v0.35.5

Added

• An optional RuVector backend for memory and retrieval, installed with the ruvector extra
  (lionagi[ruvector], which depends on the ruvector SDK and numpy) and imported explicitly
  from lionagi.integrations.ruvector; the default import lionagi does not load it.
  RuVectorMemoryStore implements the MemoryStore protocol for Branch(memory=...) over a native
  HNSW index. A text query ranks by cosine similarity, approximately through the index when no tags
  or filters are given and exactly over every eligible record when they are; tags match any
  supplied tag and metadata filters must all match. RuVectorStore.as_tool gives a branch a
  read-only ruvector_search tool. Embeddings come from the application, and the index lives in
  process memory with no persistence. Contributed by @ruvnet (#3451).
• A roadmap (docs/roadmap.md, linked from the README) and its design record, ADR-0125, set out
  the planned migration of lionagi's architecture towards lion1, the runtime kept in this
  repository's lion1/ directory, in seven steps that keep public imports, call shapes and return
  types working. These are documents only; no runtime code changes for the migration in this
  release.

Fixed

• Cancelling an engine run no longer leaves a CLI process running when it was spawned during the
  reap and given the memory address of a process already reaped: the rescan for late spawns took
  it for one it had already handled and returned with it still tracked.
• fail_at and move_on_at in lionagi.ln.concurrency put the deadline they are given on the
  cancel scope. They converted it to a duration and read the clock again, so the scope ended later
  than asked by however long passed between the two reads.