feat: pluggable embedding providers (OpenAI + Gemini)

[aloysiusmartis]

Summary

• Adds a provider-agnostic EmbeddingProvider interface so gbrain can use Gemini (gemini-embedding-001, 1–3072 Matryoshka dims) or OpenAI (text-embedding-3-large, 1536 dims) interchangeably
• Public embed/embedBatch API in embedding.ts is unchanged — existing call sites see no diff
• Fixes a critical silent bug: operations.ts put_page had !process.env.OPENAI_API_KEY hardcoded, so Gemini users got no embeddings on every page import

Architecture

src/core/embedding-provider.ts      — EmbeddingProvider interface + factory
src/core/providers/openai-embedder.ts  — OpenAI impl (extracted from embedding.ts)
src/core/providers/gemini-embedder.ts  — Gemini impl with Matryoshka dim support
src/core/providers/retry-utils.ts      — shared exponentialDelay + sleep
src/commands/migrate-provider.ts       — gbrain migrate --provider openai|gemini

getActiveProvider() reads GBRAIN_EMBEDDING_PROVIDER env and returns the right singleton. isEmbeddingAvailable() replaces all !process.env.OPENAI_API_KEY checks so hybrid search and page ingestion work correctly regardless of which provider is active.

New command: gbrain migrate --provider

gbrain migrate --provider gemini               # migrate existing brain
gbrain migrate --provider gemini --dimensions 512  # Matryoshka sub-dim
gbrain migrate --provider openai               # migrate back to OpenAI
gbrain migrate --provider gemini --dry-run     # preview only

• ALTER TABLE only when dims actually change (same provider/same dims = re-embed only)
• Re-embeds all chunks with the new provider after schema change
• Updates config_table + ~/.gbrain/config.json
• Remote guard: CLI-only, errors if called via MCP

Init-time provider selection

gbrain init --provider gemini               # new Gemini brain (768 dims default)
gbrain init --provider gemini --dimensions 512
gbrain init --provider openai               # explicit OpenAI (1536 dims default)

Schema

getPGLiteSchema(dims, model) replaces hardcoded vector(1536) in PGLite DDL so gbrain init --provider gemini creates a vector(768) schema from the start.

Config persistence

GBrainConfig gains embedding_provider and embedding_dimensions. loadConfig() propagates them to env vars at startup so subsequent sessions use the same provider without repeating the flag.

Tests

• test/embedding-provider.test.ts — 22 unit + 3 live (skipped without API key): factory, fallback, unknown provider error, boundary dims
• test/pglite-schema-provider.test.ts — 6 tests for getPGLiteSchema() substitutions
• test/config-embedding-provider.test.ts — 4 tests for env-var propagation (no-override behavior)
• test/migrate-provider-args.test.ts — 8 tests for dims-change logic and API key guard

All 2627 unit tests pass (bun test, 0 fail).

Relation to PR #197

This overlaps with trymhaak's voyage embedding PR (#197). The approach here uses an interface/factory pattern (EmbeddingProvider) so adding future providers (voyage, cohere, local) is additive — no call site changes. The factory is a single getActiveProvider() call; all embed/embedBatch callers go through embedding.ts unchanged.

If #197 lands first, this PR could subsume it by adding a VoyageEmbedder to src/core/providers/. Happy to coordinate.

Checklist

• Public API unchanged (embed, embedBatch, EMBEDDING_MODEL, EMBEDDING_DIMENSIONS in embedding.ts)
• All existing tests pass (2627/2627)
• No // FORK: comments in this diff
• COORDINATION.md and fork-only scripts not included
• Remote guard on migrate --provider (MCP callers get a clear error)
• NaN guard on --dimensions arg parsing