…#1494) (#1503)

* feat(reminders): operator visibility into reminder failures and skips (#1494)

Reminder failures were invisible: a stuck/failing reminder only emitted a daemon
log warning, so the gotowebinar incident skipped 49 fires in a day unnoticed.
This surfaces failures where operators look.

- Failure → destination channel: when a Channel-delivery reminder execution
  fails for any reason (including the #1500 stall backstop, or auto-disable), a
  plain-language notice is posted to the reminder's own destination channel via a
  new IReminderChannelNotifier seam — interface in Netclaw.Actors (transport-
  agnostic, hands over a resolved ChannelDeliveryTargetInfo), implemented in
  Netclaw.Daemon over IChannelRegistry. Bounded by the auto-disable threshold, so
  no channel spam. Skips are NOT posted (too noisy) — only counted.

- Skipped-duplicate visibility: the two skip sites (scheduled + deferred queue)
  now increment an in-memory per-reminder counter (since daemon start) instead of
  only logging.

- `netclaw reminder status <id>`: new CLI subcommand + GET /api/reminders/{id}/
  status REST endpoint, backed by a new GetReminderStatusQuery the manager
  answers from live state — enabled, in-flight, next fire, consecutive failures,
  skipped fires, and recent run history.

The actor stays transport-agnostic: ReminderManagerActor resolves the channel
target (same assembly) and the daemon notifier owns the channel post; a
NullReminderChannelNotifier Null Object serves tests / no-channel daemons (a
required dependency, not a nullable one).

Tests: ReminderChannelFailureNotifier posts the right ChannelSendRequest
(destination + DM kinds); status query returns per-reminder health and not-found.
Full Netclaw.Actors.Tests (2482) and Netclaw.Daemon.Tests (754) green. Updates the
netclaw-operations skill (scheduling reference, 2.19.0 → 2.20.0).

Note: the behavioral eval suite (skill-content edit) needs a live model endpoint
not available here — run in CI / before merge.

* fix(reminders): address code-review findings on #1494

Verified findings from the workflow code review:

- LlmSessionTestBase: register IReminderChannelNotifier (the new required
  reminder-manager dependency) so DI-resolved reminder managers in session
  integration tests initialize instead of dying with ActorInitializationException.
- HandleGetStatusAsync: a transient read failure (schedule fetch / history IO)
  no longer impersonates "not found" — it faults the Ask (Status.Failure → the
  endpoint maps to 5xx) so the operator sees a real error instead of being told a
  wedged reminder was deleted. That was a silent fallback on the very feature
  meant to expose them.
- Status query: read and display the same RecentHistoryCount (5) records, newest
  first — the CLI previously showed the OLDEST 5 of a last-10 window, hiding the
  most recent (failing) runs.
- Channel notice: post the per-failure notice only below the auto-disable
  threshold; on the threshold-hitting failure the disabled notice already carries
  the last error, so both was double noise on the most important event.
- NextFire: pass null through (FormatTimestamp renders null as "unknown"); the CLI
  shows "not scheduled" for disabled/unscheduled reminders.
- Status handler runs its two independent backend reads concurrently (Task.WhenAll).

Full Netclaw.Actors.Tests (2482) green; reminder + Daemon notifier tests green.

Not changed (verified lower-severity / pre-existing patterns): whole-list
scheduler fetch (matches existing GetReminder), notifier↔ChannelSendTools and
CLI-subcommand scaffolding duplication.