v0.4.0 — recall returns memories

• Recall returns memories. res["hits"] becomes res["memories"], and
  each entry is one whole memory with its content rather than a scattering
  of its sections. matched names the arms that found it, sections the
  headings that matched, via what pulled in a neighbour.
• Scores mean something. score is calibrated in [0, 1] and comparable
  across queries, replacing a fused rank that only ordered one response.
  scores["coverage"] says how much of the query a memory accounted for.
• recall takes filters — tiers, paths, tags, tags_all, since,
  until, min_score, context, detail, include_expired — applied
  inside every retrieval arm server-side rather than after the fact.
• answer(query) returns one text answer: a fast model over a retrieval,
  or with agentic=True a stronger model that searches the memory itself with
  tools and returns its trace. Raises GitloomError("no_answer") rather
  than returning an empty string. Both meter as chats.
• memory.vocab — learn, list, lookup, forget. A learned alias
  makes a query for any surface form find memories written with another;
  matched definitions come back as defined.
• memory.skills — store and find, stored as memories under the
  skills/ tier.
• candidates, filtered_out and timings report what retrieval did.