test(scheduler): nail H7 progress keep + H8 deferred task:state

Address Michał NEED-CHANGES on !7 tip f408da93: fail/defer omit-progress
preserves progress/stat_date; seed + quiet-window defer assert
status=deferred and future due_at on task:state. Soft: seeded due≠wall;
cleanup failure reads task:state.

DEVELOPMENT.md

@@ -456,6 +456,9 @@ P1 实现说明：
 
 - 时区使用固定 UTC+8（中国无夏令时），避免 Windows 缺少 `tzdata`；
 - 午夜 `seed_at`（默认 `00:00`）把 `snapshot:last_completed` 的 ok=1 行克隆为当日壳，写入 `snapshot:seeded:<date>=1` / `fans:seeded:<date>=1`，**不**推进完成日；
+- 也可手动：`python app.py snapshot-seed [--date YYYY-MM-DD]`（Debian / `docker compose exec re-xxm python app.py snapshot-seed`）；
+- 目标日已有 `snapshot:completed:<date>` 时 seed 记 `already_completed` 成功（即使 `last_completed` 已等于当日），**不得**误判 `no_source`；
+- 若全站 `snapshot` 正在跑，或目标日未完成且 `last_completed` 暂不可用（`no_source`），seed **不记成功**，约 10 分钟后重试直到真正克隆成功；
 - 晚间 `snapshot_at`（默认 `23:15`）对 seeded 日 force 重采；成功后写 `snapshot:completed:<date>=1`、推进 `last_completed` 并清除 seeded；
 - 当日全部视频与粉丝采集成功后在 `kv` 写入 `snapshot:completed:<date>=1`；启动时未完成会补跑（纯 seed 日不误标完成）；
 - 完成标记不是唯一门槛：每次检查 `pending_snapshot_aids(date)`，新入库或从 `dropped` 恢复的视频会立即补采当天快照；
@@ -466,6 +469,15 @@ P1 实现说明：
 - Cookie 登录态通过 `x/web-interface/nav` 检查，失效只告警，不阻塞匿名 `view`；
 - Telegram 摘要推迟到 P3，P1 先写结构化 JSON 日志和 `kv` 状态。
 
+调度避让与补跑：
+
+- 常驻启动时日更任务（`snapshot_seed` / `snapshot` / `fans`）按墙钟下一档 `seed_at` / `snapshot_at` 排期（可参考上次成功/失败记录），**不得** `next_run=0` 重启即跑；
+- 仅到达当日 `snapshot_at` 后（或 CLI `--force`）才对 seed 壳 `force` 打穿；白天重启不因 `snapshot:seeded` 全量强采；
+- 全站 `snapshot` **不因**其它任务 running 而 skip；到期/强制必须执行；
+- 上海时区 **23:00–01:00** 为全站快照窗：`discovery` / `snapshot_recent` / `candidate_cleanup` 不执行，延后到 01:00 后；`fans` 短耗时，不避开任何任务；
+- 任务 skip（含静默窗延后）写入 `ok=false` + `skipped=true`，**不得**记成成功或推进「当日已完成」；
+- skip / fail 后短间隔补跑（约 5 分钟；`snapshot_seed` 避让/无源为约 10 分钟）；日更成功才推到次日 `snapshot_at` / `seed_at`。
+
 ---
 
 ## 8. Telegram 与入库操作

project.md

@@ -800,7 +800,7 @@ secrets.example.json            凭据样例（auth 回写目标）
 **进程结构**
 
 ```
-scheduler 线程：discovery(每 60min) / snapshot_seed(每日 00:00) / snapshot+fans(每日 23:15) / snapshot_recent(小时) / candidate_cleanup(每日)
+scheduler 线程：discovery(每 discover_minutes，避开 23:00–01:00) / snapshot_seed(每日 00:00；snapshot 占用或 no_source 则 10 分钟重试，CLI: snapshot-seed) / snapshot(每日 23:15，不因其它任务 skip) / fans(每日 23:15，不避开) / snapshot_recent(小时，避开 23:00–01:00) / candidate_cleanup(每日，避开 23:00–01:00)
 web 线程：ThreadingHTTPServer :8080
 telegram：getUpdates long polling（回调 + 私聊文本入库）
 storage：SQLite WAL，单写线程

reaction_board/scheduler.py

@@ -3,6 +3,9 @@
 
 P0 只注册收藏夹同步任务；P1/P3 会在同一调度器上追加快照、粉丝和发现任务。
 调度只负责“什么时候运行”，具体任务保持幂等，并在自己的服务层处理重试。
+
+任务 ``fn`` 可返回 :class:`TaskControl` 以覆盖 ``ok`` / ``next_run``
+（用于静默窗延后、失败补跑等）；未返回时仍按 ``next_run_fn`` / ``interval`` 排期。
 """
 
 from __future__ import annotations
@@ -12,6 +15,15 @@ from dataclasses import dataclass
 from typing import Any, Callable, Sequence
 
 
+@dataclass
+class TaskControl:
+    """任务可选返回值：覆盖成功标记与下次运行时间。"""
+
+    ok: bool = True
+    next_run: float | None = None
+    error: str | None = None
+
+
 @dataclass
 class ScheduledTask:
     name: str
@@ -70,23 +82,48 @@ class Scheduler:
                 continue
             started = current
             error: str | None = None
+            control: TaskControl | None = None
             try:
-                task.fn()
+                raw = task.fn()
             except Exception as exc:  # noqa: BLE001  单任务失败不终止调度
                 error = str(exc)
-                self._log("error", "scheduler.task_error", task=task.name, error=error)
             else:
-                self._log("info", "scheduler.task_ok", task=task.name)
+                if isinstance(raw, TaskControl):
+                    control = raw
             finished = self._clock()
-            if task.next_run_fn is not None:
+            if control is not None and control.next_run is not None:
+                task.next_run = float(control.next_run)
+            elif task.next_run_fn is not None:
                 task.next_run = task.next_run_fn(started)
             else:
                 task.next_run = started + task.interval
+            ok = error is None
+            if control is not None:
+                ok = bool(control.ok) and error is None
+                if control.error and error is None:
+                    error = control.error
+            # Skip/fail ≠ success：先定 ok/error，再记 scheduler 日志。
+            if ok:
+                self._log("info", "scheduler.task_ok", task=task.name)
+            elif control is not None:
+                self._log(
+                    "warning",
+                    "scheduler.task_fail",
+                    task=task.name,
+                    error=error,
+                )
+            else:
+                self._log(
+                    "error",
+                    "scheduler.task_error",
+                    task=task.name,
+                    error=error,
+                )
             result = TaskResult(
                 name=task.name,
                 started_at=started,
                 finished_at=finished,
-                ok=error is None,
+                ok=ok,
                 error=error,
             )
             results.append(result)

reaction_board/services/snapshot.py

@@ -51,6 +51,7 @@ class SnapshotSummary:
     completed: bool = False
     already_completed: bool = False
     finished_at: int = 0
+    reason: str | None = None
     errors: list[str] = field(default_factory=list)
 
     def as_dict(self) -> dict[str, Any]:
@@ -73,6 +74,7 @@ class SnapshotSummary:
             "backup_error": self.backup_error,
             "completed": self.completed,
             "already_completed": self.already_completed,
+            "reason": self.reason,
             "errors": self.errors,
         }
 
@@ -153,7 +155,10 @@ class SnapshotService:
             raise ValueError(f"日期格式必须是 YYYY-MM-DD: {stat_date!r}")
         seeded = self.repo.kv_get(self.seeded_key(stat_date)) == "1"
         if force:
