@oneuptime/common 12.0.32 → 12.0.34

package/Tests/Server/Services/NetworkAlertPolicyEngineService.test.ts

@@ -0,0 +1,1849 @@
+/*
+ * Contract under test — the Network Alert Policy engine, the ONE place in the
+ * product that creates or deletes a policy-owned Monitor.
+ *
+ * Monitors are billed and they own incident history, so nearly every case
+ * here is a case about NOT doing something. The engine is reached from a
+ * dozen directions — a device created, re-sited, re-labelled, archived,
+ * switched to monitor-backed, handed a probe; a policy saved, enabled,
+ * re-scoped, re-pointed, deleted; a five-minute sweep — and each of them
+ * ends in `reconcileDevice`, which computes DESIRED against ACTUAL for one
+ * device under that device's own Redis lock. What this file pins is the
+ * behaviour that would otherwise fail silently and expensively:
+ *
+ *   - TWO POLICIES ON ONE DEVICE each get their own monitor, and neither
+ *     pass touches the other's row. Ownership is the PAIR (policy,
+ *     template), not the device.
+ *   - A SCOPE THAT SHRINKS deletes only monitors stamped with THAT policy.
+ *     A hand-made Network Device monitor on the same device, and an
+ *     auto-import rule's, carry no networkAlertPolicyId and are therefore
+ *     not even in ACTUAL — the query, not a filter afterwards, is what makes
+ *     that true.
+ *   - AN UNOWNED (device, template) MONITOR IS ADOPTED, never duplicated.
+ *     Monitor's partial unique index on (autoProvisionedNetworkDeviceId,
+ *     monitorTemplateId) would refuse the duplicate on this pass and on
+ *     every pass after it, so the policy would report a failure forever
+ *     beside a perfectly good monitor.
+ *   - A DEVICE THAT STOPS BEING PROVISIONABLE — archived, monitor-backed, or
+ *     with no probe — loses every policy monitor. A monitor-backed device's
+ *     health is a bound monitor's; a Network Device monitor on it is billed
+ *     for a poll that will never happen.
+ *   - A DISABLED POLICY PAUSES its monitors instead of deleting them, and
+ *     enabling reverses it. Deleting would lose the incident history and
+ *     re-bill the fleet on the way back.
+ *   - THE PLAN IS ASKED ONCE PER RUN and the run stops at the first refusal,
+ *     so a project that cannot hold another monitor gets one sentence in
+ *     lastSyncError rather than five hundred identical failures.
+ *   - THE LOCK is per device, taken before anything is read, released even
+ *     when the body throws, and a lock somebody else holds means this pass
+ *     does nothing at all.
+ *
+ * The collaborating singleton services — and the Semaphore module, which
+ * would otherwise reach Redis — are stubbed at the MODULE level before the
+ * engine is imported: their real files reach Postgres through
+ * DatabaseService, and nothing here should touch any of it.
+ *
+ * MonitorService.findBy and NetworkDeviceService.findBy are backed by small
+ * in-memory tables rather than by call-order mocks, because the engine asks
+ * them several different questions per pass and a `mockResolvedValueOnce`
+ * chain would pass for the wrong reason the moment the query order changed.
+ */
+
+jest.mock("../../../Server/EnvironmentConfig", () => {
+  return {
+    ...(jest.requireActual("../../../Server/EnvironmentConfig") as Record<
+      string,
+      unknown
+    >),
+    IsBillingEnabled: true,
+    AllowedActiveMonitorCountInFreePlan: 10,
+  };
+});
+
+jest.mock("../../../Server/Services/MonitorService", () => {
+  return {
+    __esModule: true,
+    default: {
+      create: jest.fn(),
+      findBy: jest.fn(),
+      findOneBy: jest.fn(),
+      countBy: jest.fn(),
+      deleteBy: jest.fn(),
+      updateBy: jest.fn(),
+      updateColumnsByIdWithoutHooks: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Services/MonitorTemplateService", () => {
+  return {
+    __esModule: true,
+    default: {
+      findOneBy: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Services/NetworkAlertPolicyService", () => {
+  return {
+    __esModule: true,
+    default: {
+      findBy: jest.fn(),
+      findOneBy: jest.fn(),
+      updateColumnsByIdWithoutHooks: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Services/NetworkDeviceService", () => {
+  return {
+    __esModule: true,
+    default: {
+      findBy: jest.fn(),
+      findOneBy: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Services/ProjectService", () => {
+  return {
+    __esModule: true,
+    default: {
+      getCurrentPlan: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Infrastructure/Semaphore", () => {
+  return {
+    __esModule: true,
+    default: {
+      lock: jest.fn(),
+      release: jest.fn(),
+    },
+  };
+});
+
+jest.mock("../../../Server/Utils/Logger", () => {
+  return {
+    __esModule: true,
+    default: {
+      debug: jest.fn(),
+      info: jest.fn(),
+      warn: jest.fn(),
+      error: jest.fn(),
+      trace: jest.fn(),
+    },
+  };
+});
+
+import NetworkAlertPolicyEngineService, {
+  MAX_MONITORS_PER_POLICY_SYNC,
+  POLICY_DEVICE_LOCK_NAMESPACE,
+  PolicyRunContext,
+  policyDeviceLockKey,
+} from "../../../Server/Services/NetworkAlertPolicyEngineService";
+import MonitorService from "../../../Server/Services/MonitorService";
+import MonitorTemplateService from "../../../Server/Services/MonitorTemplateService";
+import NetworkAlertPolicyService from "../../../Server/Services/NetworkAlertPolicyService";
+import NetworkDeviceService from "../../../Server/Services/NetworkDeviceService";
+import ProjectService from "../../../Server/Services/ProjectService";
+import Semaphore from "../../../Server/Infrastructure/Semaphore";
+import Label from "../../../Models/DatabaseModels/Label";
+import Monitor from "../../../Models/DatabaseModels/Monitor";
+import MonitorTemplate from "../../../Models/DatabaseModels/MonitorTemplate";
+import NetworkAlertPolicy from "../../../Models/DatabaseModels/NetworkAlertPolicy";
+import NetworkDevice from "../../../Models/DatabaseModels/NetworkDevice";
+import FindBy from "../../../Server/Types/Database/FindBy";
+import FindOneBy from "../../../Server/Types/Database/FindOneBy";
+import { PlanType } from "../../../Types/Billing/SubscriptionPlan";
+import BadDataException from "../../../Types/Exception/BadDataException";
+import MonitorStep from "../../../Types/Monitor/MonitorStep";
+import MonitorSteps from "../../../Types/Monitor/MonitorSteps";
+import MonitorType from "../../../Types/Monitor/MonitorType";
+import NetworkDeviceMonitoringMethod from "../../../Types/NetworkDevice/NetworkDeviceMonitoringMethod";
+import NetworkAlertPolicyScope from "../../../Types/NetworkDevice/NetworkAlertPolicyScope";
+import ObjectID from "../../../Types/ObjectID";
+import PositiveNumber from "../../../Types/PositiveNumber";
+import { beforeEach, describe, expect, it, jest } from "@jest/globals";
+
+const PROJECT_ID: ObjectID = new ObjectID(
+  "11111111-1111-4111-8111-111111111111",
+);
+const DEVICE_ID: ObjectID = new ObjectID(
+  "22222222-2222-4222-8222-222222222222",
+);
+const OTHER_DEVICE_ID: ObjectID = new ObjectID(
+  "2b2b2b2b-2b2b-4b2b-8b2b-2b2b2b2b2b2b",
+);
+const POLICY_ID: ObjectID = new ObjectID(
+  "33333333-3333-4333-8333-333333333333",
+);
+const OTHER_POLICY_ID: ObjectID = new ObjectID(
+  "3b3b3b3b-3b3b-4b3b-8b3b-3b3b3b3b3b3b",
+);
+const TEMPLATE_ID: ObjectID = new ObjectID(
+  "44444444-4444-4444-8444-444444444444",
+);
+const OTHER_TEMPLATE_ID: ObjectID = new ObjectID(
+  "4b4b4b4b-4b4b-4b4b-8b4b-4b4b4b4b4b4b",
+);
+const PROBE_ID: ObjectID = new ObjectID("55555555-5555-4555-8555-555555555555");
+const SITE_ID: ObjectID = new ObjectID("66666666-6666-4666-8666-666666666666");
+const OTHER_SITE_ID: ObjectID = new ObjectID(
+  "6b6b6b6b-6b6b-4b6b-8b6b-6b6b6b6b6b6b",
+);
+const LABEL_ID: ObjectID = new ObjectID("77777777-7777-4777-8777-777777777777");
+
+// What the mocked Semaphore.lock hands back and release must get back.
