feat(schedule): 任务支持「每天定时 / 每周定时」，用选择器而不是手写 cron

- schedule.py: 新增调度计算模块（纯函数，便于单测）。三种模式：interval（每 N 分钟）/ daily（选钟点）/ weekly（选星期 + 钟点）
- 钟点模式是「固定时刻」而非「固定延迟」——从日历重算，所以某轮跑晚了不会把之后每一轮都拖晚。interval 保持原语义：从上一轮开始计时
- 抖动只加给 interval。给「每天 9:00」也加抖动就成了 9:00–9:01 随机触发，操作者选的时间被悄悄改掉，只会像 bug
- 模型 / schema / service: 新增 schedule_mode / schedule_hours / schedule_days / schedule_minute。时钟字段存逗号分隔文本——几个小整数、永远整体读写，单开一张表只会换来 join。interval_minutes 保留且仍是默认值，已有任务不受影响
- service: 改动任何调度字段都按合并后的状态重算 next_run_at。重新启用也算改动，否则停用一个月再打开会带着一个月前的 next_run_at，一保存就立即触发
- 前端: 运行方式三选一 + 小时/星期胶囊多选 + 分钟下拉，并实时预览结果句子。任务卡片改显示后端拼好的 schedule_label，避免列表和编辑器对同一计划给出两种说法
- 校验: 钟点模式至少选一个时间，按周至少选一个星期

tests/test_schedule.py 新增 24 个用例，含「恰好等于当前时刻的档位归属下一天」这个会让调度器自循环的边界。

api/monitor/scheduler.py

@@ -23,8 +23,15 @@ is enough here: there is exactly one process, one global crawler subprocess, and
 therefore no concurrency to coordinate -- a cron-style library would add a
 dependency without adding a capability.
 
-Scheduling is **fixed-delay**, not fixed-rate: ``next_run_at`` is set from the
-moment a run starts, so a slow run cannot make its task fire back-to-back.
+Two families of schedule, and the difference matters:
+
+* ``interval`` is **fixed-delay**, not fixed-rate -- ``next_run_at`` is measured
+  from the moment a run starts, so a slow run cannot make its task fire
+  back-to-back.
+* the clock modes (``daily``/``weekly``) are **fixed-time** -- recomputed from the
+  calendar, so a run that starts late does not drag every later run with it.
+
+The arithmetic for both lives in schedule.py.
 """
 
 import asyncio
@@ -37,7 +44,7 @@ from sqlalchemy import select
 from tools.time_util import get_current_timestamp
 
 from ..services import crawler_manager
-from . import app_settings
+from . import app_settings, schedule
 from .db import get_session
 from .models import MonitorRun, MonitorTask, RUN_INTERRUPTED, RUN_RUNNING
 from .runner import execute_task
@@ -45,10 +52,9 @@ from .settings import get_cookie
 
 POLL_INTERVAL_SECONDS = 20
 # Spread tasks sharing an interval so they do not all come due on the same tick.
+# Applied to interval mode only -- see the advance step below.
 JITTER_SECONDS = 60
 
-_MS_PER_MINUTE = 60_000
-
 
 class MonitorScheduler:
     """Polls the task table and runs whatever is due."""
@@ -168,11 +174,34 @@ class MonitorScheduler:
             # Advance before running so a crash mid-run cannot cause an immediate
             # re-fire, and so a long outage coalesces into a single run instead
             # of one run per missed interval.
-            task.next_run_at = (
-                get_current_timestamp()
-                + task.interval_minutes * _MS_PER_MINUTE
-                + random.randint(0, JITTER_SECONDS) * 1000
+            now = get_current_timestamp()
+            following = schedule.next_occurrence(
+                mode=task.schedule_mode,
+                interval_minutes=task.interval_minutes,
+                hours=schedule.parse_hours(task.schedule_hours),
+                days=schedule.parse_days(task.schedule_days),
+                minute=task.schedule_minute,
+                after_ms=now,
             )
+
+            if following is None:
+                # A clock schedule with no times can never fire. The API rejects
+                # that shape, so this guards against a hand-edited row: park the
+                # task with no next run rather than leaving it permanently due and
+                # re-running it on every tick.
+                task.next_run_at = None
+                print(
+                    f"[monitor.scheduler] task {task.id} has no usable schedule "
+                    f"and will not run until one is set"
+                )
+            elif task.schedule_mode == schedule.MODE_INTERVAL:
+                # Jitter belongs to the interval mode only. Spreading identical
+                # intervals apart is the point; nudging a time the operator
+                # explicitly picked is not -- it just looks like a broken clock.
+                task.next_run_at = following + random.randint(0, JITTER_SECONDS) * 1000
+            else:
+                task.next_run_at = following
+
             task_id = task.id
 
         try: