chore(tui): polish demo + regenerate screenshots

Rework the VHS demo so the README screenshots actually entice a download.

Demo data / tooling:
- seed.yaml: real, reachable service URLs (detail now shows nextcloud.com,
  not example.com); Auth Portal -> non-resolving home.arpa host so it reads
  as a believable, reliably-DOWN monitor
- backfill: transient outages for Nextcloud/Jellyfin/Immich aligned with their
  state changes (uptime % now matches); log timestamps derived from now so the
  Logs view reads chronologically; real SSL warning; three probe nodes across
  regions; seeded alert send health
- demo.tape: shorter warm-up, added Nodes + theme captures, ordered so every
  shot stays inside the 60s node-freshness window (consistent probe count)
- vhs/crop: new tool to trim the empty terminal border around each screenshot
- setup.sh: build backfill up front for deterministic timing; UPTOP_DEMO=1

Supporting code:
- persist alert send health (new alert_health table, load on startup,
  best-effort save on send) so health/last-sent survive restarts
- latency Min/Avg/Max ignore failed checks (no more "Min 0ms")
- correct "probe"/"probes" pluralization
- stable status dot instead of an animated spinner under UPTOP_DEMO

internal/models/models.go

@@ -79,6 +79,17 @@ type ProbeNode struct {
 	Version  string
 }
 
+// AlertHealthRecord is the persisted send health of an alert channel. It lets the
+// "last sent" / health indicators survive restarts instead of resetting to "never".
+type AlertHealthRecord struct {
+	AlertID    int
+	LastSendAt time.Time
+	LastSendOK bool
+	LastError  string
+	SendCount  int
+	FailCount  int
+}
+
 type MaintenanceWindow struct {
 	ID          int
 	MonitorID   int