The first pass read 'alarmaId'/'tipo' from the channel payload while the
native side sends 'alarmId'/'type'/'atMillis' (AlarmScheduler.kt:1389).
Every entry would have been dropped silently in production.
The tests passed because the fake was seeded with the same guessed keys,
so they confirmed the mistake instead of catching it. Decoding now goes
through FalloProgramacionNativo.fromMap -- the single place native key
names appear -- and the fixtures build through that same constructor.
Completes the bridge the native side already exposed. AlarmScheduler and
PluriWaveAlarmService record a pre-notice that could not be armed, a
refused foreground-service start, and a per-alarm reschedule that failed
after a reboot -- but nothing read them, so all three still ended at
logcat.
EstadoAlarmas now drains them at startup and turns each into a per-alarm
exception, which the card already knows how to mark. An alarm that never
reached the OS stops looking identical to one that did.
The read is deliberately tolerant: a failure to read is logged and
swallowed, never surfaced as an alarm error, so a diagnostics gap cannot
masquerade as a scheduling problem.
FreeTLab