-            return self.run_videos(date=stat_date, force=True)
+            # 晚间 seal / 续跑：打穿 seed，保留已真实采集的 ok=1 作为进度。
+            return self.run_videos(
+                date=stat_date, force=True, keep_scraped=True,
+            )
 
         completed = self.repo.kv_get(self.completed_key(stat_date)) == "1"
         if seeded and not completed:
@@ -167,6 +172,7 @@ class SnapshotService:
                 finished_at=now,
                 completed=False,
                 already_completed=False,
+                reason="seeded_pending_seal",
             )
         pending = self.repo.pending_snapshot_aids(stat_date)
         if completed and not pending:
@@ -203,8 +209,19 @@ class SnapshotService:
         summary = self.run_videos(date=date, force=force)
         return self.run_fans(date=summary.stat_date, force=force, summary=summary)
 
-    def run_videos(self, *, date: str | None = None, force: bool = False) -> SnapshotSummary:
-        """采集视频七项指标并在全部成功后标记完成日 / 备份。"""
+    def run_videos(
+        self,
+        *,
+        date: str | None = None,
+        force: bool = False,
+        keep_scraped: bool = False,
+    ) -> SnapshotSummary:
+        """采集视频七项指标并在全部成功后标记完成日 / 备份。
+
+        ``force``：忽略完成标记，重跑。
+        ``keep_scraped``：force 时仍跳过非 seed 的 ok=1（seal 断点续跑）；
+        CLI ``--force`` 不传此标志，会重采全部成功行。
+        """
         cfg = self.config_store.get()
         client = self._client()
         now = self._now()
@@ -225,11 +242,19 @@ class SnapshotService:
         summary.total = len(videos)
         for index, aid in enumerate(videos, 1):
             existing = self.repo.get_snapshot(stat_date, aid)
-            if existing is not None and int(existing.get("ok") or 0) == 1 and not force:
-                # 断点续跑：同日已成功则跳过。小时任务刷新过的行也是 ok=1，
-                # 自然覆盖「1 小时内 captured_at 已更新则不再次更新」的要求。
-                summary.skipped += 1
-                continue
+            if existing is not None and int(existing.get("ok") or 0) == 1:
+                src = str(existing.get("src") or "")
+                if keep_scraped and src == "seed":
+                    # seal 续跑：必须打穿克隆壳
+                    pass
+                elif keep_scraped and src != "seed":
+                    summary.skipped += 1
+                    continue
+                elif not force:
+                    # 普通续跑：同日已成功则跳过（含小时任务刚刷新的）
+                    summary.skipped += 1
+                    continue
+                # force 且非 keep_scraped：CLI 全量重采，不 skip
 
             stat, error, unavailable = self._fetch_view(client, aid)
             if unavailable:
@@ -423,6 +448,7 @@ class SnapshotService:
                     completed=False,
                     already_completed=False,
                     fans_skipped=len(cfg.track_fans),
+                    reason="seeded_pending_seal",
                 )
             return self.run_fans(date=stat_date, force=True)
         if not self._pending_fan_mids(stat_date, cfg.track_fans):
@@ -536,15 +562,8 @@ class SnapshotService:
             raise ValueError(f"日期格式必须是 YYYY-MM-DD: {stat_date!r}")
 
         summary = SnapshotSummary(stat_date=stat_date, started_at=now)
-        source = self.repo.kv_get("snapshot:last_completed")
-        if not source or source >= stat_date:
-            self._log(
-                "info", "snapshot.seed_skip",
-                stat_date=stat_date, reason="no_source", source=source,
-            )
-            summary.finished_at = self._now()
-            summary.completed = True
-            return summary
+        # 目标日已完成：先于 no_source。全站快照当天完成后
+        # last_completed == stat_date，不可误判为无源再 10 分钟重试。
         if self.repo.kv_get(self.completed_key(stat_date)) == "1":
             self._log(
                 "info", "snapshot.seed_skip",
@@ -553,6 +572,17 @@ class SnapshotService:
             summary.finished_at = self._now()
             summary.completed = True
             summary.already_completed = True
+            summary.reason = "already_completed"
+            return summary
+        source = self.repo.kv_get("snapshot:last_completed")
+        if not source or source >= stat_date:
+            self._log(
+                "info", "snapshot.seed_skip",
+                stat_date=stat_date, reason="no_source", source=source,
+            )
+            summary.finished_at = self._now()
+            summary.completed = False
+            summary.reason = "no_source"
             return summary
         if self.repo.kv_get(self.seeded_key(stat_date)) == "1":
             self._log(
@@ -562,6 +592,7 @@ class SnapshotService:
             summary.finished_at = self._now()
             summary.completed = True
             summary.already_completed = True
+            summary.reason = "already_seeded"
             return summary
 
         copied = self.repo.copy_ok_snapshots(
@@ -583,6 +614,7 @@ class SnapshotService:
                 self.repo.bump_cache_version("fans")
         summary.finished_at = self._now()
         summary.completed = True  # seed 任务本身成功；不等于快照落盘完成
+        summary.reason = "seeded"
         self._log(
             "info", "snapshot.seed_done",
             stat_date=stat_date, source=source,

reaction_board/task_state.py

@@ -0,0 +1,291 @@
+# -*- coding: utf-8 -*-
+"""调度任务状态持久化（跨进程重启 / 容器重建）。
+
+每条任务一条 kv 记录 ``task:state:{name}``，字段见 :class:`TaskState`。
+调度器只认落盘的 ``due_at`` / ``status`` / ``stat_date`` / ``progress``，
+不再靠启动时把 ``next_run=0`` 猜「要不要立刻跑」。
+"""
+
+from __future__ import annotations
+
+import json
+import time
+from dataclasses import asdict, dataclass, field
+from typing import Any, Mapping
+
+from .repository import Repository
+
+STATUS_PENDING = "pending"
+STATUS_RUNNING = "running"
+STATUS_DONE = "done"
+STATUS_FAILED = "failed"
+STATUS_DEFERRED = "deferred"
+
+_ACTIVE = frozenset({STATUS_PENDING, STATUS_RUNNING, STATUS_FAILED, STATUS_DEFERRED})
+
+
+@dataclass
+class TaskState:
+    name: str
+    status: str
+    due_at: float
+    stat_date: str | None = None
+    progress: dict[str, Any] | None = None