+const FAKE_MUTEX: { id: string } = { id: "fake-device-mutex" };
+
+const monitorCreateMock: jest.Mock =
+  MonitorService.create as unknown as jest.Mock;
+const monitorFindByMock: jest.Mock =
+  MonitorService.findBy as unknown as jest.Mock;
+const monitorFindOneByMock: jest.Mock =
+  MonitorService.findOneBy as unknown as jest.Mock;
+const monitorCountByMock: jest.Mock =
+  MonitorService.countBy as unknown as jest.Mock;
+const monitorDeleteByMock: jest.Mock =
+  MonitorService.deleteBy as unknown as jest.Mock;
+const monitorUpdateByMock: jest.Mock =
+  MonitorService.updateBy as unknown as jest.Mock;
+const monitorStampMock: jest.Mock =
+  MonitorService.updateColumnsByIdWithoutHooks as unknown as jest.Mock;
+const templateFindOneByMock: jest.Mock =
+  MonitorTemplateService.findOneBy as unknown as jest.Mock;
+const policyFindByMock: jest.Mock =
+  NetworkAlertPolicyService.findBy as unknown as jest.Mock;
+const policyFindOneByMock: jest.Mock =
+  NetworkAlertPolicyService.findOneBy as unknown as jest.Mock;
+const policyStampMock: jest.Mock =
+  NetworkAlertPolicyService.updateColumnsByIdWithoutHooks as unknown as jest.Mock;
+const deviceFindByMock: jest.Mock =
+  NetworkDeviceService.findBy as unknown as jest.Mock;
+const deviceFindOneByMock: jest.Mock =
+  NetworkDeviceService.findOneBy as unknown as jest.Mock;
+const currentPlanMock: jest.Mock =
+  ProjectService.getCurrentPlan as unknown as jest.Mock;
+const semaphoreLockMock: jest.Mock = Semaphore.lock as unknown as jest.Mock;
+const semaphoreReleaseMock: jest.Mock =
+  Semaphore.release as unknown as jest.Mock;
+
+/*
+ * A query value the engine could have written. Understanding the raw
+ * operators (rather than only exact values) is what lets the fake tables
+ * answer the SAME questions the database would — including the two that
+ * carry the whole safety story here, `networkAlertPolicyId IS NOT NULL` and
+ * `IS NULL`.
+ */
+function matchesQueryValue(actual: unknown, expected: unknown): boolean {
+  const findOperator: {
+    getSql?: (alias: string) => string;
+    objectLiteralParameters?: Record<string, unknown>;
+  } = expected as {
+    getSql?: (alias: string) => string;
+    objectLiteralParameters?: Record<string, unknown>;
+  };
+
+  if (findOperator && typeof findOperator.getSql === "function") {
+    const sql: string = findOperator.getSql("column");
+    const parameters: Array<unknown> = Object.values(
+      findOperator.objectLiteralParameters || {},
+    ).flat();
+
+    if (sql.includes("IS NOT NULL")) {
+      return actual !== null && actual !== undefined;
+    }
+
+    if (sql.includes("IS NULL")) {
+      return actual === null || actual === undefined;
+    }
+
+    if (sql.includes("IN (")) {
+      return parameters.some((parameter: unknown): boolean => {
+        return parameter?.toString() === actual?.toString();
+      });
+    }
+
+    if (sql.includes(">")) {
+      return (
+        actual !== null &&
+        actual !== undefined &&
+        String(actual) > String(parameters[0])
+      );
+    }
+
+    throw new Error(`Unhandled query operator in test matcher: ${sql}`);
+  }
+
+  if (expected === null || expected === undefined) {
+    return actual === null || actual === undefined;
+  }
+
+  if (typeof expected === "boolean") {
+    return Boolean(actual) === expected;
+  }
+
+  return actual?.toString() === expected.toString();
+}
+
+function matchesQuery(
+  row: Record<string, unknown>,
+  query: Record<string, unknown>,
+): boolean {
+  for (const [column, expected] of Object.entries(query)) {
+    if (column === "labels") {
+      /*
+       * The many-to-many spelling: `labels: [id, id]` is "carries any of
+       * these labels", which is exactly what the scope means by a label list.
+       */
+      const wanted: Array<unknown> = expected as Array<unknown>;
+      const carried: Array<Label> = (row["labels"] as Array<Label>) || [];
+
+      const hasAny: boolean = carried.some((label: Label): boolean => {
+        return wanted.some((id: unknown): boolean => {
+          return id?.toString() === label.id?.toString();
+        });
+      });
+
+      if (!hasAny) {
+        return false;
+      }
+
+      continue;
+    }
+
+    const actual: unknown = column === "_id" ? row["_id"] : row[column];
+
+    if (!matchesQueryValue(actual, expected)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+// The fake tables. Rebuilt per test.
