SuperLocalMemory v4.2.17

The last fixes of this cycle, batched into one release. No new settings and no
data migration: the store stays at schema 68, so 4.2.16 and 4.2.17 can open
each other's stores.

Added

• slm hide <fact_id>... keeps memories out of recall answers and session
  starts without deleting or editing them; slm hide --undo shows them again.
  Both are stored as your own decision, which the automatic bookkeeping rule
  and every repair leave alone. Until now this needed hand-written SQL.
• Jev can order the memories a session starts with. If you chose Jev as the
  answer check and turned on its reordering (with its consent), the SessionStart
  hook and session_init ask Jev once for the order of their memories, against
  your first message or the project. Only the order is used: the answer-check
  verdict is discarded, so a session start never says "that isn't in
  SuperLocalMemory", and nothing is hidden. It waits at most 1.5 seconds and
  keeps the usual order on any failure. With reordering off, a session start
  makes no Jev request and nothing changes.

Fixed

• A replaced memory no longer stays pinned at session start. When
  remember(replaces=...) or a correction retired a fact, recall and pinned
  facts already left it out, but the session-start context (slm session-context, which the Claude Code and Codex SessionStart hooks print)
  and the Core Memory blocks read the store with their own queries and kept
  showing the old fact under "Core Memory (pinned, high-value)". Both now skip
  retired facts, and a block compiled before its fact was retired is not shown.
  Nothing is deleted: the retired fact stays in history and audit.
• A question that names a version is no longer answered by another version.
  "SLM 4.2.15 release" ranked "SLM 4.2.0 released" and six more other-version
  notes first, because embeddings and the entity graph see 4.2.0 and 4.2.15 as
  the same thing. A result that names only other versions now goes below the
  rest, both before recall cuts to your limit and in the final order. An exact
  match still comes first, and nothing is dropped.
• Recall records which memories it returned again. 3.8.6 made recall
  write-free so it never waits behind remember, and the record in
  fact_access_log went with it; on one store the last recall row was dated
  26 July 2026. Retention's access clock and the "active decisions" Core Memory
  block read that table, so every recall since then counted as unread: a memory
  you kept recalling still aged as unused, and no new decision reached the
  block. The record is now queued on SLM's background writer, so recall itself
  stays write-free. SLM's own warm-up recalls are not counted.
• PromptInjector.inject_into_context ignored its documented total_budget
  override; it is now split between soft prompts and memories in the configured
  ratio.
• Cloud backup sync reports the size it uploaded (size_mb).

Changed

• The code base is lint-clean (ruff check src tests reports nothing), and CI
  now runs a pinned ruff on every change. Line length is 120.