v0.97.0

Added

• Graph retrieval without loading an index: IGraphChunkRetriever. RetrieveAsync(query, GraphRetrievalOptions) returns the
  chunks the persisted entity graph connects to the entities a query names — seeds looked up by exact name in one store round
  trip (word runs, possessives, and substrings of CJK/Hangul words so particles do not hide a name), a bounded walk of
  MaxHops around them (HopDecay per hop), scored per chunk by the entities it mentions — within the same metadata
  Filters the vector and keyword legs take (a document scope also cuts entities before chunks are fetched). Every graph
  store registration registers it (AddGraphChunkRetrieval).
• The persisted graph as a third leg of a scoped hybrid search: HybridSearchOptions.GraphOptions. Set a
  GraphLegOptions (the retriever's options plus Weight, default 0.3, added on top of the vector and keyword weights) and
  IHybridSearchService — the core service and the Qdrant service alike — fuses the graph's chunks with the other two
  legs under the same Filters (the leg's own Filters win when set). Rows carry GraphScore and GraphRank
  (int.MaxValue when the graph did not find the row) and AppliedFusion.GraphWeight. RRF, WeightedSum,
  RelativeScoreFusion and Maximum take the third leg; Product or HarmonicMean set with it throws ArgumentException, and a
  service-chosen one becomes RRF. Asking for the leg without a registered graph store throws InvalidOperationException.
  A failing graph backend empties the leg like the other two; leaving GraphOptions null changes nothing.

Changed

• Breaking: SearchSource is a [Flags] enum with explicit values — None = 0, Vector = 1, Sparse = 2,
  Both = 3, Graph = 4 (before: Vector = 0, Sparse = 1, Both = 2). A row the graph leg also found is
  Vector | Graph and so on. Migration: compare with HasFlag (or & SearchSource.Both) where the graph leg can be on,
  and re-read any stored numeric value.
• SQLite graph traversal reads what it visited in two queries instead of one per entity and one per relationship
  (IGraphStore.TraverseAsync), which the graph leg calls once per seed.
• The Qdrant hybrid search service is registered Scoped (was Singleton), like the core service: it now takes the graph
  leg's retriever, which takes a scoped graph store on SQLite and PostgreSQL.
• PostgreSQL: indexing new vocabulary is several times faster. bm25_terms.term now uses the "C" collation (byte
  order). Terms are only compared for equality and sorted to give writers one insert order, so the database's
  linguistic collation bought nothing and cost every term insert and document-frequency update a locale comparison per
  index level — measured on Korean text under en_US.utf8, indexing 2,000 chunks with a fresh vocabulary went from
  107 s to 31–33 s. A table created by an earlier release is converted once on first open (the column type does not
  change; its unique index is rebuilt under a short exclusive lock). Search results are unchanged.

Fixed

• A rule-extracted or custom-pattern entity is found by a lower-case name. GraphEntity.NormalizedName is now always
  case-folded; the rule-based extractor and custom patterns used to store it as written («Apple Inc»), so the SQLite and
  PostgreSQL name lookups (which compare lower-case) and query-entity matching never found it. A graph stored by an earlier
  release keeps working: the next build that extracts such an entity joins the stored node (it also looks the name up as
  written) and writes it back folded; nodes read back from the store key folded.

Dependencies

• Re-pinned sibling package(s) FluxCurator 0.14.0 -> 0.15.0.
• Re-pinned sibling package(s) FluxGuard.Remote 0.21.0 -> 0.22.0.