守住 AutoHeal inline 告警去重

docs/memory/claude_inventory_validation_20260513.md

@@ -36,6 +36,7 @@
 - Cron 盲區清單多數已補 `_notify_scheduler_failure()`；ROI 月報已避開 09:00 改 09:05，AI smoke 已是 09:10。
 - V2 指出的 9 個 cron 盲區已補回歸守門：8 個 `run_scheduler.py` wrapper 必須呼叫 `_notify_scheduler_failure()`，`scheduler.py::run_promo_event_task` 必須呼叫 `notify_failure()`。
 - 09:00 排程衝突已有回歸守門：`daily_report=09:00`、`roi_monthly_report=09:05`、`ai_smoke_daily_summary=09:10` 必須保持錯開。
+- AutoHeal inline host transition 告警已有 DB transition dedup 守門：`run_host_health_probe()` 只在健康狀態翻轉且 1h 內沒有同方向 transition 時推送。
 - CD migration 全範圍冪等已有回歸守門：`.gitea/workflows/cd.yaml` 必須維持 024-099 pattern、`sort | uniq` 與 `for m in $V5_MIGRATIONS` apply loop。
 - CD Observability production smoke 已補 timeout 守門：`quick_review.sh --observability-smoke` 必須帶 `--timeout 12`。
 - 0-byte `database/momo*.db` 迷惑檔已不存在；真實 SQLite 僅在 `data/momo_database.db`。

tests/test_run_scheduler_embed_consistency.py

@@ -121,6 +121,18 @@ def test_scheduler_observability_wrappers_notify_on_exception(monkeypatch):
         assert "_notify_scheduler_failure(" in source
 
 
+def test_host_health_transition_alert_keeps_db_dedup_window(monkeypatch):
+    run_scheduler = _load_run_scheduler(monkeypatch)
+    source = inspect.getsource(run_scheduler.run_host_health_probe)
+
+    assert "prev_healthy != rec['healthy']" in source
+    assert "recent_transition" in source
+    assert "INTERVAL '1 hour'" in source
+    assert "INTERVAL '90 minutes'" in source
+    assert "if recent_transition is None:" in source
+    assert "_push_host_transition_alert(tr)" in source
+
+
 def test_v2_cron_blind_spot_list_has_failure_notifications(monkeypatch):
     run_scheduler = _load_run_scheduler(monkeypatch)