+let monitorTable: Array<Monitor> = [];
+let deviceTable: Array<NetworkDevice> = [];
+let policyTable: Array<NetworkAlertPolicy> = [];
+
+function sortAndPage<TRow extends { _id?: string }>(
+  rows: Array<TRow>,
+  findBy: FindBy<never>,
+): Array<TRow> {
+  const sorted: Array<TRow> = [...rows].sort(
+    (left: TRow, right: TRow): number => {
+      return (left._id || "").localeCompare(right._id || "");
+    },
+  );
+
+  const limit: number =
+    typeof findBy.limit === "number" ? findBy.limit : sorted.length;
+
+  return sorted.slice(0, limit);
+}
+
+function installFakeTables(): void {
+  monitorFindByMock.mockImplementation(
+    async (findBy: FindBy<Monitor>): Promise<Array<Monitor>> => {
+      return sortAndPage(
+        monitorTable.filter((monitor: Monitor): boolean => {
+          return matchesQuery(
+            monitor as unknown as Record<string, unknown>,
+            findBy.query as Record<string, unknown>,
+          );
+        }),
+        findBy as unknown as FindBy<never>,
+      );
+    },
+  );
+
+  monitorFindOneByMock.mockImplementation(
+    async (findOneBy: FindOneBy<Monitor>): Promise<Monitor | null> => {
+      return (
+        monitorTable.find((monitor: Monitor): boolean => {
+          return matchesQuery(
+            monitor as unknown as Record<string, unknown>,
+            findOneBy.query as Record<string, unknown>,
+          );
+        }) || null
+      );
+    },
+  );
+
+  deviceFindByMock.mockImplementation(
+    async (findBy: FindBy<NetworkDevice>): Promise<Array<NetworkDevice>> => {
+      return sortAndPage(
+        deviceTable.filter((device: NetworkDevice): boolean => {
+          return matchesQuery(
+            device as unknown as Record<string, unknown>,
+            findBy.query as Record<string, unknown>,
+          );
+        }),
+        findBy as unknown as FindBy<never>,
+      );
+    },
+  );
+
+  deviceFindOneByMock.mockImplementation(
+    async (
+      findOneBy: FindOneBy<NetworkDevice>,
+    ): Promise<NetworkDevice | null> => {
+      return (
+        deviceTable.find((device: NetworkDevice): boolean => {
+          return matchesQuery(
+            device as unknown as Record<string, unknown>,
+            findOneBy.query as Record<string, unknown>,
+          );
+        }) || null
+      );
+    },
+  );
+
+  policyFindByMock.mockImplementation(
+    async (
+      findBy: FindBy<NetworkAlertPolicy>,
+    ): Promise<Array<NetworkAlertPolicy>> => {
+      return policyTable.filter((policy: NetworkAlertPolicy): boolean => {
+        return matchesQuery(
+          policy as unknown as Record<string, unknown>,
+          findBy.query as Record<string, unknown>,
+        );
+      });
+    },
+  );
+
+  policyFindOneByMock.mockImplementation(
+    async (
+      findOneBy: FindOneBy<NetworkAlertPolicy>,
+    ): Promise<NetworkAlertPolicy | null> => {
+      return (
+        policyTable.find((policy: NetworkAlertPolicy): boolean => {
+          return matchesQuery(
+            policy as unknown as Record<string, unknown>,
+            findOneBy.query as Record<string, unknown>,
+          );
+        }) || null
+      );
+    },
+  );
+
+  /*
+   * Writes land back in the fake tables so a second pass in the same test
+   * sees what the first one did — which is the only way to assert that
+   * reconciliation is idempotent.
+   */
+  monitorCreateMock.mockImplementation(
+    async (createBy: { data: Monitor }): Promise<Monitor> => {
+      const monitor: Monitor = createBy.data;
+      monitor.id = ObjectID.generate();
+      monitorTable.push(monitor);
+
+      return monitor;
+    },
+  );
+
+  monitorDeleteByMock.mockImplementation(
+    async (deleteBy: { query: Record<string, unknown> }): Promise<number> => {
+      const remaining: Array<Monitor> = monitorTable.filter(
+        (monitor: Monitor): boolean => {
+          return !matchesQuery(
+            monitor as unknown as Record<string, unknown>,
+            deleteBy.query,
+          );
+        },
+      );
+
+      const deleted: number = monitorTable.length - remaining.length;
+      monitorTable = remaining;
+
+      return deleted;
+    },
+  );
+
+  monitorUpdateByMock.mockImplementation(
+    async (updateBy: {
+      query: Record<string, unknown>;
+      data: Record<string, unknown>;
+    }): Promise<number> => {
+      let updated: number = 0;
+
+      for (const monitor of monitorTable) {
+        if (
+          matchesQuery(
+            monitor as unknown as Record<string, unknown>,
+            updateBy.query,
+          )
+        ) {
+          Object.assign(monitor, updateBy.data);
+          updated++;
+        }
+      }
+
+      return updated;
+    },
+  );
+
+  monitorStampMock.mockImplementation(
+    async (input: {
+      id: ObjectID;
+      data: Record<string, unknown>;
+    }): Promise<void> => {
+      const monitor: Monitor | undefined = monitorTable.find(
+        (candidate: Monitor): boolean => {
+          return candidate.id?.toString() === input.id.toString();
+        },
+      );
+
+      if (monitor) {
+        Object.assign(monitor, input.data);
+      }
+    },
+  );
+
+  policyStampMock.mockResolvedValue(undefined);
+  templateFindOneByMock.mockImplementation(
+    async (
+      findOneBy: FindOneBy<MonitorTemplate>,
+    ): Promise<MonitorTemplate | null> => {
+      const queriedId: unknown = (findOneBy.query as Record<string, unknown>)[
+        "_id"
+      ];
+
+      return queriedId?.toString() === TEMPLATE_ID.toString()
+        ? monitorTemplate(TEMPLATE_ID, "Reachability")
+        : queriedId?.toString() === OTHER_TEMPLATE_ID.toString()
+          ? monitorTemplate(OTHER_TEMPLATE_ID, "Interfaces")
+          : null;
+    },
+  );
+
+  semaphoreLockMock.mockResolvedValue(FAKE_MUTEX);
+  semaphoreReleaseMock.mockResolvedValue(undefined);
+  currentPlanMock.mockResolvedValue({
+    plan: PlanType.Growth,
+    isSubscriptionUnpaid: false,
+  });
+  monitorCountByMock.mockResolvedValue(new PositiveNumber(0));
+}
+
+function monitorTemplate(id: ObjectID, name: string): MonitorTemplate {
+  const template: MonitorTemplate = new MonitorTemplate();
+  template.id = id;
+  template.projectId = PROJECT_ID;
+  template.monitorType = MonitorType.NetworkDevice;
+  template.monitorName = name;
+
+  const step: MonitorStep = new MonitorStep();
+  step.data!.networkDeviceMonitor = {
+    // A design-time placeholder; buildMonitor rebinds it to the real device.
+    networkDeviceId: ObjectID.generate().toString(),
+    monitorInterfaces: false,
+    oids: [],
+  };
+  template.monitorSteps = new MonitorSteps();
+  template.monitorSteps.data = { monitorStepsInstanceArray: [step] };
+
+  return template;
+}
+
+function device(
+  overrides: Partial<NetworkDevice> = {},
+  id: ObjectID = DEVICE_ID,
+): NetworkDevice {
+  const networkDevice: NetworkDevice = new NetworkDevice();
+  networkDevice.id = id;
+  networkDevice.projectId = PROJECT_ID;
+  networkDevice.name = "Core switch";
+  networkDevice.isArchived = false;
+  networkDevice.monitoringMethod = NetworkDeviceMonitoringMethod.Probe;
+  networkDevice.probeId = PROBE_ID;
+  networkDevice.siteId = SITE_ID;
+  Object.assign(networkDevice, overrides);
+
+  return networkDevice;
+}
+
+function policy(
+  overrides: Partial<NetworkAlertPolicy> = {},
+  id: ObjectID = POLICY_ID,
+): NetworkAlertPolicy {
+  const alertPolicy: NetworkAlertPolicy = new NetworkAlertPolicy();
+  alertPolicy.id = id;
+  alertPolicy.projectId = PROJECT_ID;
+  alertPolicy.name = "Warehouse switches";
+  alertPolicy.isEnabled = true;
+  alertPolicy.monitorTemplateId = TEMPLATE_ID;
+  alertPolicy.scope = {} as NetworkAlertPolicyScope;
+  Object.assign(alertPolicy, overrides);
+
+  return alertPolicy;
+}
+
+function monitorRow(overrides: Partial<Monitor> = {}): Monitor {
+  const monitor: Monitor = new Monitor();
+  monitor.id = ObjectID.generate();
+  monitor.projectId = PROJECT_ID;
+  monitor.monitorType = MonitorType.NetworkDevice;
+  monitor.disableActiveMonitoring = false;
+  Object.assign(monitor, overrides);
+
+  return monitor;
+}
+
+function labelled(labelId: ObjectID): Array<Label> {
+  const label: Label = new Label();
+  label.id = labelId;
+
+  return [label];
+}
+
+function monitorsOfDevice(deviceId: ObjectID = DEVICE_ID): Array<Monitor> {
+  return monitorTable.filter((monitor: Monitor): boolean => {
+    return (
+      monitor.autoProvisionedNetworkDeviceId?.toString() === deviceId.toString()
+    );
+  });
+}
+
+beforeEach(() => {
+  jest.clearAllMocks();
+  monitorTable = [];
+  deviceTable = [];
+  policyTable = [];
+  installFakeTables();
+});
+
+describe("reconcileDevice provisions what the policies ask for", () => {
+  it("creates one monitor for a device an enabled policy covers", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+    expect(monitorsOfDevice()[0]!.networkAlertPolicyId?.toString()).toBe(
+      POLICY_ID.toString(),
+    );
+    expect(monitorsOfDevice()[0]!.monitorTemplateId?.toString()).toBe(
+      TEMPLATE_ID.toString(),
+    );
+  });
+
+  /*
+   * The stamp goes on the CREATE, not in a follow-up update. A monitor that
+   * exists for a moment without its owner is one a concurrent pass of a
+   * different policy could adopt.
+   */
+  it("stamps the owning policy in the same write that creates the monitor", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const created: { data: Monitor } = monitorCreateMock.mock.calls[0]![0] as {
+      data: Monitor;
+    };
+
+    expect(created.data.networkAlertPolicyId?.toString()).toBe(
+      POLICY_ID.toString(),
+    );
+  });
+
+  it("is idempotent: a second pass over an already-covered device writes nothing", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+    monitorCreateMock.mockClear();
+    monitorDeleteByMock.mockClear();
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(monitorDeleteByMock).not.toHaveBeenCalled();
+    expect(monitorsOfDevice()).toHaveLength(1);
+  });
+
+  it("provisions nothing for a device outside every policy's scope", async () => {
+    deviceTable = [device({ siteId: OTHER_SITE_ID })];
+    policyTable = [policy({ scope: { siteIds: [SITE_ID.toString()] } })];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+  });
+
+  it("matches a device by label as well as by site", async () => {
+    deviceTable = [device({ labels: labelled(LABEL_ID) })];
+    policyTable = [policy({ scope: { labelIds: [LABEL_ID.toString()] } })];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+  });
+});
+
+/*
+ * THE DANGEROUS CASE. Two policies covering one device is the ordinary
+ * situation ("everything gets reachability" plus "warehouse switches get
+ * interface alerts"), and it is the one where a wrong ownership key destroys
+ * a monitor: if the engine keyed on the DEVICE rather than on (policy,
+ * template), each pass would see a monitor that is "not mine" and delete it,
+ * and the two policies would take turns removing each other's alerting.
+ */
+describe("two policies on one device", () => {
+  beforeEach(() => {
+    deviceTable = [device()];
+    policyTable = [
+      policy(),
+      policy(
+        {
+          name: "Interface alerts",
+          monitorTemplateId: OTHER_TEMPLATE_ID,
+        },
+        OTHER_POLICY_ID,
+      ),
+    ];
+  });
+
+  it("gives the device one monitor per policy", async () => {
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const owners: Array<string> = monitorsOfDevice().map(
+      (monitor: Monitor): string => {
+        return monitor.networkAlertPolicyId!.toString();
+      },
+    );
+
+    expect(owners.sort()).toEqual(
+      [POLICY_ID.toString(), OTHER_POLICY_ID.toString()].sort(),
+    );
+  });
+
+  it("leaves the other policy's monitor alone when one policy's scope stops matching", async () => {
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    // The first policy narrows to a site this device is not in.
+    policyTable[0]!.scope = {
+      siteIds: [OTHER_SITE_ID.toString()],
+    } as NetworkAlertPolicyScope;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const survivors: Array<Monitor> = monitorsOfDevice();
+
+    expect(survivors).toHaveLength(1);
+    expect(survivors[0]!.networkAlertPolicyId?.toString()).toBe(
+      OTHER_POLICY_ID.toString(),
+    );
+  });
+});
+
+/*
+ * THE OTHER DANGEROUS CASE. A shrinking scope is the ordinary way a policy
+ * stops covering a device, and it is the moment the engine deletes. What it
+ * must delete is exactly the rows stamped with the policy that shrank —
+ * never the monitor a person built by hand, and never the one an auto-import
+ * rule provisioned, both of which sit on the same device and carry the same
+ * `autoProvisionedNetworkDeviceId`.
+ */
+describe("a scope that shrinks", () => {
+  it("deletes only the monitors the shrinking policy owns", async () => {
+    const handMade: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      // No networkAlertPolicyId: nobody's policy made this.
+      monitorTemplateId: OTHER_TEMPLATE_ID,
+    });
+
+    monitorTable = [handMade];
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(2);
+
+    policyTable[0]!.scope = {
+      siteIds: [OTHER_SITE_ID.toString()],
+    } as NetworkAlertPolicyScope;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const survivors: Array<Monitor> = monitorsOfDevice();
+
+    expect(survivors).toHaveLength(1);
+    expect(survivors[0]!.id?.toString()).toBe(handMade.id?.toString());
+  });
+
+  /*
+   * ...and the hand-made monitor is not merely filtered out afterwards: the
+   * QUERY that reads ACTUAL asks for `networkAlertPolicyId IS NOT NULL`, so
+   * a monitor with no owner is never even a candidate for deletion.
+   */
+  it("never reads an unowned monitor into the set it may delete", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const actualQuery: Record<string, unknown> = (
+      monitorFindByMock.mock.calls[0]![0] as {
+        query: Record<string, unknown>;
+      }
+    ).query;
+
+    expect(
+      (
+        actualQuery["networkAlertPolicyId"] as {
+          getSql: (alias: string) => string;
+        }
+      ).getSql("column"),
+    ).toContain("IS NOT NULL");
+    expect(actualQuery["autoProvisionedNetworkDeviceId"]?.toString()).toBe(
+      DEVICE_ID.toString(),
+    );
+  });
+});
+
+/*
+ * Adoption. A (device, template) monitor that already exists without an owner
+ * occupies the pair Monitor's partial unique index protects, so creating a
+ * second one is impossible — not slow, impossible, on this pass and on every
+ * pass after it.
+ */
+describe("an unowned monitor for the same (device, template)", () => {
+  it("is adopted by stamping the policy, not duplicated", async () => {
+    const unowned: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: TEMPLATE_ID,
+    });
+
+    monitorTable = [unowned];
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(monitorsOfDevice()).toHaveLength(1);
+    expect(unowned.networkAlertPolicyId?.toString()).toBe(POLICY_ID.toString());
+  });
+
+  /*
+   * The stamp is a provenance marker, not an operational change: no hooks,
+   * no workflow, no realtime event, no audit row. The monitor keeps doing
+   * exactly what it was doing.
+   */
+  it("is adopted without running the monitor update pipeline", async () => {
+    monitorTable = [
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+      }),
+    ];
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({
+          networkAlertPolicyId: expect.anything(),
+        }),
+      }),
+    );
+    expect(monitorUpdateByMock).not.toHaveBeenCalled();
+  });
+
+  /*
+   * A monitor another policy owns is not adoptable and not deletable. The
+   * one-policy-per-template rule makes this unreachable through the API, but
+   * the engine must never be the thing that turns a data anomaly into a lost
+   * monitor.
+   */
+  it("is left alone when a different policy already owns it", async () => {
+    const foreign: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: TEMPLATE_ID,
+      networkAlertPolicyId: OTHER_POLICY_ID,
+    });
+
+    monitorTable = [foreign];
+    deviceTable = [device()];
+    policyTable = [
+      policy(),
+      policy(
+        { name: "Other", scope: { siteIds: [SITE_ID.toString()] } },
+        OTHER_POLICY_ID,
+      ),
+    ];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(foreign.networkAlertPolicyId?.toString()).toBe(
+      OTHER_POLICY_ID.toString(),
+    );
+    expect(monitorTable).toContain(foreign);
+  });
+});
+
+/*
+ * Provisionability is a property of the DEVICE, and losing it removes every
+ * policy monitor. The monitor-backed case is the one that matters most: the
+ * device's health is now a bound monitor's, nothing polls it, and a Network
+ * Device monitor left behind is billed for a verdict that will never arrive.
+ */
+describe("a device that stops being provisionable", () => {
+  async function reconcileAfter(
+    change: Partial<NetworkDevice>,
+  ): Promise<Array<Monitor>> {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+
+    Object.assign(deviceTable[0]!, change);
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    return monitorsOfDevice();
+  }
+
+  it("loses its policy monitors when it becomes monitor-backed", async () => {
+    expect(
+      await reconcileAfter({
+        monitoringMethod: NetworkDeviceMonitoringMethod.Monitor,
+      }),
+    ).toHaveLength(0);
+  });
+
+  it("loses its policy monitors when it is archived", async () => {
+    expect(await reconcileAfter({ isArchived: true })).toHaveLength(0);
+  });
+
+  /*
+   * A device with no probe is polled by nothing, so its monitor would sit
+   * Pending forever — and still be billed.
+   */
+  it("loses its policy monitors when its probe is cleared", async () => {
+    expect(
+      await reconcileAfter({ probeId: undefined as unknown as ObjectID }),
+    ).toHaveLength(0);
+  });
+
+  it("gets them back when it becomes probe-polled again", async () => {
+    deviceTable = [
+      device({ monitoringMethod: NetworkDeviceMonitoringMethod.Monitor }),
+    ];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(0);
+
+    deviceTable[0]!.monitoringMethod = NetworkDeviceMonitoringMethod.Probe;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+  });
+
+  /*
+   * The legacy free-text values — NULL, "" and "SNMP" — all mean "the probe
+   * polls it". A device written before ping-first polling must not be read
+   * as monitor-backed and quietly stripped of its alerting.
+   */
+  it("treats a legacy SNMP monitoringMethod as probe-polled", async () => {
+    deviceTable = [device({ monitoringMethod: "SNMP" })];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+  });
+});
+
+describe("a disabled policy", () => {
+  it("keeps its monitors and pauses them rather than deleting them", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    policyTable[0]!.isEnabled = false;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    /*
+     * The row survives — with its id, its incident history and its
+     * dependencies — which is the whole difference between pausing a policy
+     * and deleting one.
+     */
+    expect(monitorsOfDevice()).toHaveLength(1);
+    expect(monitorsOfDevice()[0]!.disableActiveMonitoring).toBe(true);
+  });
+
+  it("provisions nothing new while it is off", async () => {
+    deviceTable = [device()];
+    policyTable = [policy({ isEnabled: false })];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+  });
+
+  /*
+   * A template-less policy — only reachable through the FK's SET NULL
+   * backstop, because MonitorTemplateService refuses to delete a template a
+   * policy uses — is treated exactly like a disabled one. Nothing to
+   * provision from, so nothing is provisioned and nothing is destroyed.
+   */
+  it("behaves the same when its template was deleted out from under it", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    delete policyTable[0]!.monitorTemplateId;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(1);
+  });
+});
+
+/*
+ * Re-pointing a policy at a different template is a re-clone: the monitors
+ * from the old template are the old alerting configuration and must go, and
+ * new ones are provisioned from the template the policy names now. This is
+ * why ownership is the pair, not the policy id alone.
+ */
+describe("a policy re-pointed at a different template", () => {
+  it("replaces the monitors made from the old template", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const originalMonitorId: string = monitorsOfDevice()[0]!.id!.toString();
+
+    policyTable[0]!.monitorTemplateId = OTHER_TEMPLATE_ID;
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    const survivors: Array<Monitor> = monitorsOfDevice();
+
+    expect(survivors).toHaveLength(1);
+    expect(survivors[0]!.id!.toString()).not.toBe(originalMonitorId);
+    expect(survivors[0]!.monitorTemplateId?.toString()).toBe(
+      OTHER_TEMPLATE_ID.toString(),
+    );
+  });
+});
+
+/*
+ * A monitor whose owning policy row is gone is nobody's. In practice the
+ * policy delete path removed it already; this is the backstop for the row
+ * that outlived it, and without it that monitor would be billed forever with
+ * nothing left to explain it.
+ */
+describe("a monitor whose policy no longer exists", () => {
+  it("is removed on the next pass", async () => {
+    monitorTable = [
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+        networkAlertPolicyId: OTHER_POLICY_ID,
+      }),
+    ];
+    deviceTable = [device()];
+    policyTable = [];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(monitorsOfDevice()).toHaveLength(0);
+  });
+});
+
+describe("the per-device lock", () => {
+  it("is taken on the device, by name, before anything is read", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(semaphoreLockMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        key: policyDeviceLockKey(DEVICE_ID),
+        namespace: POLICY_DEVICE_LOCK_NAMESPACE,
+        /*
+         * One attempt, never a queue: whoever holds it is computing the same
+         * difference from fresher data than this pass would.
+         */
+        acquireAttemptsLimit: 1,
+      }),
+    );
+    expect(policyDeviceLockKey(DEVICE_ID)).toBe(
+      `NetworkAlertPolicy:Device:${DEVICE_ID.toString()}`,
+    );
+  });
+
+  /*
+   * A pass that cannot take the lock does NOTHING — it does not read, and
+   * above all it does not write. That is what makes "two policies racing on
+   * one device" safe: the loser of the race is not a second writer, it is a
+   * no-op.