+    updated_at: float = 0.0
+    error: str | None = None
+
+    def as_dict(self) -> dict[str, Any]:
+        payload = asdict(self)
+        # JSON 友好：去掉 None progress 也可保留
+        return payload
+
+
+def state_key(name: str) -> str:
+    return f"task:state:{name}"
+
+
+def load_state(repo: Repository, name: str) -> TaskState | None:
+    raw = repo.kv_get(state_key(name))
+    if not raw:
+        return None
+    try:
+        data = json.loads(raw)
+    except (TypeError, ValueError, json.JSONDecodeError):
+        return None
+    if not isinstance(data, dict) or not data.get("name"):
+        return None
+    try:
+        due_at = float(data.get("due_at") or 0)
+        updated_at = float(data.get("updated_at") or 0)
+    except (TypeError, ValueError):
+        return None
+    progress = data.get("progress")
+    if progress is not None and not isinstance(progress, dict):
+        progress = None
+    return TaskState(
+        name=str(data.get("name") or name),
+        status=str(data.get("status") or STATUS_PENDING),
+        due_at=due_at,
+        stat_date=(str(data["stat_date"]) if data.get("stat_date") else None),
+        progress=progress,
+        updated_at=updated_at,
+        error=(str(data["error"]) if data.get("error") else None),
+    )
+
+
+def save_state(repo: Repository, state: TaskState) -> None:
+    repo.kv_set(state_key(state.name), json.dumps(state.as_dict(), ensure_ascii=False))
+
+
+def _now(now: float | None) -> float:
+    return time.time() if now is None else float(now)
+
+
+def recover_stale_running(
+    repo: Repository,
+    name: str,
+    *,
+    now: float | None = None,
+) -> TaskState | None:
+    """进程已死但仍为 ``running``：改为 ``pending``，保留 progress / stat_date。
+
+    ``due_at`` 取 ``min(原 due_at, now)``，使过期/中断任务可立即续跑，
+    未到期的（极少见）仍尊重原 due_at。
+    """
+    current = _now(now)
+    state = load_state(repo, name)
+    if state is None or state.status != STATUS_RUNNING:
+        return state
+    state.status = STATUS_PENDING
+    state.due_at = min(float(state.due_at or current), current)
+    state.updated_at = current
+    save_state(repo, state)
+    return state
+
+
+def mark_running(
+    repo: Repository,
+    name: str,
+    *,
+    stat_date: str | None,
+    now: float | None = None,
+    progress: Mapping[str, Any] | None = None,
+    due_at: float | None = None,
+) -> TaskState:
+    current = _now(now)
+    prev = load_state(repo, name)
+    state = TaskState(
+        name=name,
+        status=STATUS_RUNNING,
+        due_at=float(due_at if due_at is not None else (prev.due_at if prev else current)),
+        stat_date=stat_date if stat_date is not None else (prev.stat_date if prev else None),
+        progress=dict(progress) if progress is not None else (prev.progress if prev else None),
+        updated_at=current,
+        error=None,
+    )
+    save_state(repo, state)
+    return state
+
+
+def mark_done(
+    repo: Repository,
+    name: str,
+    *,
+    due_at: float,
+    now: float | None = None,
+    stat_date: str | None = None,
+    progress: Mapping[str, Any] | None = None,
+) -> TaskState:
+    """成功：记 done，并写入下一次 ``due_at``（下一条墙钟/间隔）。"""
+    current = _now(now)
+    state = TaskState(
+        name=name,
+        status=STATUS_DONE,
+        due_at=float(due_at),
+        stat_date=stat_date,
+        progress=dict(progress) if progress is not None else None,
+        updated_at=current,
+        error=None,
+    )
+    save_state(repo, state)
+    return state
+
+
+def mark_failed(
+    repo: Repository,
+    name: str,
+    *,
+    due_at: float,
+    now: float | None = None,
+    stat_date: str | None = None,
+    progress: Mapping[str, Any] | None = None,
+    error: str | None = None,
+) -> TaskState:
+    current = _now(now)
+    prev = load_state(repo, name)
+    state = TaskState(
+        name=name,
+        status=STATUS_FAILED,
+        due_at=float(due_at),
+        stat_date=stat_date if stat_date is not None else (prev.stat_date if prev else None),
+        progress=(
+            dict(progress) if progress is not None
+            else (prev.progress if prev else None)
+        ),
+        updated_at=current,
+        error=error,
+    )
+    save_state(repo, state)
+    return state
+
+
+def mark_deferred(
+    repo: Repository,
+    name: str,
+    *,
+    due_at: float,
+    now: float | None = None,
+    stat_date: str | None = None,
+    progress: Mapping[str, Any] | None = None,
+    error: str | None = None,
+) -> TaskState:
+    """主动延后（静默窗 / seed 避让等）：保留进度，写未来 due_at。"""
+    current = _now(now)
+    prev = load_state(repo, name)
+    state = TaskState(
+        name=name,
+        status=STATUS_DEFERRED,
+        due_at=float(due_at),
+        stat_date=stat_date if stat_date is not None else (prev.stat_date if prev else None),
+        progress=(
+            dict(progress) if progress is not None
+            else (prev.progress if prev else None)
+        ),
+        updated_at=current,
+        error=error,
+    )
+    save_state(repo, state)
+    return state
+
+
+def ensure_pending(
+    repo: Repository,
+    name: str,
+    *,
+    due_at: float,
+    stat_date: str | None,
+    now: float | None = None,
+    overwrite_done: bool = True,
+) -> TaskState:
+    """保证存在一条 pending（如 seed 成功后挂上当晚 seal）。
+
+    若已有未完成状态且 due_at 更早，保留更早的 due_at；
+    ``overwrite_done`` 时覆盖已完成记录。
+    """
+    current = _now(now)
+    prev = load_state(repo, name)
+    if prev is not None and prev.status in _ACTIVE and prev.status != STATUS_RUNNING:
+        # 未完成：取更早的 due_at，补上缺失的 stat_date
+        due = min(float(prev.due_at or due_at), float(due_at))
+        sd = prev.stat_date or stat_date
+        if due == prev.due_at and sd == prev.stat_date and prev.status == STATUS_PENDING:
+            return prev
+        prev.due_at = due
+        prev.stat_date = sd
+        prev.status = STATUS_PENDING
+        prev.updated_at = current
+        save_state(repo, prev)
+        return prev
+    if prev is not None and prev.status == STATUS_RUNNING:
