SuperLocalMemory 4.2.5 — every sentence you save can be found, and saves no longer hold up the store

No migration. Nothing is moved or deleted.

Every sentence is findable

• A long memory kept only a few of its sentences as facts. A 767-character
  launch note was saved as the whole note plus four facts, its first two
  bullets; asking "Product Hunt license fix" found nothing, because that line
  lived only inside the whole-note fact. Three hard-coded limits of five facts
  per memory overrode the configured ten. The configured limit is back, and
  after extraction every sentence or bullet that no extracted fact already
  holds is kept as a fact of its own, word for word (up to 40 per memory;
  greetings, fragments and markup lines are skipped). Word-for-word text cannot
  misstate what was saved, so it needs no model call.
• Memories saved before this release can gain the same, and this is off
  by default. On a 27,700-fact store it adds about 14,800 facts (half again
  the store), and its effect on recall time has not yet been measured on a
  quiet machine, so it is not switched on for everyone. To turn it on, add
  "encoding": {"coverage_backfill_per_memory": 10} to config.json in the
  data folder and run slm restart. It then adds up to ten facts per memory,
  newest first, in small batches that give way to every recall. It only
  touches memories whose full text is already searchable, so nothing becomes
  findable that was not before; anything erased, archived, corrected or held
  for review is left alone. Each added fact is listed in
  backfill/coverage-backfill.jsonl in the data folder.

Saves no longer hold up the store

On a copy of a 27,700-fact store, two steps of every save scanned the whole
store while holding the write lock, so under load other writes waited, mesh
heartbeats failed with "database is locked" and accepted memories retried:

Step, same store	4.2.4	4.2.5
Duplicate check on each save	1.3-1.7 s (5-30 s under load)	under 1 ms
Keeping a scene's member list (1,607 members)	4.26 s	16 ms
Lock holds over 1.5 s while 52 memories were saved (load 25-53)	41, longest 5.9 s	0

• A write whose caller has already given up now stops at its deadline and lets
  go of the lock, instead of finishing work that is rolled back anyway.
• The background re-read of upcoming plans reads and classifies outside the
  write lock; it locks only for its own update.

Fixes

• slm doctor warned about memories saved minutes ago as "missing from the
  search index". They are indexed in the background, a minute or two behind
  on a busy machine. Memories from the last ten minutes are now reported as
  being indexed; older ones still warn.
• Repair checks were skipped on a machine up less than half an hour. "Never
  checked" was stored as 0 and compared with a clock that counts from boot.