fix(ai): suppress sealed slo watchdog meta noise

apps/api/src/jobs/ai_slo_watchdog_job.py

@@ -118,11 +118,18 @@ async def _check_once() -> None:
         report = await AiSloCalculator().calculate()
         if report.any_violated:
             violated = [m.name for m in report.metrics if m.violated]
-            w1_line, w1_cause = _format_slo_violation_for_alert(report, violated)
-            violations.append(w1_line)
-            if w1_cause:
-                probable_causes.append(w1_cause)
-            violation_codes.append(f"W1:slo_violated:{','.join(sorted(violated))}")
+            if _is_observation_only_slo_violation(report, violated):
+                logger.info(
+                    "watchdog_w1_slo_observation_only",
+                    violated=violated,
+                    reason="sealed_waiting_rolling_window",
+                )
+            else:
+                w1_line, w1_cause = _format_slo_violation_for_alert(report, violated)
+                violations.append(w1_line)
+                if w1_cause:
+                    probable_causes.append(w1_cause)
+                violation_codes.append(f"W1:slo_violated:{','.join(sorted(violated))}")
     except Exception as e:
         logger.warning("watchdog_w1_slo_check_failed", error=str(e))
 
@@ -354,6 +361,23 @@ def _format_slo_violation_for_alert(report, violated: list[str]) -> tuple[str, s
     return line, "\n".join(cause_parts)
 
 
+def _is_observation_only_slo_violation(report, violated: list[str]) -> bool:
+    """已封口且只等 rolling window 的 W-1，不再升成 Meta System 告警。"""
+    if set(violated) != {"auto_execute_success_rate"}:
+        return False
+
+    diagnostics = getattr(report, "diagnostics", {}) or {}
+    diag = diagnostics.get("auto_execute_success_rate") or {}
+    try:
+        open_groups = int(diag.get("open_failure_group_count") or 0)
+    except (TypeError, ValueError):
+        open_groups = 0
+    return (
+        diag.get("status") == "sealed_waiting_window"
+        and open_groups == 0
+    )
+
+
 def _short_taipei_time(value: str | None) -> str | None:
     if not value:
         return None

apps/api/tests/test_ai_slo_calculator.py

@@ -3,7 +3,10 @@ from __future__ import annotations
 from datetime import datetime, timedelta
 from zoneinfo import ZoneInfo
 
-from src.jobs.ai_slo_watchdog_job import _format_slo_violation_for_alert
+from src.jobs.ai_slo_watchdog_job import (
+    _format_slo_violation_for_alert,
+    _is_observation_only_slo_violation,
+)
 from src.services.ai_slo_calculator import (
     SLO_AUTO_SUCCESS_MIN,
     SloMetric,
@@ -176,3 +179,92 @@ def test_watchdog_formats_auto_execute_diagnostics_for_meta_alert():
     assert cause is not None
     assert "Top failure groups" in cause
     assert "不需要重啟服務或改寫歷史資料" in cause
+
+
+def test_watchdog_treats_sealed_auto_execute_slo_as_observation_only():
+    report = SloReport(
+        metrics=[
+            SloMetric(
+                name="auto_execute_success_rate",
+                value=0.83,
+                threshold=SLO_AUTO_SUCCESS_MIN,
+                direction="above",
+                sample_count=53,
+                violated=True,
+            )
+        ],
+        any_violated=True,
+        diagnostics={
+            "auto_execute_success_rate": {
+                "status": "sealed_waiting_window",
+                "open_failure_group_count": 0,
+            }
+        },
+    )
+
+    assert _is_observation_only_slo_violation(
+        report,
+        ["auto_execute_success_rate"],
+    )
+
+
+def test_watchdog_keeps_auto_execute_slo_alert_when_open_groups_remain():
+    report = SloReport(
+        metrics=[
+            SloMetric(
+                name="auto_execute_success_rate",
+                value=0.83,
+                threshold=SLO_AUTO_SUCCESS_MIN,
+                direction="above",
+                sample_count=53,
+                violated=True,
+            )
+        ],
+        any_violated=True,
+        diagnostics={
+            "auto_execute_success_rate": {
+                "status": "needs_investigation",
+                "open_failure_group_count": 1,
+            }
+        },
+    )
+
+    assert not _is_observation_only_slo_violation(
+        report,
+        ["auto_execute_success_rate"],
+    )
+
+
+def test_watchdog_keeps_slo_alert_when_other_metric_violates():
+    report = SloReport(
+        metrics=[
+            SloMetric(
+                name="auto_execute_success_rate",
+                value=0.83,
+                threshold=SLO_AUTO_SUCCESS_MIN,
+                direction="above",
+                sample_count=53,
+                violated=True,
+            ),
+            SloMetric(
+                name="human_override_rate",
+                value=0.25,
+                threshold=0.20,
+                direction="below",
+                sample_count=10,
+                violated=True,
+            ),
+        ],
+        any_violated=True,
+        diagnostics={
+            "auto_execute_success_rate": {
+                "status": "sealed_waiting_window",
+                "open_failure_group_count": 0,
+            }
+        },
+    )
+
+    assert not _is_observation_only_slo_violation(
+        report,
+        ["auto_execute_success_rate", "human_override_rate"],
+    )

