Enhance vector index handling with improved validation and background repair mechanisms

docs/core-concepts/storage.md

@@ -59,7 +59,7 @@ Those projections include:
 
 This is why search and recall can continue to work even when one retrieval path is incomplete. The storage layer keeps multiple doors into the same memory.
 
-Semantic vectors are guarded by the retrieval text hash. The durable vector blob lives in `target_embeddings`; sqlite-vec rows are tracked separately in `vector_index_entries` and copied into `.konteks/vectors.sqlite`. The acceleration index can be rebuilt from the durable blob. If text changes, old vectors are not used for scoring; changed extraction repairs missing or stale vectors for changed sections and modules without rebuilding unrelated retrieval documents. Corpus-wide embedding generation and sqlite-vec repair run during `konteks init` and `konteks rebuild`, keeping automatic warm-up and save refresh work proportional to changed files.
+Semantic vectors are guarded by the retrieval text hash. The durable vector blob lives in `target_embeddings`; sqlite-vec rows are tracked separately in `vector_index_entries` and copied into `.konteks/vectors.sqlite`. The acceleration index can be rebuilt from the durable blob. If text changes, old vectors are not used for scoring; changed extraction repairs missing or stale vectors for changed sections and modules without rebuilding unrelated retrieval documents. Corpus-wide embedding generation and sqlite-vec repair run during `konteks init` and `konteks rebuild`, keeping automatic warm-up and save refresh work proportional to changed files. Search validates sqlite-vec against durable vectors once per process. When native rows are incomplete, search returns the bounded durable-vector fallback immediately and repairs sqlite-vec in the background.
 
 Konteks loads sqlite-vec through Bun's SQLite extension loader when available, then falls back to Node's built-in SQLite loader. If neither runtime can load the native extension, indexing reports an actionable dependency error.
 

docs/development/database-erd.md

@@ -247,7 +247,7 @@ but their table options live in `src/database/utils/migrations`.
 `memory_fts`.
 
 Semantic retrieval uses sqlite-vec virtual tables created per embedding dimension through Bun's SQLite extension loader when available, then falls back to Node's built-in SQLite extension loader. `target_embeddings` remains the
-durable compatibility store and exact fallback source. `vector_index_entries` records which durable embeddings have been copied into the sqlite-vec virtual table for their dimension in `.konteks/vectors.sqlite`. A vector is considered fresh only when its model-aware `embedding_hash` matches the current retrieval text and the sqlite-vec metadata matches the same hash; stale or missing index rows are repaired from stored blobs during embedding generation and ignored during retrieval.
+durable compatibility store and exact fallback source. `vector_index_entries` records which durable embeddings have been copied into the sqlite-vec virtual table for their dimension in `.konteks/vectors.sqlite`. A vector is considered fresh only when its model-aware `embedding_hash` matches the current retrieval text and the sqlite-vec metadata matches the same hash. Search validates native rows against durable vectors once per process; unhealthy groups use bounded exact fallback immediately and schedule best-effort background repair from stored blobs.
 
 ## Main Domains