+        return prev
+    if prev is not None and prev.status == STATUS_DONE and not overwrite_done:
+        return prev
+    state = TaskState(
+        name=name,
+        status=STATUS_PENDING,
+        due_at=float(due_at),
+        stat_date=stat_date,
+        progress=None,
+        updated_at=current,
+        error=None,
+    )
+    save_state(repo, state)
+    return state
+
+
+def startup_next_run(
+    repo: Repository,
+    name: str,
+    *,
+    wall_due_at: float,
+    now: float | None = None,
+) -> float:
+    """常驻启动时的 ``ScheduledTask.next_run``：只认落盘状态。
+
+    - 恢复残留 ``running`` → ``pending``
+    - ``due_at > now`` → 原样返回（失败延时不会被提前）
+    - ``due_at <= now`` 且未完成 → 立刻可跑（返回 ``due_at``，可为过去时间）
+    - ``done`` 且 ``due_at > now`` → 等下一次
+    - ``done`` 且 ``due_at <= now`` → 到期，返回 ``due_at``
+    - 无状态 → 墙钟 ``wall_due_at``
+    """
+    current = _now(now)
+    state = recover_stale_running(repo, name, now=current)
+    if state is None:
+        return float(wall_due_at)
+    due = float(state.due_at or 0)
+    if due <= 0:
+        return float(wall_due_at)
+    # 无论 status：due_at 是唯一时间门闩
+    if due > current:
+        return due
+    # 已到期：未完成要续跑；done 也表示下一轮已到期
+    return due
+
+
+def resume_stat_date(repo: Repository, name: str, *, fallback: str) -> str:
+    """续跑时优先用状态里的归属日，避免跨日重启跑错天。"""
+    state = load_state(repo, name)
+    if state is not None and state.stat_date and state.status in _ACTIVE:
+        return state.stat_date
+    return fallback

reaction_board/timeutil.py

@@ -57,3 +57,39 @@ def day_bounds(date: str) -> tuple[float, float]:
         raise ValueError(f"日期格式必须是 YYYY-MM-DD: {date!r}")
     start = datetime.strptime(date, "%Y-%m-%d").replace(tzinfo=SHANGHAI_TZ)
     return start.timestamp(), (start + timedelta(days=1)).timestamp()
+
+
+# 全站快照静默窗：上海时区 [23:00, 01:00)。此间 discovery / recent / cleanup 不执行。
+SNAPSHOT_QUIET_START_HOUR = 23
+SNAPSHOT_QUIET_END_HOUR = 1
+
+
+def in_snapshot_quiet_window(ts: float | None = None) -> bool:
+    """是否处于全站快照静默窗 [23:00, 01:00)（上海时区，跨日）。"""
+    dt = shanghai_datetime(ts)
+    return dt.hour >= SNAPSHOT_QUIET_START_HOUR or dt.hour < SNAPSHOT_QUIET_END_HOUR
+
+
+def next_after_snapshot_quiet_window(ts: float) -> float:
+    """若 ``ts`` 落在静默窗内，推到窗结束后的 01:00；否则原样返回。"""
+    if not in_snapshot_quiet_window(ts):
+        return ts
+    dt = shanghai_datetime(ts)
+    if dt.hour >= SNAPSHOT_QUIET_START_HOUR:
+        target = (dt + timedelta(days=1)).replace(
+            hour=SNAPSHOT_QUIET_END_HOUR, minute=0, second=0, microsecond=0,
+        )
+    else:
+        target = dt.replace(
+            hour=SNAPSHOT_QUIET_END_HOUR, minute=0, second=0, microsecond=0,
+        )
+    if target.timestamp() <= ts:
+        target += timedelta(days=1)
+    return target.timestamp()
+
+
+def interval_next_run(now: float, interval: float) -> float:
+    """``now + interval``，若落入静默窗则推到 01:00 之后。"""
+    if interval <= 0:
+        raise ValueError("interval 必须大于 0")
+    return next_after_snapshot_quiet_window(now + interval)

reaction_board/web/queries.py

@@ -34,7 +34,14 @@ from ..analytics import (
 )
 from ..config import ConfigStore
 from ..database import Database
-from ..timeutil import day_bounds, next_daily_run, shanghai_date, shanghai_datetime
+from ..timeutil import (
+    day_bounds,
+    interval_next_run,
+    next_after_snapshot_quiet_window,
+    next_daily_run,
+    shanghai_date,
+    shanghai_datetime,
+)
 from .cache import DASHBOARD_CACHE, TTL_FANS, TTL_VIDEOS, build_cache_key
 
 TRACKED_STATUSES = ("active", "unfaved")
@@ -1971,9 +1978,27 @@ class DashboardQueries:
                 "fans": int(next_daily_run(now, cfg.schedule.snapshot_at)),
                 "snapshot": int(next_daily_run(now, cfg.schedule.snapshot_at)),
                 "snapshot_seed": int(next_daily_run(now, cfg.schedule.seed_at)),
-                "discover": int(now + discover_interval),
-                "snapshot_recent": int(now + recent_interval),
+                "discover": int(interval_next_run(now, float(discover_interval))),
+                "snapshot_recent": int(interval_next_run(now, float(recent_interval))),
             }
+            retry_seconds = 300
+            seed_retry_seconds = 600
+            # 优先 task:state.due_at（跨重启唯一时间门闩）；无则回退 task:last 推导。
+            state_rows = conn.execute(
+                "SELECT k, v FROM kv WHERE k LIKE 'task:state:%'"
+            ).fetchall()
+            for row in state_rows:
+                key = str(row["k"]).removeprefix("task:state:")
+                ui_name = "discover" if key == "discovery" else key
+                if ui_name not in next_runs:
+                    continue
+                try:
+                    state = json.loads(str(row["v"]))
+                    due = float(state.get("due_at") or 0)
+                except (TypeError, ValueError, json.JSONDecodeError):
+                    continue
+                if due > 0:
+                    next_runs[ui_name] = int(due)
             for name, key, at in (
                 ("fans", "fans", cfg.schedule.snapshot_at),
                 ("snapshot", "snapshot", cfg.schedule.snapshot_at),
@@ -1981,23 +2006,58 @@ class DashboardQueries:
                 ("discover", "discovery", None),
                 ("snapshot_recent", "snapshot_recent", None),
             ):
+                # 已有落盘 due_at 则不再用 last 覆盖
+                state_raw = self._kv(conn, f"task:state:{key}")
+                if state_raw:
+                    try:
+                        if float(json.loads(state_raw).get("due_at") or 0) > 0:
+                            continue
+                    except (TypeError, ValueError, json.JSONDecodeError):
+                        pass
                 record = tasks.get(key)
                 if not (isinstance(record, dict) and record.get("finished_at")):
                     continue
                 finished_at = int(record["finished_at"])
+                summary = record.get("summary") if isinstance(record.get("summary"), dict) else {}
