Summary

Fresh gbrain init --pglite on v0.36.3.0 lands in a state where embedding_model config says openai:text-embedding-3-large but the embed pipeline tries ZeroEntropy (zembed-1 1280d), causing every page to fail with expected 1536 dimensions, not 1280.

The state is sticky — wiping ~/.gbrain/brain.pglite and re-running init does NOT reset it. Same error persists even after unset ZEROENTROPY_API_KEY.

Environment

• gbrain version: 0.36.3.0
• Repo commit: 1d5f69f
• bun: 1.x (installed via Homebrew)
• macOS: 26.3.1 (M-series Mac)

Steps to reproduce

git clone https://github.com/garrytan/gbrain.git ~/sandbox/gbrain
cd ~/sandbox/gbrain
bun install && bun link

export OPENAI_API_KEY=sk-proj-...
export ZEROENTROPY_API_KEY=ze_...

gbrain init --pglite
gbrain import /some/markdown/dir/ --no-embed   # 31 files, 234 chunks ok
gbrain embed --stale

Result:

Error embedding entities/nvda: expected 1536 dimensions, not 1280
Error embedding concepts/ai-model-release-cycle: expected 1536 dimensions, not 1280
...
[embed.pages] 31/31 (100%)
Embedded 0 chunks across 31 pages

Investigation

• gbrain config get embedding_model returns text-embedding-3-large (no provider prefix)
• gbrain config get embedding_dimensions returns 1536
• gbrain config set embedding_model openai:text-embedding-3-large succeeds, but embed pipeline still tries ZE
• gbrain providers list shows openai + zeroentropyai both ready
• gbrain doctor warns embedding_provider: zeroentropyai:zembed-1 responds (925ms, 1280 dims) but: DB dimension mismatch: column is vector(1536) but provider returns 1280-dim

Workarounds tried (none worked)

Suspected root cause (from reading source)

src/core/retrieval-upgrade-planner.ts defines a ZE-as-default switch flow (v0.36.0.0). On fresh init when OPENAI_API_KEY is set, isLegacyDefault=true triggers zeSwitchOffered=true, but ze_switch_applied is never written, leaving the embed pipeline in a half-migrated state where it routes to ZE_TARGET_EMBEDDING_MODEL even though gbrain config get embedding_model returns the OpenAI default.

loadConfigWithEngine in src/core/config.ts doesn't merge embedding_model from the DB plane (only embedding_multimodal, embedding_columns, search_embedding_column, etc.), so user config set embedding_model writes to the DB but the embed code reads the gateway default.

Impact

Cannot complete fresh PGLite install following INSTALL_FOR_AGENTS.md.

Suggested fix directions (no PR yet)

1. loadConfigWithEngine should merge embedding_model from DB plane the same way other embedding_* keys are merged.
2. gbrain init --pglite should either complete the ZE switch (if ZE key set) or hard-refuse it (if dim mismatch), not leave half-state.
3. gbrain ze-switch should print what it's doing (currently silent, no diagnostic output).
4. gbrain init --embedding-dimensions N flag should be honored or removed from CLI help.