docs/LOGBOOK.md

@@ -1,3 +1,38 @@
+## 2026-06-01｜W-1 已封口 SLO 改為觀察中，停止 Meta 誤擾
+
+**背景**：
+
+- 正式 `/api/v1/ai/slo?force_refresh=true` 仍顯示 `auto_execute_success_rate=44/53=83.0%`，低於 `85%` 門檻。
+- 但 diagnostics 已確認 9 筆失敗全屬兩個已封口群組：`DockerContainerUnhealthy` 已改走 `ssh_docker_restart/write` MCP grant，`StockWoooWorkDown` 已阻擋外部站台誤配 K3s node PlayBook。
+- `open_failure_group_count=0`，狀態是 `sealed_waiting_window`；這代表目前是 7 日 rolling window 歷史包袱，不是新故障。若仍送 `META SYSTEM | 飛輪核心異常`，會造成 operator 誤判與 Telegram 噪音。
+
+**本次調整**：
+
+- `apps/api/src/jobs/ai_slo_watchdog_job.py`：
+  - 新增 `_is_observation_only_slo_violation()`。
+  - 當且僅當違反項目只有 `auto_execute_success_rate`，且 diagnostics 為 `sealed_waiting_window`、`open_failure_group_count=0` 時，W-1 降成 observation-only，不再升成 Meta System 告警。
+  - 若存在未封口群組、診斷資料不足，或其他 SLO 一起違反，仍維持原本 Meta System 告警行為。
+- `apps/api/tests/test_ai_slo_calculator.py`：
+  - 新增已封口 observation-only 測試。
+  - 新增 open group / other metric 仍告警的防回歸測試。
+
+**驗證**：
+
+- Production SLO 查核：`any_violated=true`，`auto_execute_success_rate=83.0%`，`sealed_failure_group_count=2`，`open_failure_group_count=0`，`projected_green_at=2026-06-03T15:07:04.658109+00:00`。
+- Production health：Public `/api/v1/health` 回 `HTTP 200`，約 `0.19s`；`ollama_route_order=["ollama_gcp_a","ollama_gcp_b","ollama_local"]`，GCP-A / GCP-B up。
+- Local validation：
+  - `python3 -m py_compile apps/api/src/jobs/ai_slo_watchdog_job.py apps/api/src/services/ai_slo_calculator.py apps/api/tests/test_ai_slo_calculator.py`
+  - `DATABASE_URL=sqlite+aiosqlite:///:memory: PYTHONPATH=apps/api /Users/ogt/.pyenv/shims/pytest apps/api/tests/test_ai_slo_calculator.py apps/api/tests/test_trust_drift_watchdog.py -q` → `19 passed`
+  - `DATABASE_URL=sqlite+aiosqlite:///:memory: PYTHONPATH=apps/api /Users/ogt/.pyenv/shims/pytest apps/api/tests/test_ai_slo_calculator.py apps/api/tests/test_trust_drift_watchdog.py apps/api/tests/test_auto_repair_service.py apps/api/tests/test_alert_rule_engine_validation.py -q` → `84 passed`
+  - `python3 scripts/security/security-mirror-progress-guard.py --root .` → `SECURITY_MIRROR_PROGRESS_GUARD_OK`
+  - `git diff --check`
+
+**目前整體進度（本階段完成後）**：
+
+- W-1 Meta 告警降噪：約 `96%`；已封口 rolling-window 狀態不再升級為飛輪核心異常，剩餘需 production watchdog 下一輪實跑確認沒有再推同型 META。
+- W-1 自動修復 SLO 可解釋化：約 `92%`；可查、可視、可判斷人工/觀察邊界，但 SLO 數值仍需等歷史失敗滑出或新增真實成功樣本才回綠。
+- 完整 AI 自動化飛輪總進度：維持 `61%`；這次是降低錯誤告警與 operator 誤判，不等於新增 verified auto-repair 能力。
+
 ## 2026-06-01｜Alert Chain post-deploy API Health 重試化
 
 **背景**：