+                skipped = bool(record.get("skipped") or summary.get("skipped"))
+                failed = record.get("ok") is False
                 if name in {"discover", "snapshot_recent"}:
-                    # 只按上次完成推下一档；进行中由 running 展示，不按墙钟抢跑。
+                    # 上次完成推下一档；skip/失败短补跑；静默窗推到 01:00 后。
                     interval = (
                         recent_interval if name == "snapshot_recent"
                         else discover_interval
                     )
-                    next_runs[name] = finished_at + interval
+                    if skipped:
+                        next_runs[name] = int(
+                            next_after_snapshot_quiet_window(float(finished_at))
+                        )
+                    elif failed:
+                        next_runs[name] = int(finished_at + retry_seconds)
+                    else:
+                        next_runs[name] = int(
+                            interval_next_run(float(finished_at), float(interval))
+                        )
                 else:
-                    next_runs[name] = int(next_daily_run(finished_at, at))
+                    if skipped or failed:
+                        # 日更任务失败/skip：短间隔补跑，绝不按成功推到次日。
+                        # snapshot_seed：全站快照占用 / no_source 时 10 分钟补跑。
+                        gap = (
+                            seed_retry_seconds if name == "snapshot_seed"
+                            else retry_seconds
+                        )
+                        next_runs[name] = int(finished_at + gap)
+                    else:
+                        next_runs[name] = int(next_daily_run(finished_at, at))
             errors = [
                 {"task": name, **record}
                 for name, record in tasks.items()
-                if isinstance(record, dict) and record.get("ok") is False
+                if isinstance(record, dict)
+                and record.get("ok") is False
+                and not record.get("skipped")
+                and not (
+                    isinstance(record.get("summary"), dict)
+                    and record["summary"].get("skipped")
+                )
             ]
             counts = {str(row["status"]): int(row["n"]) for row in conn.execute(
                 "SELECT status, COUNT(*) AS n FROM video GROUP BY status"

reaction_board/web/static/views/system.js

@@ -5,6 +5,15 @@ import {
   escapeHtml, formatCount, formatDateTime, formatNumber,
 } from "../format.js";
 
+function taskResultLabel(record) {
+  const summary = record.summary && typeof record.summary === "object" ? record.summary : {};
+  if (record.skipped || summary.skipped) {
+    return `<span class="muted">跳过</span>`;
+  }
+  if (record.ok) return "成功";
+  return `<span class="danger">失败</span>`;
+}
+
 function taskRows(tasks) {
   const entries = Object.entries(tasks || {});
   if (!entries.length) return emptyState("暂无任务记录", "运行一次同步、快照或发现后会写入最近结果。");
@@ -15,12 +24,16 @@ function taskRows(tasks) {
         ${entries.map(([name, record]) => {
           const summary = record.summary && typeof record.summary === "object"
             ? Object.entries(record.summary)
-              .filter(([key]) => ["added", "removed", "candidates", "notified", "ok", "failed", "total", "completed", "backup_error"].includes(key))
+              .filter(([key]) => [
+                "added", "removed", "candidates", "notified", "ok", "failed",
+                "total", "completed", "backup_error", "skipped", "already_completed",
+                "reason", "elapsed", "success",
+              ].includes(key))
               .map(([key, value]) => `${key}=${value}`).join(" ")
             : "";
           return `<tr>
             <td>${escapeHtml(name)}</td>
-            <td>${record.ok ? "成功" : `<span class="danger">失败</span>`}</td>
+            <td>${taskResultLabel(record)}</td>
             <td>${escapeHtml(formatDateTime(record.finished_at))}</td>
             <td class="summary-cell">${escapeHtml(summary || record.error || "—")}</td>
           </tr>`;
@@ -100,7 +113,7 @@ export async function renderSystem(root) {
         <div><span>新视频小时同步</span><strong>${running.snapshot_recent ? "运行中" : escapeHtml(formatDateTime(next.snapshot_recent))}</strong></div>
         <div><span>新视频搜索</span><strong>${running.discover || running.discovery ? "运行中" : escapeHtml(formatDateTime(next.discover))}</strong></div>
       </div>
-      <p class="section-note">进行中显示「运行中」；空闲时按该任务上次完成时间推算下次运行，不按墙钟抢跑。新视频小时同步与全站快照互斥。</p>
+      <p class="section-note">进行中显示「运行中」；空闲时按该任务上次完成时间推算下次运行。23:00–01:00（上海）为全站快照窗：发现 / 小时同步 / 候选清理会延后到 01:00 后；粉丝更新不避开。跳过或失败不会记成成功，会短间隔补跑。</p>
     </section>
 
     <section class="panel">

tests/test_scheduler.py

@@ -8,7 +8,7 @@ import threading
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 sys.stdout.reconfigure(encoding="utf-8")
 
-from reaction_board.scheduler import ScheduledTask, Scheduler  # noqa: E402
+from reaction_board.scheduler import ScheduledTask, Scheduler, TaskControl  # noqa: E402
 
 
 def test_run_when_due_and_reschedule():
@@ -128,9 +128,111 @@ def test_next_run_fn_controls_reschedule():
     assert calls == [1_000.0, 1_123.0]
 
 
+
+def test_task_control_overrides_ok_and_next_run():
+    clock = [1_000.0]
+    task = ScheduledTask(
+        "defer",
+        60,
+        lambda: TaskControl(ok=False, next_run=1_500.0, error="quiet"),
+    )
+    scheduler = Scheduler([task], clock=lambda: clock[0])
+    results = scheduler.run_due()
+    assert len(results) == 1
+    assert results[0].ok is False
+    assert results[0].error == "quiet"
+    assert task.next_run == 1_500.0
+    # 未到 next_run 不重跑
+    clock[0] = 1_499.0
+    assert scheduler.run_due() == []
+    clock[0] = 1_500.0
+    task.fn = lambda: TaskControl(ok=True, next_run=2_000.0)
+    results = scheduler.run_due()
+    assert results[0].ok is True
+    assert task.next_run == 2_000.0
+
+
+
+
+class _CaptureLogger:
+    def __init__(self):
+        self.events = []
+
+    def info(self, event, **fields):
