v0.4.4

Fixed

• nitrite_vector — a selective metadata filter could silently return zero hits.
  RagStore::search(...).filter(...) fetched a fixed k * oversample (default 4) nearest
  neighbours from the ANN index and applied the metadata filter afterwards. That works when
  the filter keeps a large fraction of the index, but fails completely when it is selective —
  the common RAG shape, "search only this document's chunks": every one of the k * oversample
  candidates can fail the filter, and the search returns nothing while matching documents sit
  further out. Measured on a real index: 41 records, one chunk each, k = 5, filtered to a
  single record, query semantically closer to the other 40 — 0 hits. The failure scales
  with index size and is invisible in small-index testing. oversample was never a fix, only a
  knob: the multiple needed is index_size / matching_docs, which the caller cannot know in
  advance.

  The traversal now widens and re-queries until k matching hits are found or the index is
  exhausted, turning a silently wrong answer into a correct one; oversample becomes the
  starting window rather than a hard ceiling. A filter matching very few documents therefore
  costs a full index scan — for the single-document case an exact scan over that document's
  chunks (collection().find(meta_filter)) remains both faster and more accurate than any ANN
  path. Widening deliberately does not apply to min_score: a score cutoff is monotone in
  the ANN ranking, so the hits it drops are the tail and fetching further out cannot recover
  any.