Summary

POST /api/index returns 200 OK even when every file in the batch fails to embed. Errors are logged to stderr only; the Web UI and CLI both treat the response as success.

Reproduce

With a base install (no [onnx] extra) and embedding.provider=onnx:

POST /api/index HTTP/1.1 200 OK
ERROR Embedding failed for .../MEMORY.md (14 chunks): fastembed is required
      for the ONNX embedding provider. Install it with: pip install memtomem[onnx]
ERROR Embedding failed for .../feedback_ai_attribution_opt_in.md (1 chunks): ...
... (hundreds of errors across every indexed file)

Web UI's stats counter shows 0 chunks but the index button gave no visible failure signal.

Source

packages/memtomem/src/memtomem/engine.py:520 (approx) logs the error and continues. The /api/index route returns its summary without aggregating per-file embedding failures.

Expected

Response body should include a counter for embedding failures (e.g. {indexed: N, embedding_failed: M, errors: [...first few paths]}) so both the Web UI and CLI can surface a loud "indexed but embeddings failed — install memtomem[onnx]" state instead of a clean-looking 200 OK.

Related

#353 fixes the upstream wizard gap that leads users here. This issue covers the downstream error-reporting gap that would remain even if the wizard surfaced everything perfectly (e.g. user who deletes fastembed after setup).