+        self.events.append(("info", event, fields))
+
+    def warning(self, event, **fields):
+        self.events.append(("warning", event, fields))
+
+    def error(self, event, **fields):
+        self.events.append(("error", event, fields))
+
+
+def test_task_control_ok_false_logs_fail_not_ok():
+    """TaskControl(ok=False) 不得记 scheduler.task_ok（Skip ≠ success）。"""
+    clock = [1_000.0]
+    logger = _CaptureLogger()
+    task = ScheduledTask(
+        "quiet",
+        60,
+        lambda: TaskControl(ok=False, next_run=1_500.0, error="snapshot_quiet_window"),
+    )
+    scheduler = Scheduler([task], clock=lambda: clock[0], logger=logger)
+    results = scheduler.run_due()
+    assert results[0].ok is False
+    assert results[0].error == "snapshot_quiet_window"
+    assert ("info", "scheduler.task_ok", {"task": "quiet"}) not in logger.events
+    fail_events = [e for e in logger.events if e[1] == "scheduler.task_fail"]
+    assert len(fail_events) == 1
+    assert fail_events[0][0] == "warning"
+    assert fail_events[0][2]["task"] == "quiet"
+    assert fail_events[0][2]["error"] == "snapshot_quiet_window"
+
+
+def test_task_control_ok_true_still_logs_task_ok():
+    clock = [1_000.0]
+    logger = _CaptureLogger()
+    task = ScheduledTask(
+        "ok_task",
+        60,
+        lambda: TaskControl(ok=True, next_run=1_200.0),
+    )
+    scheduler = Scheduler([task], clock=lambda: clock[0], logger=logger)
+    results = scheduler.run_due()
+    assert results[0].ok is True
+    assert ("info", "scheduler.task_ok", {"task": "ok_task"}) in logger.events
+    assert not any(e[1] == "scheduler.task_fail" for e in logger.events)
+
+
+def test_raised_exception_logs_task_fail():
+    clock = [100.0]
+    logger = _CaptureLogger()
+
+    def boom():
+        raise RuntimeError("db down")
+
+    scheduler = Scheduler(
+        [ScheduledTask("bad", 10, boom)],
+        clock=lambda: clock[0],
+        logger=logger,
+    )
+    results = scheduler.run_due()
+    assert results[0].ok is False
+    assert results[0].error == "db down"
+    assert not any(e[1] == "scheduler.task_ok" for e in logger.events)
+    assert not any(e[1] == "scheduler.task_fail" for e in logger.events)
+    err_events = [e for e in logger.events if e[1] == "scheduler.task_error"]
+    assert len(err_events) == 1
+    assert err_events[0][0] == "error"
+    assert err_events[0][2]["error"] == "db down"
+
 def main():
     test_run_when_due_and_reschedule()
     test_next_run_fn_controls_reschedule()
+    test_task_control_overrides_ok_and_next_run()
+    test_task_control_ok_false_logs_fail_not_ok()
+    test_task_control_ok_true_still_logs_task_ok()
+    test_raised_exception_logs_task_fail()
     test_task_error_does_not_block_other_tasks()
     test_run_forever_with_stop_event()
     test_max_iterations_and_validation()

tests/test_snapshot.py

@@ -537,6 +537,36 @@ def test_full_snapshot_skips_ok_rows_including_hourly_refresh():
 
 
 
+def test_seed_same_day_completed_is_already_completed_not_no_source():
+    """全站完成后 last_completed==target：应 already_completed，不误判 no_source。"""
+    clock = 1_800_000_000
+    with tempfile.TemporaryDirectory() as tmp:
+        transport = SnapshotTransport({1: view_for(1)})
+        db, repo, service = make_service(tmp, transport, track_fans=())
+        try:
+            add_video(repo, 1)
+            today = "2026-10-31"
+            repo.upsert_snapshot(
+                today, 1,
+                metrics={"view": 10, "like": 1, "coin": 0, "favorite": 0,
+                         "share": 0, "reply": 0, "danmaku": 0},
+                ok=True, err=None, captured_at=clock,
+            )
+            # 正常收尾态：last_completed 已推进到当日，且 completed 标记在位
+            repo.kv_set("snapshot:last_completed", today)
+            repo.kv_set(service.completed_key(today), "1")
+
+            summary = service.run_seed_from_last_completed(date=today)
+            assert summary.reason == "already_completed"
+            assert summary.already_completed is True
+            assert summary.completed is True
+            assert summary.reason != "no_source"
+            assert transport.count("/web-interface/view") == 0
+        finally:
+            db.close()
+
+
+
 def test_seed_from_last_completed_does_not_mark_completed():
     clock = 1_800_000_000
     with tempfile.TemporaryDirectory() as tmp:
@@ -631,6 +661,50 @@ def test_seeded_day_skips_without_force_and_seals_with_force():
             db.close()
 
 
+def test_seal_keep_scraped_skips_view_progress_punches_seed():
+    """晚间 seal（keep_scraped）：已 view 的 ok=1 作进度跳过，seed 壳仍打穿。"""
+    clock = 1_800_000_000
+    with tempfile.TemporaryDirectory() as tmp:
+        transport = SnapshotTransport({
+            1: view_for(1, view=111),
+            2: view_for(2, view=222),
+        })
+        db, repo, service = make_service(tmp, transport)
+        try:
+            add_video(repo, 1)
+            add_video(repo, 2)
+            today = DATE
+            # 两条 seed 壳
+            for aid, view in ((1, 10), (2, 20)):
+                repo.upsert_snapshot(
+                    today, aid,
+                    metrics={"view": view, "like": 1, "coin": 0, "favorite": 0,
+                             "share": 0, "reply": 0, "danmaku": 0},
+                    ok=True, err=None, captured_at=clock, src="seed",
+                )
+            # 模拟 seal 中途：aid=1 已真实采集
+            repo.upsert_snapshot(
+                today, 1,
+                metrics={"view": 111, "like": 1, "coin": 0, "favorite": 0,
+                         "share": 0, "reply": 0, "danmaku": 0},
+                ok=True, err=None, captured_at=clock + 1, src="view",
+            )
+            repo.kv_set(service.seeded_key(today), "1")
+
+            sealed = service.run_if_needed(date=today, force=True)
+            assert sealed.completed is True
+            assert sealed.skipped >= 1  # aid=1 进度跳过
+            assert sealed.success >= 1  # aid=2 seed 打穿
