fix(rag): guard against None embeddings in LlamaIndex pipeline

[kagura-agent]

Summary

Fixes #346 — RAG queries crash with TypeError: unsupported operand type(s) for *: 'NoneType' and 'float' when a stored embedding vector is None.

Root Cause

When an embedding provider returns {"embedding": null} for a chunk, two things go wrong:

1. _extract_embeddings_from_response uses item.get("embedding", []) — but dict.get() only returns the default when the key is absent, not when the value is explicitly None. So None passes through.
2. CustomEmbedding._get_text_embeddings trusts the result without validation, allowing None vectors to be stored in the index and crash np.dot at query time.

Fix

Two-layer defense:

1. Adapter layer (openai_compatible.py): Changed item.get("embedding", []) → item.get("embedding") or [] so explicit None values are caught.
2. Pipeline layer (llamaindex.py): Added post-embed validation in _get_text_embeddings — any None vectors are replaced with zero vectors and logged as errors. This prevents silent storage corruption regardless of which adapter is used.

Testing

• Added unit test for None embedding extraction in test_extract_embeddings.py
• All 22 extraction tests pass
• All 45 passing embedding/RAG tests still pass (10 pre-existing failures due to missing async plugin — unrelated)