Loki Mode v7.99.1

Memory retrieval: dimension-mismatch degrades to keyword search

• Vector search no longer crashes on an embedding dimension change
  (memory/retrieval.py): when the embedding engine has fallen back to a
  different model/dimension since a vector index was built, the query vector no
  longer matches the index and VectorSearchIndex.search raises ValueError.
  retrieve_by_similarity now catches that and degrades to keyword search (the
  same honest fallback the staleness path uses) instead of crashing or returning
  wrong-dimension neighbors. Regression test added. (Supersedes the reverted
  embeddings-only signal from the wave-7 batch, which was inert.)