+            # 只应请求 aid=2 的 view（aid=1 已是 view 进度）
+            assert transport.count("/web-interface/view") == 1
+            assert int(repo.get_snapshot(today, 1)["view"]) == 111
+            assert repo.get_snapshot(today, 1)["src"] == "view"
+            assert int(repo.get_snapshot(today, 2)["view"]) == 222
+            assert repo.get_snapshot(today, 2)["src"] == "view"
+        finally:
+            db.close()
+
+
 def main():
     test_snapshot_success_resume_and_force()
     test_resume_skips_ok_and_retries_failed()
@@ -646,8 +720,10 @@ def main():
     test_backup_failure_retries_without_api()
     test_run_recent_videos_refreshes_without_completion_markers()
     test_full_snapshot_skips_ok_rows_including_hourly_refresh()
+    test_seed_same_day_completed_is_already_completed_not_no_source()
     test_seed_from_last_completed_does_not_mark_completed()
     test_seeded_day_skips_without_force_and_seals_with_force()
+    test_seal_keep_scraped_skips_view_progress_punches_seed()
     print("OK: 每日快照、断点续跑、失败占位与粉丝采集测试全部通过")
 
 

tests/test_task_state.py

@@ -0,0 +1,266 @@
+# -*- coding: utf-8 -*-
+"""task:state 持久化与启动续跑: uv run python tests/test_task_state.py"""
+
+import json
+import os
+import sys
+import tempfile
+from datetime import datetime
+from pathlib import Path
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+sys.stdout.reconfigure(encoding="utf-8")
+
+from reaction_board.database import Database  # noqa: E402
+from reaction_board.repository import Repository  # noqa: E402
+from reaction_board.scheduler import ScheduledTask, Scheduler  # noqa: E402
+from reaction_board.task_state import (  # noqa: E402
+    STATUS_DEFERRED,
+    STATUS_DONE,
+    STATUS_FAILED,
+    STATUS_PENDING,
+    STATUS_RUNNING,
+    ensure_pending,
+    load_state,
+    mark_deferred,
+    mark_done,
+    mark_failed,
+    mark_running,
+    recover_stale_running,
+    resume_stat_date,
+    startup_next_run,
+)
+from reaction_board.timeutil import SHANGHAI_TZ, next_daily_run  # noqa: E402
+
+
+def _repo(tmp: str) -> tuple[Database, Repository]:
+    db = Database(Path(tmp) / "app.db")
+    db.initialize()
+    return db, Repository(db)
+
+
+def test_restart_mid_run_resumes_same_stat_date():
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 23, 40, tzinfo=SHANGHAI_TZ).timestamp()
+            mark_running(
+                repo, "snapshot", stat_date="2026-11-03", now=now,
+                progress={"done": 10, "total": 100},
+            )
+            # 模拟进程崩溃：running 残留
+            recovered = recover_stale_running(repo, "snapshot", now=now + 60)
+            assert recovered is not None
+            assert recovered.status == STATUS_PENDING
+            assert recovered.stat_date == "2026-11-03"
+            assert recovered.progress == {"done": 10, "total": 100}
+            assert recovered.due_at <= now + 60
+            assert resume_stat_date(repo, "snapshot", fallback="2026-11-04") == "2026-11-03"
+            nxt = startup_next_run(
+                repo, "snapshot",
+                wall_due_at=next_daily_run(now + 60, "23:15"),
+                now=now + 60,
+            )
+            assert nxt <= now + 60
+        finally:
+            db.close()
+
+
+def test_seeded_before_snapshot_at_does_not_force_run():
+    """seed 后挂 pending；白天重启认落盘 due_at（刻意 ≠ 墙钟）且未到不跑。"""
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 10, 30, tzinfo=SHANGHAI_TZ).timestamp()
+            # 刻意偏离墙钟 23:15，证明认落盘而非 fallback
+            slot = datetime(2026, 11, 3, 22, 0, tzinfo=SHANGHAI_TZ).timestamp()
+            wall = next_daily_run(now, "23:15")
+            assert slot != wall
+            ensure_pending(
+                repo, "snapshot", due_at=slot, stat_date="2026-11-03", now=now,
+            )
+            repo.kv_set("snapshot:seeded:2026-11-03", "1")
+            nxt = startup_next_run(
+                repo, "snapshot",
+                wall_due_at=wall,
+                now=now,
+            )
+            assert nxt == slot
+            assert nxt != wall
+            assert nxt > now
+            calls = []
+
+            def run_snapshot():
+                calls.append("ran")
+                return None
+
+            clock = [now]
+            sched = Scheduler(
+                [ScheduledTask("snapshot", 86400.0, run_snapshot, next_run=nxt)],
+                clock=lambda: clock[0],
+            )
+            assert sched.run_due() == []
+            assert calls == []
+        finally:
+            db.close()
+
+
+def test_failed_future_due_at_does_not_run_early():
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 12, 0, tzinfo=SHANGHAI_TZ).timestamp()
+            due = now + 600
+            mark_failed(
+                repo, "snapshot", due_at=due, now=now,
+                stat_date="2026-11-03", error="boom",
+            )
+            nxt = startup_next_run(
+                repo, "snapshot",
+                wall_due_at=next_daily_run(now, "23:15"),
+                now=now,
+            )
+            assert nxt == due
+            clock = [now]
+            calls = []
+            sched = Scheduler(
+                [
+                    ScheduledTask(
+                        "snapshot", 86400.0, lambda: calls.append(1), next_run=nxt,
+                    )
+                ],
+                clock=lambda: clock[0],
+            )
+            assert sched.run_due() == []
+            assert calls == []
+        finally:
+            db.close()
+
+
+def test_due_at_passed_runs_including_failed_retry():
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 12, 10, tzinfo=SHANGHAI_TZ).timestamp()
+            due = now - 30  # 已过重试点
+            mark_failed(
+                repo, "snapshot", due_at=due, now=due - 300,
+                stat_date="2026-11-03", error="boom",
+            )
+            nxt = startup_next_run(
+                repo, "snapshot",
+                wall_due_at=next_daily_run(now, "23:15"),
+                now=now,
+            )
+            # 不得退回今晚墙钟；应立刻可跑
+            assert nxt == due
