Summary

• add optional Gemini embeddings support alongside the existing OpenAI path
• keep embeddings at 1536 dimensions so existing GBrain vector schema remains compatible
• document the new provider setup and add focused config tests

What changed

• add embedding_provider, embedding_model, and gemini_api_key support in config loading
• make the embedding service provider-aware (openai or gemini)
• use Gemini batchEmbedContents with outputDimensionality: 1536
• normalize Gemini vectors before storing them
• let hybrid search run whenever any embedding provider is configured
• persist the embedding model used on chunks during import/embed flows
• add gbrain init --provider openai|gemini --key <api_key> support
• update README/help text for Gemini usage
• add targeted tests around provider selection/config overrides

Why this is safe

• OpenAI remains the default behavior
• no schema changes are required
• Gemini is configured to emit 1536-dimensional embeddings, matching the current vector size

Verification

• bun test test/embedding-provider.test.ts test/config.test.ts test/cli.test.ts
• live local validation against a real Gemini API key:
  • direct embedding call returned 1536-dimensional vectors
  • embed --stale completed successfully on a local PGLite brain
  • semantic search worked after embedding

Notes

• repo-wide TypeScript already has some unrelated pre-existing errors outside this change, so verification here is focused on the affected paths.