SuperLocalMemory 4.2.3 — Recall answers on large memories from every agent, hook and harness

A fix release. Nothing is moved or deleted, and no setting changes.

Fixes

• A quick look at memory came back empty on large stores. An agent's
  recall at depth=0 (the cheap one-line glance the tool suggests starting
  with) was treated like the prompt hook and given its three-second limit. On a
  store where full recall takes about 2.5 seconds (27,000 memories), it timed
  out and returned nothing, with every search channel marked "timeout". A
  glance is now a question like any other and waits for the full answer.
• session_init lost its search on large stores. The first call every
  agent makes loaded only pinned rules: its search ran under the same short
  limit and timed out. It now gets the full time. It is still never judged by
  the answer check.
• When the prompt hook ran out of time it injected nothing. The keyword
  backup only started after the full search was given up, with about 0.2
  seconds left, so it never finished. Now, when a hook's search is still running
  a second before its limit, the keyword backup starts beside it, and whichever
  is ready first answers. The hook still never holds up a prompt longer than
  before. This is the same for every harness (Claude Code, Codex, Cursor,
  Copilot, Antigravity, Hermes).
• Hermes prefetch gave up before recall answered. Hermes waits 3 seconds
  for the memory it adds to each turn, but asked for a full recall. It now sends
  its time budget (budget_s, a new optional recall parameter), and recall
  answers inside it.
• prestage_context always said recall_error. It recalled inside the
  MCP process, which may not; it now recalls through the memory service, as
  session_init does.
• Smaller tool fixes found by an end-to-end run of every MCP tool:
  skill_lineage reports an empty history instead of an error on a store where
  no skill has evolved; correct_pattern refuses a pattern that does not exist
  instead of reporting success; review_correction no longer tells an agent to
  retry a review the memory service refused (an unknown or settled case).
• Docs: the slm-loop skill and the MCP tool reference now say what
  slm_loop_run needs to finish: an operator-registered check
  (gate_contract). A matching memory alone never ends a loop.