feat(backend): Phase 1-4 全新開發完成，37/37 TDD 通過

[Phase 0 Reset]
- 清除舊版 app/、alembic/versions/、雜亂測試腳本
- 新 requirements.txt (移除 caldav/redis/keycloak-lib，加入 apscheduler/croniter/docker/paramiko/ping3/dnspython)

[Phase 1 資料庫]
- 9 張資料表 SQLAlchemy Models：tenants / accounts / schedules / schedule_logs /
  tenant_schedule_results / account_schedule_results / servers / server_status_logs / system_status_logs
- Alembic migration 001_create_all_tables (已套用到 10.1.0.20:5433/virtual_mis)
- seed.py：schedules 初始 3 筆 / servers 初始 4 筆

[Phase 2 CRUD API]
- GET/POST/PUT/DELETE: /api/v1/tenants / accounts / servers / schedules
- /api/v1/system-status
- 帳號編碼自動產生 (prefix + seq_no 4碼左補0)
- 燈號 (lights) 從最新排程結果取得

[Phase 3 Watchdog]
- APScheduler interval 3分鐘，原子 UPDATE status=Going 防重複執行
- 手動觸發 API: POST /api/v1/schedules/{id}/run

[Phase 4 Service Clients]
- KeycloakClient：vmis-admin realm，REST API (不用 python-keycloak)
- MailClient：Docker Mailserver @ 10.1.0.254:8080，含 MX DNS 驗證
- DockerClient：docker-py 本機 + paramiko SSH 遠端 compose
- NextcloudClient：OCS API user/quota
- SystemChecker：功能驗證 (traefik routers>0 / keycloak token / SMTP EHLO / DB SELECT 1 / ping)

[TDD]
- 37 tests / 37 passed (2.11s)
- SQLite in-memory + StaticPool，無需外部 DB

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/app/services/scheduler/watchdog.py

@@ -0,0 +1,107 @@
+"""
+Watchdog: APScheduler BackgroundScheduler，每 3 分鐘掃描 schedules 表。
+防重複執行：原子 UPDATE status='Going'，影響 0 筆則跳過。
+"""
+import logging
+from datetime import datetime
+from apscheduler.schedulers.background import BackgroundScheduler
+from croniter import croniter
+from sqlalchemy import update
+from sqlalchemy.orm import Session
+
+from app.core.database import SessionLocal
+from app.models.schedule import Schedule, ScheduleLog
+
+logger = logging.getLogger(__name__)
+
+_scheduler = BackgroundScheduler(timezone="Asia/Taipei")
+
+
+def _watchdog_tick():
+    db: Session = SessionLocal()
+    try:
+        due = (
+            db.query(Schedule)
+            .filter(
+                Schedule.status == "Waiting",
+                Schedule.next_run_at <= datetime.utcnow(),
+            )
+            .all()
+        )
+        for schedule in due:
+            # Atomic lock: only one process wins
+            affected = db.execute(
+                update(Schedule)
+                .where(Schedule.id == schedule.id, Schedule.status == "Waiting")
+                .values(status="Going")
+            ).rowcount
+            db.commit()
+
+            if affected == 0:
+                # Another process already grabbed it
+                continue
+
+            log = ScheduleLog(
+                schedule_id=schedule.id,
+                schedule_name=schedule.name,
+                started_at=datetime.utcnow(),
+                status="running",
+            )
+            db.add(log)
+            db.commit()
+            db.refresh(log)
+
+            try:
+                from app.services.scheduler.runner import dispatch_schedule
+                dispatch_schedule(schedule.id, log.id, db)
+                final_status = "ok"
+            except Exception as e:
+                logger.exception(f"Schedule {schedule.name} failed: {e}")
+                final_status = "error"
+
+            # Update log
+            log.ended_at = datetime.utcnow()
+            log.status = final_status
+
+            # Recalculate next_run_at
+            try:
+                cron = croniter(schedule.cron_timer, datetime.utcnow())
+                next_run = cron.get_next(datetime)
+            except Exception:
+                next_run = None
+
+            # Reset schedule
+            db.execute(
+                update(Schedule)
+                .where(Schedule.id == schedule.id)
+                .values(
+                    status="Waiting",
+                    last_run_at=datetime.utcnow(),
+                    next_run_at=next_run,
+                    last_status=final_status,
+                )
+            )
+            db.commit()
+
+    except Exception as e:
+        logger.exception(f"Watchdog tick error: {e}")
+        db.rollback()
+    finally:
+        db.close()
+
+
+def start_watchdog():
+    _scheduler.add_job(
+        _watchdog_tick,
+        trigger="interval",
+        minutes=3,
+        id="watchdog",
+        replace_existing=True,
+    )
+    _scheduler.start()
+    logger.info("Watchdog scheduler started")
+
+
+def stop_watchdog():
+    _scheduler.shutdown(wait=False)
+    logger.info("Watchdog scheduler stopped")