Fix durable memory update acknowledgements for 0.14.10

[jsflax]

Memory updates could report success after an unchecked transaction failed, and updates selected from synced or group stores did not consistently use the connection that owns the selected row. Use checked transactions on that connection and acknowledge success only after commit. Undelete now checks edge restoration too; a later graph failure remains an uncertain partial outcome.

The restricted session learner recognizes the exact typed pre-transaction busy receipt for update, while preserving reconciliation for uncertain writes. Version 0.14.10 also includes the already merged learner schema and stop-after-error fixes.

Validation

• 113 focused gateway/runner tests, 335 assembled-plugin tests, and 50 release protocol, workflow and packaging tests passed.
• Independent source and fixture review accepted.
• Native tests passed: 32 synthetic storage cases plus 20 existing update, restore and privacy regressions (24 Swift Testing methods). Existing regressions use the real bundled model in an opt-in CPU fixture; production model settings are unchanged.
• The same six busy-BEGIN cases fail on unchanged main: it returns success while independent SQLite reads show unchanged data. The local-query recovery check also exposes the old connection bug. The fixed code passes these cases.

The synthetic cases cover local, synced and group stores by ID and query, busy BEGIN, setter failures and rollback, spoofed error text, independent SQLite readback, and partial undelete outcomes. Deferred COMMIT behavior is covered by the pinned Lattice SDK tests and the reviewed handler guard rather than a new handler-level COMMIT fault fixture.