Backup restore validates the entire legacy array or version-one JSON envelope before writing, rejects malformed records, duplicate IDs and unsupported future versions, and previews additions versus existing IDs. Existing pieces always win, regardless of backup dates. Confirmation re-reads the current database and adds missing pieces in one IndexedDB transaction; quota/errors abort the entire import. Missing, repeated and self connections on added pieces are counted in the preview and omitted. Source text, writing, accepted organization, grounded evidence, rejection history and compatible maintenance metadata are retained. Pending imported connection/organization reviews start paused; restored pieces do not trigger automatic legacy topic assignment. Existing review controls and explicit topic sorting remain available. Backup export reads a fresh database snapshot and includes export time/app version.
Legacy topic assignment also commits against the latest collection, protecting edits, source replacements, manual categories and deletions while a provider response is in flight.
Monthly request counters record provider starts, successful cache hits/live shared results, and capacity/daily/pending waits in the existing per-user aggregate ledger. They store no source text, piece IDs or request fingerprints. The existing spending details show reuse and waits; counters begin at v45, count requests rather than unique pieces, and do not assert dollars saved or increase cost estimates. Existing v44 allowances, cache retention and observed-token accounting remain intact.
A developer-only quality comparison harness uses original synthetic sources through production organization/breakdown, connection and grouping paths. Replay/scoring and regression detection run without credentials in tests. Live comparisons require explicit –live, a model and a key, make at most ten serial provider calls, and are never run by CI. The report includes rubric checks, literal evidence, context sizes, latency and observed token estimates. No live-model superiority or quality parity is claimed; semantic review is still required. Model defaults, evidence budgets, grouping rules, map layout and everyday navigation are unchanged.
After the Worker deployment, refresh the app to v45. No existing library migration or full reprocessing is required. JSON backups remain important because the library lives in device/browser storage. Installed iPhone Share Sheet and physical VoiceOver behavior remain manual checks beyond automated Chromium/WebKit coverage.