[KeyCOrigin]

Summary

In HybridRetriever._retrieve, the vector-search step is wrapped by
recover_read and can be retried when the vector store read fails. Before
this change, every retry called embed_query(query) again, even though the
query text never changes across retries.

This PR caches the query embedding for the duration of a single retrieval:

• Compute embed_query(query) once, before the first vector query.
• Reuse the same embedding on subsequent vector-search retries.
• If the embedding step itself fails, the cache stays empty and the next
  attempt re-embeds (no stale value is reused).

Changes

• datamind/capabilities/kb/providers/hybrid_retriever.py
  • Added a per-retrieval query_vector local that holds the completed
    embedding across recover_read retries.
  • Moved stage = "kb.embedding" to run only when the embedding is
    actually computed.
  • Kept the store query unchanged otherwise (same filters, top-k,
    versioned reads, and fallback behaviour).

Notes

• Scope is intentionally limited to a single retrieval call; no
  cross-request embedding cache is added here.
• No behaviour change for the happy path or for the BM25 fallback path.

Verification

• python3 -m py_compile datamind/capabilities/kb/providers/hybrid_retriever.py
  passes.