+   */
+  it("makes a pass that cannot acquire it a complete no-op", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    semaphoreLockMock.mockRejectedValue(new Error("lock held"));
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(deviceFindOneByMock).not.toHaveBeenCalled();
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(monitorDeleteByMock).not.toHaveBeenCalled();
+  });
+
+  it("is released even when the reconciliation throws", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    deviceFindOneByMock.mockRejectedValue(new Error("database is on fire"));
+
+    await NetworkAlertPolicyEngineService.reconcileDevice({
+      projectId: PROJECT_ID,
+      deviceId: DEVICE_ID,
+    });
+
+    expect(semaphoreReleaseMock).toHaveBeenCalledWith(FAKE_MUTEX);
+  });
+
+  /*
+   * ...and the throw is recorded rather than propagated. reconcileDevice is
+   * called from a device create's detached chain; a template that stopped
+   * validating must not become a failed device import.
+   */
+  it("records the failure instead of throwing it at the caller", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    deviceFindOneByMock.mockRejectedValue(new Error("database is on fire"));
+
+    const context: PolicyRunContext =
+      await NetworkAlertPolicyEngineService.reconcileDevice({
+        projectId: PROJECT_ID,
+        deviceId: DEVICE_ID,
+      });
+
+    expect(context.failures[0]).toContain("database is on fire");
+  });
+});
+
+/*
+ * Prerequisite 5. A project whose plan cannot hold another monitor must
+ * produce ONE message an operator can act on, not five hundred identical
+ * create failures — and the run must stop, because every remaining create
+ * would fail the same way.
+ */
+describe("the plan check", () => {
+  it("stops the run before the first create when the subscription is unpaid", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    currentPlanMock.mockResolvedValue({
+      plan: PlanType.Growth,
+      isSubscriptionUnpaid: true,
+    });
+
+    const context: PolicyRunContext =
+      await NetworkAlertPolicyEngineService.reconcileDevice({
+        projectId: PROJECT_ID,
+        deviceId: DEVICE_ID,
+      });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(context.isStopped).toBe(true);
+    expect(context.planException).toContain("subscription is unpaid");
+  });
+
+  it("stops the run when the free plan's monitor limit is already reached", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    currentPlanMock.mockResolvedValue({
+      plan: PlanType.Free,
+      isSubscriptionUnpaid: false,
+    });
+    monitorCountByMock.mockResolvedValue(new PositiveNumber(10));
+
+    const context: PolicyRunContext =
+      await NetworkAlertPolicyEngineService.reconcileDevice({
+        projectId: PROJECT_ID,
+        deviceId: DEVICE_ID,
+      });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(context.planException).toContain("free plan");
+  });
+
+  /*
+   * Asked ONCE per run. Fifty devices in one sweep must not be fifty billing
+   * lookups; that is the difference between a five-minute sweep and a
+   * five-minute outage.
+   */
+  it("asks the plan once for a whole run", async () => {
+    deviceTable = [device(), device({}, OTHER_DEVICE_ID)];
+    policyTable = [policy()];
+
+    const context: PolicyRunContext =
+      NetworkAlertPolicyEngineService.createRunContext(PROJECT_ID);
+
+    await NetworkAlertPolicyEngineService.reconcileDevices({
+      projectId: PROJECT_ID,
+      deviceIds: [DEVICE_ID, OTHER_DEVICE_ID],
+      context: context,
+    });
+
+    expect(monitorCreateMock).toHaveBeenCalledTimes(2);
+    expect(currentPlanMock).toHaveBeenCalledTimes(1);
+  });
+
+  /*
+   * The ceiling can be crossed DURING a run: the precheck passed at nine of
+   * ten monitors and the tenth create is the one that is refused. The verdict
+   * is re-asked after a failed create — never pattern-matched out of the
+   * error text, which would drift the moment MonitorService reworded its
+   * refusal — and the run stops at the first refusal instead of the five
+   * hundredth.
+   */
+  it("stops mid-run when a create fails and the plan now refuses another monitor", async () => {
+    deviceTable = [device(), device({}, OTHER_DEVICE_ID)];
+    policyTable = [policy()];
+    currentPlanMock.mockResolvedValue({
+      plan: PlanType.Free,
+      isSubscriptionUnpaid: false,
+    });
+    monitorCountByMock
+      .mockResolvedValueOnce(new PositiveNumber(9))
+      .mockResolvedValue(new PositiveNumber(10));
+    monitorCreateMock.mockRejectedValueOnce(
+      new BadDataException("monitor limit reached"),
+    );
+
+    const context: PolicyRunContext =
+      NetworkAlertPolicyEngineService.createRunContext(PROJECT_ID);
+
+    await NetworkAlertPolicyEngineService.reconcileDevices({
+      projectId: PROJECT_ID,
+      deviceIds: [DEVICE_ID, OTHER_DEVICE_ID],
+      context: context,
+    });
+
+    expect(context.isStopped).toBe(true);
+    expect(context.planException).toContain("free plan");
+    // The second device was never attempted.
+    expect(monitorCreateMock).toHaveBeenCalledTimes(1);
+  });
+
+  /*
+   * A billing lookup that FAILS is not a plan refusal. A Stripe blip must not
+   * stop a fleet from being provisioned; it is recorded and the run carries
+   * on.
+   */
+  it("treats a billing lookup failure as an ordinary failure, not a refusal", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    currentPlanMock.mockRejectedValue(new Error("stripe timed out"));
+
+    const context: PolicyRunContext =
+      await NetworkAlertPolicyEngineService.reconcileDevice({
+        projectId: PROJECT_ID,
+        deviceId: DEVICE_ID,
+      });
+
+    expect(context.isStopped).toBe(false);
+    expect(context.failures[0]).toContain("stripe timed out");
+    expect(monitorCreateMock).toHaveBeenCalledTimes(1);
+  });
+});
+
+describe("the run's monitor budget", () => {
+  it("stops writing once MAX_MONITORS_PER_POLICY_SYNC monitors have been written", async () => {
+    deviceTable = [device(), device({}, OTHER_DEVICE_ID)];
+    policyTable = [policy()];
+
+    const context: PolicyRunContext =
+      NetworkAlertPolicyEngineService.createRunContext(PROJECT_ID, 1);
+
+    await NetworkAlertPolicyEngineService.reconcileDevices({
+      projectId: PROJECT_ID,
+      deviceIds: [DEVICE_ID, OTHER_DEVICE_ID],
+      context: context,
+    });
+
+    /*
+     * One monitor, and the run marked truncated rather than failed: nothing
+     * is half-written, and the next sweep computes the same difference and
+     * continues.
