skillmem 0.13.0 — approve rules in chat

Rules an agent saves for you now get approved, instead of sitting unapproved
for weeks — and recall finds your curated notes under a pile of session recaps.

• Approve in chat. A new Stop hook, pending-rules, asks the person (as a
  systemMessage, in Russian and English) about every user/feedback rule
  the last turn saved, at most three, in full, read from the database. If their
  very next message is just одобряю / approve, the UserPromptSubmit hook
  approves exactly that text (trusted_by='owner-prompt') and keeps the word
  from the agent. Any other reply withdraws the offer. Longer rules are left for
  review. A second owner signal, documented with its limits in INVARIANTS §0.
• pending-rules imports the auto-memory .md files the turn changed, so a
  rule reaches the database (and recall) before the session ends instead of on
  a manual skillmem migrate.
• skillmem review walks every unapproved user/feedback memory
  (--types for more), shows it in full like trust, and asks y / n / q.
  Owner-only. The session briefing now asks the agent to tell the user about it.
• Recall ranking. Session recaps (note/session, ~90% of a long-lived
  database) get half their RRF score: they retell every topic and pushed the one
  curated note that answered to #28. Own 30-question set hit@3 18 → 26/30;
  LongMemEval hit@5 unchanged (0.889).
• tool-recall runs lexical-only: loading the embedder cost ~1.2 s on every
  Bash/Edit/Write call (1.45 s → 0.2 s).
• init --claude-code adds the Stop hook and deny rules for review, running
  hooks from Bash, and owner_prompt.