Tier 1 invariants held on live run 2026-06-24: the observe-alert-acknowledge loop for
critical hyperkalemia (K+ 6.8 mEq/L) completed entirely on the ward gateway while the
central API was stopped. Ten sync items buffered locally during the outage; central
received no new observations until reconnect. After central restart, `SyncUploaderService`
drained the buffer within the 180s verification window — gateway reported buffer depth 0,
and central showed the critical alert with `acknowledgedBy: RN-Wu` and
`syncedFromGateway: true`. Central paging was suppressed for the gateway-synced alert
(no duplicate critical pages). The run used instant replay (`--speed 0`) rather than a
full 30-minute wall-clock outage, but the architectural property — ward-local safety
with deferred idempotent reconciliation — was demonstrated end-to-end via
`run-phase24-verification.sh`.