+     */
+    expect(monitorCreateMock).toHaveBeenCalledTimes(1);
+    expect(context.isTruncated).toBe(true);
+    expect(context.failures).toHaveLength(0);
+  });
+
+  it("defaults to the documented cap", () => {
+    expect(
+      NetworkAlertPolicyEngineService.createRunContext(PROJECT_ID)
+        .monitorBudget,
+    ).toBe(MAX_MONITORS_PER_POLICY_SYNC);
+  });
+});
+
+describe("syncPolicy over a policy's whole fleet", () => {
+  it("provisions every matching device and stamps the covered count", async () => {
+    deviceTable = [device(), device({}, OTHER_DEVICE_ID)];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    expect(monitorTable).toHaveLength(2);
+    expect(policyStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        id: POLICY_ID,
+        data: expect.objectContaining({
+          coveredDeviceCount: 2,
+          lastSyncError: null,
+        }),
+      }),
+    );
+  });
+
+  /*
+   * The covered count is what the settings table prints beside the scope
+   * sentence, so it counts the devices the policy actually provisions for —
+   * not every device the scope's SQL half matches. An archived or
+   * monitor-backed device is in the site, and gets no monitor.
+   */
+  it("counts only the devices it can actually provision for", async () => {
+    deviceTable = [
+      device(),
+      device(
+        { monitoringMethod: NetworkDeviceMonitoringMethod.Monitor },
+        OTHER_DEVICE_ID,
+      ),
+    ];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    expect(policyStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({ coveredDeviceCount: 1 }),
+      }),
+    );
+  });
+
+  it("removes the monitors of devices that have left the scope", async () => {
+    deviceTable = [device(), device({}, OTHER_DEVICE_ID)];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+    expect(monitorTable).toHaveLength(2);
+
+    // The second device moves to a site the policy does not cover.
+    deviceTable[1]!.siteId = OTHER_SITE_ID;
+    policyTable[0]!.scope = {
+      siteIds: [SITE_ID.toString()],
+    } as NetworkAlertPolicyScope;
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    expect(monitorTable).toHaveLength(1);
+    expect(monitorTable[0]!.autoProvisionedNetworkDeviceId?.toString()).toBe(
+      DEVICE_ID.toString(),
+    );
+  });
+
+  it("records a failure in lastSyncError where the operator will see it", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+    // The template vanished between the policy's save and this pass.
+    templateFindOneByMock.mockResolvedValue(null);
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    expect(policyStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({
+          lastSyncError: expect.stringContaining("Monitor template"),
+        }),
+      }),
+    );
+  });
+
+  /*
+   * A disabled policy is looked at and found to have nothing to do. Stamping
+   * lastSyncAt for it is honest — and leaving lastSyncError null is what
+   * stops the settings table showing a red row for a policy somebody
+   * deliberately switched off.
+   */
+  it("stamps a disabled policy without touching a single monitor", async () => {
+    deviceTable = [device()];
+    policyTable = [policy({ isEnabled: false })];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+    expect(monitorDeleteByMock).not.toHaveBeenCalled();
+    expect(policyStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({ lastSyncError: null }),
+      }),
+    );
+  });
+
+  it("does nothing at all for a policy that has been deleted", async () => {
+    policyTable = [];
+
+    await expect(
+      NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID }),
+    ).resolves.toBeNull();
+
+    expect(policyStampMock).not.toHaveBeenCalled();
+  });
+
+  /*
+   * templateSyncedAt means "every monitor this policy owns is running the
+   * template's current configuration". A sweep that changed nothing has
+   * re-synced nothing, so it must not stamp it — otherwise an operator would
+   * read "Template Synced: 2 minutes ago" and believe a criteria edit they
+   * made an hour ago had landed.
+   */
+  it("does not stamp templateSyncedAt on an ordinary sweep", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    const stamped: Record<string, unknown> = (
+      policyStampMock.mock.calls[0]![0] as { data: Record<string, unknown> }
+    ).data;
+
+    expect(stamped["templateSyncedAt"]).toBeUndefined();
+  });
+
+  it("stamps templateSyncedAt when the fleet was rebuilt from a new template", async () => {
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({
+      policyId: POLICY_ID,
+      stampTemplateSyncedOnCleanPass: true,
+    });
+
+    const stamped: Record<string, unknown> = (
+      policyStampMock.mock.calls[0]![0] as { data: Record<string, unknown> }
+    ).data;
+
+    expect(stamped["templateSyncedAt"]).toBeInstanceOf(Date);
+  });
+
+  /*
+   * ...but not when the pass ADOPTED a monitor. An adopted monitor was built
+   * by somebody else and its criteria are not this template's, so claiming
+   * the fleet is in step with the template would be false.
+   */
+  it("does not stamp templateSyncedAt for a pass that adopted a monitor", async () => {
+    monitorTable = [
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+      }),
+    ];
+    deviceTable = [device()];
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({
+      policyId: POLICY_ID,
+      stampTemplateSyncedOnCleanPass: true,
+    });
+
+    const stamped: Record<string, unknown> = (
+      policyStampMock.mock.calls[0]![0] as { data: Record<string, unknown> }
+    ).data;
+
+    expect(stamped["templateSyncedAt"]).toBeUndefined();
+  });
+});
+
+/*
+ * Prerequisite 9's engine half. A scope is validated at write, so a stale or
+ * malformed id only ever arrives here through a hand-edited row — and the
+ * engine's answer is "match nothing", never "throw" and never "match
+ * everything".
+ */
+describe("scope ids the engine cannot use", () => {
+  it("matches nothing when a kind holds only unusable ids", async () => {
+    deviceTable = [device()];
+    policyTable = [
+      policy({
+        scope: { siteIds: ["not-a-uuid"] } as NetworkAlertPolicyScope,
+      }),
+    ];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    /*
+     * Dropping the last unusable id and leaving the kind EMPTY would be the
+     * dangerous bug: an empty kind matches every device, so a typo'd scope
+     * would silently provision the whole estate.
+     */
+    expect(monitorCreateMock).not.toHaveBeenCalled();
+  });
+
+  it("adds no clause at all for a kind that is genuinely empty", async () => {
+    deviceTable = [device()];
+    policyTable = [policy({ scope: {} as NetworkAlertPolicyScope })];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    const deviceQuery: Record<string, unknown> = (
+      deviceFindByMock.mock.calls[0]![0] as { query: Record<string, unknown> }
+    ).query;
+
+    expect(deviceQuery["siteId"]).toBeUndefined();
+    expect(deviceQuery["networkDeviceRoleId"]).toBeUndefined();
+    expect(deviceQuery["labels"]).toBeUndefined();
+    expect(monitorCreateMock).toHaveBeenCalledTimes(1);
+  });
+
+  // The engine reads a scope. It never writes one back.
