fix(alerts): краш при restart-политике, гистерезис закрытия, редкие правки

Найдено смоуком на живых контейнерах (examples/, #19):
- контейнер с restart: on-failure поднимается за секунду — между раундами
  его не видно в exited; краш теперь распознаётся и по росту RestartCount
  (сравнение со снапшотами до текущего раунда), повторы — счётчик
- алерт закрывается, только когда условия нет resolve_after (60 с): иначе
  флапающий сервис получал бы «✅ восстановлено» после каждого раунда
- правки сообщений со счётчиком — не чаще раза в минуту на алерт

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app/alerts/service.py

@@ -15,6 +15,7 @@ from app.db.session import SessionMaker
 log = logging.getLogger(__name__)
 
 HISTORY = timedelta(days=7)  # AL-6
+EDIT_EVERY = timedelta(minutes=1)
 LOG_CHARS = 3000
 
 LogSource = Callable[[str, int], Awaitable[str]]  # (container id, lines) -> tail
@@ -68,6 +69,11 @@ class AlertService:
                 known = by_key.get(key)
                 if known is not None and known.resolved_at is None:
                     known.last_at = now
+                    if c.bump:  # one more incident while still open (a crash that Docker already restarted)
+                        known.count += 1
+                        known.payload = {**known.payload, **c.payload}
+                        if self._may_edit(known, now):
+                            repeated.append(known)
                     continue
                 if (
                     known is not None
@@ -76,7 +82,8 @@ class AlertService:
                 ):
                     known.resolved_at, known.last_at, known.count = None, now, known.count + 1  # back within the window
                     known.payload = {**known.payload, **c.payload}
-                    repeated.append(known)
+                    if self._may_edit(known, now):
+                        repeated.append(known)
                     continue
                 fresh = Alert(
                     kind=c.kind, service=c.service, payload=await self._payload(c), opened_at=now, last_at=now
@@ -84,12 +91,12 @@ class AlertService:
                 session.add(fresh)
                 opened.append(fresh)
             for key, a in by_key.items():
-                if a.resolved_at is None and key not in wanted:
+                # hysteresis: gone for a while, not just for one round — otherwise a flapping service
+                # would get «✅ восстановлено» after every blip
+                if a.resolved_at is None and key not in wanted and now - as_utc(a.last_at) >= self._settle:
                     a.resolved_at = now
                     resolved.append(a)
             await session.flush()
-            for a in opened:
-                await session.refresh(a, ["deliveries"])
         for a in opened:
             await self._notifier.opened(a, admins_only=False)
         for a in repeated:
@@ -112,18 +119,17 @@ class AlertService:
                 recent.count += 1
                 recent.last_at = now
                 recent.payload = {**recent.payload, **c.payload}
-                alert, new = recent, False
+                alert, new, edit = recent, False, self._may_edit(recent, now)
             else:
                 alert = Alert(
                     kind=c.kind, service=c.service, payload=dict(c.payload), opened_at=now, last_at=now, resolved_at=now
                 )
                 session.add(alert)
                 await session.flush()
-                await session.refresh(alert, ["deliveries"])
-                new = True
+                new, edit = True, False
         if new:
             await self._notifier.opened(alert, admins_only=admins_only)
-        else:
+        elif edit:
             await self._notifier.repeated(alert)
         return alert
 
@@ -140,6 +146,18 @@ class AlertService:
     def _window(self) -> timedelta:
         return timedelta(seconds=self._cfg.repeat_window)
 
+    @property
+    def _settle(self) -> timedelta:
+        return timedelta(seconds=self._cfg.resolve_after)
+
+    def _may_edit(self, alert: Alert, now: datetime) -> bool:
+        """Counter edits at most once a minute per alert; the count itself is always exact in the DB."""
+        last = alert.payload.get("edited_at")
+        if isinstance(last, str) and now - as_utc(datetime.fromisoformat(last)) < EDIT_EVERY:
+            return False
+        alert.payload = {**alert.payload, "edited_at": now.isoformat()}
+        return True
+
     async def page(self, page: int, per_page: int = 10, now: datetime | None = None) -> tuple[list[Alert], int]:
         """Active first, then the last 7 days, newest first (AL-6). Admin-only kinds are left out."""
         now = now or utcnow()