149.

[100yenadmin]

also @garrytan you really should switch to voyage-4-large default to make gbrain easier to adopt/use for norms it's smaller vector but vector size is not correlated with quality retrieval/speed/efficiency and it works just fine I'm using it.

It's the best model right now except new Gemma and you get 200k~ usages a month free making it free for 90%+ of users. not affiliated with them, just easier process for your users to setup free. also the default to anthropic is meh as it's expensive and API is now trash with 50% increase in cost on "5%+ better model.

Switch to default gpt 5.4 mini as it's free for 90%+ of your users by default and connect it into the OpenAI Codex plugin so it doesn't need a hardwire. I'll build it out for you if you give me the go ahead PR wise.

The goal here (you know this so I'm preaching to preacher and choir) is less friction for adoption and customization/recommended specs by you if you really prefer those spec in advanced config or "Garrys stack".