+  it("never rewrites the policy's scope", async () => {
+    deviceTable = [device()];
+    policyTable = [
+      policy({
+        scope: { siteIds: ["not-a-uuid"] } as NetworkAlertPolicyScope,
+      }),
+    ];
+
+    await NetworkAlertPolicyEngineService.syncPolicy({ policyId: POLICY_ID });
+
+    for (const call of policyStampMock.mock.calls) {
+      expect(
+        (call[0] as { data: Record<string, unknown> }).data["scope"],
+      ).toBeUndefined();
+    }
+
+    expect(policyTable[0]!.scope).toEqual({ siteIds: ["not-a-uuid"] });
+  });
+});
+
+describe("deleting a policy", () => {
+  it("removes exactly the monitors it owns, as root", async () => {
+    const ownMonitor: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: TEMPLATE_ID,
+      networkAlertPolicyId: POLICY_ID,
+    });
+    const otherPolicyMonitor: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: OTHER_DEVICE_ID,
+      monitorTemplateId: OTHER_TEMPLATE_ID,
+      networkAlertPolicyId: OTHER_POLICY_ID,
+    });
+    const handMade: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: OTHER_TEMPLATE_ID,
+    });
+
+    monitorTable = [ownMonitor, otherPolicyMonitor, handMade];
+
+    const deleted: number =
+      await NetworkAlertPolicyEngineService.deleteMonitorsOwnedByPolicy({
+        projectId: PROJECT_ID,
+        policyId: POLICY_ID,
+      });
+
+    expect(deleted).toBe(1);
+    expect(monitorTable).toEqual([otherPolicyMonitor, handMade]);
+    expect(monitorDeleteByMock).toHaveBeenCalledWith(
+      expect.objectContaining({ props: { isRoot: true } }),
+    );
+  });
+
+  it("does nothing when the policy owns no monitors", async () => {
+    monitorTable = [
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+      }),
+    ];
+
+    expect(
+      await NetworkAlertPolicyEngineService.deleteMonitorsOwnedByPolicy({
+        projectId: PROJECT_ID,
+        policyId: POLICY_ID,
+      }),
+    ).toBe(0);
+    expect(monitorDeleteByMock).not.toHaveBeenCalled();
+  });
+});
+
+describe("deleting devices", () => {
+  it("removes their policy-owned monitors and leaves the rest to the caller", async () => {
+    const policyMonitor: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: TEMPLATE_ID,
+      networkAlertPolicyId: POLICY_ID,
+    });
+    const ruleMonitor: Monitor = monitorRow({
+      autoProvisionedNetworkDeviceId: DEVICE_ID,
+      monitorTemplateId: OTHER_TEMPLATE_ID,
+    });
+
+    monitorTable = [policyMonitor, ruleMonitor];
+
+    const deleted: number =
+      await NetworkAlertPolicyEngineService.deletePolicyMonitorsForDevices({
+        projectId: PROJECT_ID,
+        deviceIds: [DEVICE_ID],
+      });
+
+    expect(deleted).toBe(1);
+    /*
+     * The auto-import rule's monitor survives this call: it is removed by
+     * NetworkDeviceService's own preflight, under the caller's permissions,
+     * because somebody chose that rule.
+     */
+    expect(monitorTable).toEqual([ruleMonitor]);
+  });
+
+  it("asks nothing when there are no devices", async () => {
+    expect(
+      await NetworkAlertPolicyEngineService.deletePolicyMonitorsForDevices({
+        projectId: PROJECT_ID,
+        deviceIds: [],
+      }),
+    ).toBe(0);
+    expect(monitorFindByMock).not.toHaveBeenCalled();
+  });
+});
+
+describe("pausing and resuming a policy's monitors", () => {
+  beforeEach(() => {
+    monitorTable = [
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+        networkAlertPolicyId: POLICY_ID,
+      }),
+      monitorRow({
+        autoProvisionedNetworkDeviceId: OTHER_DEVICE_ID,
+        monitorTemplateId: TEMPLATE_ID,
+        networkAlertPolicyId: POLICY_ID,
+      }),
+      // Another policy's monitor: never touched by this policy's pause.
+      monitorRow({
+        autoProvisionedNetworkDeviceId: DEVICE_ID,
+        monitorTemplateId: OTHER_TEMPLATE_ID,
+        networkAlertPolicyId: OTHER_POLICY_ID,
+      }),
+    ];
+  });
+
+  it("pauses only its own, and resuming reverses it exactly", async () => {
+    await NetworkAlertPolicyEngineService.setPolicyMonitorsPaused({
+      projectId: PROJECT_ID,
+      policyId: POLICY_ID,
+      isPaused: true,
+    });
+
+    expect(monitorTable[0]!.disableActiveMonitoring).toBe(true);
+    expect(monitorTable[1]!.disableActiveMonitoring).toBe(true);
+    expect(monitorTable[2]!.disableActiveMonitoring).toBe(false);
+
+    await NetworkAlertPolicyEngineService.setPolicyMonitorsPaused({
+      projectId: PROJECT_ID,
+      policyId: POLICY_ID,
+      isPaused: false,
+    });
+
+    expect(monitorTable[0]!.disableActiveMonitoring).toBe(false);
+    expect(monitorTable[1]!.disableActiveMonitoring).toBe(false);
+  });
+
+  /*
+   * The pause goes through updateBy, not a raw column write, so workflows,
+   * realtime and audit see it exactly as they would an operator's — and the
+   * walk fan-out, which reads this flag, stops feeding the monitor.
+   */
+  it("writes disableActiveMonitoring through the monitor service", async () => {
+    await NetworkAlertPolicyEngineService.setPolicyMonitorsPaused({
+      projectId: PROJECT_ID,
+      policyId: POLICY_ID,
+      isPaused: true,
+    });
+
+    expect(monitorUpdateByMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: { disableActiveMonitoring: true },
+        props: { isRoot: true },
+      }),
+    );
+  });
+
+  /*
+   * The page is re-read on the flag it is about to write, so a write that
+   * lands on nothing would be read again forever. It stops and says so
+   * instead of spinning the worker.
+   */
+  it("gives up rather than looping when the write lands on nothing", async () => {
+    monitorUpdateByMock.mockResolvedValue(0);
+
+    await expect(
+      NetworkAlertPolicyEngineService.setPolicyMonitorsPaused({
+        projectId: PROJECT_ID,
+        policyId: POLICY_ID,
+        isPaused: true,
+      }),
+    ).resolves.toBe(2);
+  });
+});
+
+describe("onMonitorTemplateSynced", () => {
+  it("stamps templateSyncedAt on the policy that provisions from the template", async () => {
+    policyTable = [policy()];
+
+    await NetworkAlertPolicyEngineService.onMonitorTemplateSynced({
+      monitorTemplateId: TEMPLATE_ID,
+      projectId: PROJECT_ID,
+    });
+
+    expect(policyStampMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        id: POLICY_ID,
+        data: { templateSyncedAt: expect.any(Date) },
+      }),
+    );
+  });
+
+  it("stamps nothing when no policy uses the template", async () => {
+    policyTable = [policy({ monitorTemplateId: OTHER_TEMPLATE_ID })];
+
+    await NetworkAlertPolicyEngineService.onMonitorTemplateSynced({
+      monitorTemplateId: TEMPLATE_ID,
+      projectId: PROJECT_ID,
+    });
+
+    expect(policyStampMock).not.toHaveBeenCalled();
+  });
+
+  /*
+   * Never throws. A bookkeeping column must not fail the template sync the
+   * operator actually asked for.
+   */
+  it("swallows a stamping failure", async () => {
+    policyTable = [policy()];
+    policyStampMock.mockRejectedValue(new Error("no"));
+
+    await expect(
+      NetworkAlertPolicyEngineService.onMonitorTemplateSynced({
+        monitorTemplateId: TEMPLATE_ID,
+        projectId: PROJECT_ID,
+      }),
+    ).resolves.toBeUndefined();
+  });
+});