@membook/core@0.1.1

Patch Changes

• #20 e41e553 Thanks @hiranofficial! - A model may fail to restore, but it may not destroy. LLM re-check invalidate
  verdicts now land as stale — withheld from the book, flagged for review,
  recoverable — instead of invalidating the memory outright. invalidated
  remains reachable only through deterministic evidence (the anchored file is
  gone) or a human decision in review. Measured on live data: a small model
  invalidated a still-true memory because a version number changed in its
  anchored file; restore requires string-matched evidence precisely to stop
  rubber-stamping, and this closes the same hole in the destructive direction.

• #20 e41e553 Thanks @hiranofficial! - New rememberMany on Membook: batch writes open the index once instead of
  per memory. Seeding and other bulk paths are dramatically faster on platforms
  where database open/close is expensive (Windows most of all). Additive API;
  shipped as a patch because this project reserves 0.2 for the workspace
  roadmap and the pre-1.0 minor lane signals breakage.

• #20 e41e553 Thanks @hiranofficial! - openIndex closes the database handle on every throw path. The
  metadata-mismatch throws leaked it, which was invisible on POSIX but made the
  index undeletable on Windows — so membook reindex, whose remedy for drift is
  deleting the file, failed with EBUSY. The error path no longer blocks its own
  cure.