+            assert nxt <= now
+            wall = next_daily_run(now, "23:15")
+            assert nxt != wall
+            calls = []
+            clock = [now]
+            sched = Scheduler(
+                [
+                    ScheduledTask(
+                        "snapshot", 86400.0, lambda: calls.append(1), next_run=nxt,
+                    )
+                ],
+                clock=lambda: clock[0],
+            )
+            results = sched.run_due()
+            assert len(results) == 1
+            assert calls == [1]
+        finally:
+            db.close()
+
+
+def test_done_writes_next_due_and_blocks_until_then():
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 23, 40, tzinfo=SHANGHAI_TZ).timestamp()
+            next_due = next_daily_run(now, "23:15")
+            mark_done(repo, "snapshot", due_at=next_due, now=now, stat_date="2026-11-03")
+            state = load_state(repo, "snapshot")
+            assert state is not None
+            assert state.status == STATUS_DONE
+            nxt = startup_next_run(
+                repo, "snapshot", wall_due_at=next_due, now=now + 60,
+            )
+            assert nxt == next_due
+            assert nxt > now + 60
+        finally:
+            db.close()
+
+
+def test_no_state_falls_back_to_wall_clock():
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 10, 0, tzinfo=SHANGHAI_TZ).timestamp()
+            wall = next_daily_run(now, "23:15")
+            nxt = startup_next_run(repo, "snapshot", wall_due_at=wall, now=now)
+            assert nxt == wall
+        finally:
+            db.close()
+
+
+
+def test_fail_or_defer_preserves_progress():
+    """H7: mark_failed / mark_deferred 不传 progress 时保留已有 progress 与 stat_date。"""
+    with tempfile.TemporaryDirectory() as tmp:
+        db, repo = _repo(tmp)
+        try:
+            now = datetime(2026, 11, 3, 12, 0, tzinfo=SHANGHAI_TZ).timestamp()
+            progress = {"done": 42, "total": 100, "cursor": "aid:99"}
+            mark_running(
+                repo, "snapshot", stat_date="2026-11-03", now=now,
+                progress=progress,
+            )
+            before = load_state(repo, "snapshot")
+            assert before is not None
+            assert before.progress == progress
+            assert before.stat_date == "2026-11-03"
+
+            due_fail = now + 300
+            mark_failed(
+                repo, "snapshot", due_at=due_fail, now=now + 1,
+                progress=None, error="boom",
+            )
+            failed = load_state(repo, "snapshot")
+            assert failed is not None
+            assert failed.status == STATUS_FAILED
+            assert failed.due_at == due_fail
+            assert failed.progress == progress
+            assert failed.stat_date == "2026-11-03"
+            assert failed.error == "boom"
+
+            due_defer = now + 600
+            mark_deferred(
+                repo, "snapshot", due_at=due_defer, now=now + 2,
+                progress=None, error="snapshot_quiet_window",
+            )
+            deferred = load_state(repo, "snapshot")
+            assert deferred is not None
+            assert deferred.status == STATUS_DEFERRED
+            assert deferred.due_at == due_defer
+            assert deferred.progress == progress
+            assert deferred.stat_date == "2026-11-03"
+            assert deferred.error == "snapshot_quiet_window"
+        finally:
+            db.close()
+
+
+def main_test():
+    test_restart_mid_run_resumes_same_stat_date()
+    test_seeded_before_snapshot_at_does_not_force_run()
+    test_failed_future_due_at_does_not_run_early()
+    test_due_at_passed_runs_including_failed_retry()
+    test_done_writes_next_due_and_blocks_until_then()
+    test_no_state_falls_back_to_wall_clock()
+    test_fail_or_defer_preserves_progress()
+    print("OK: task_state persistence / resume / due_at gates")
+
+
+if __name__ == "__main__":
+    main_test()

tests/test_timeutil.py

@@ -11,7 +11,10 @@ sys.stdout.reconfigure(encoding="utf-8")
 from reaction_board.timeutil import (  # noqa: E402
     SHANGHAI_TZ,
     day_bounds,
+    in_snapshot_quiet_window,
+    interval_next_run,
     is_date,
+    next_after_snapshot_quiet_window,
     next_daily_run,
     parse_hhmm,
     shanghai_date,
@@ -72,11 +75,40 @@ def test_day_bounds_and_date_validation():
         raise AssertionError("非法日期应失败")
 
 
+
+def test_snapshot_quiet_window():
+    assert in_snapshot_quiet_window(ts("2026-11-02 23:00:00"))
+    assert in_snapshot_quiet_window(ts("2026-11-02 23:59:59"))
+    assert in_snapshot_quiet_window(ts("2026-11-03 00:00:00"))
+    assert in_snapshot_quiet_window(ts("2026-11-03 00:59:59"))
+    assert not in_snapshot_quiet_window(ts("2026-11-03 01:00:00"))
+    assert not in_snapshot_quiet_window(ts("2026-11-02 22:59:59"))
+    assert not in_snapshot_quiet_window(ts("2026-11-02 12:00:00"))
+
+    assert next_after_snapshot_quiet_window(ts("2026-11-02 22:00:00")) == ts(
+        "2026-11-02 22:00:00"
+    )
+    assert next_after_snapshot_quiet_window(ts("2026-11-02 23:15:00")) == ts(
+        "2026-11-03 01:00:00"
+    )
+    assert next_after_snapshot_quiet_window(ts("2026-11-03 00:30:00")) == ts(
+        "2026-11-03 01:00:00"
+    )
+    # 22:50 + 60min = 23:50 → 推到次日 01:00
+    assert interval_next_run(ts("2026-11-02 22:50:00"), 3600) == ts(
+        "2026-11-03 01:00:00"
+    )
+    assert interval_next_run(ts("2026-11-02 12:00:00"), 3600) == ts(
+        "2026-11-02 13:00:00"
+    )
+
+
 def main():
     test_shanghai_date_boundary()
     test_parse_hhmm()
     test_next_daily_run()
     test_day_bounds_and_date_validation()
+    test_snapshot_quiet_window()
     print("OK: 上海时区与每日时间计算测试全部通过")
 
 

tests/test_web_queries.py

@@ -610,12 +610,15 @@ def test_growth_auto_falls_back_to_one_day():
 
 def test_snapshot_recent_next_run_follows_completion():
     import json
+    from datetime import datetime
+    from reaction_board.timeutil import SHANGHAI_TZ, interval_next_run
 
     with tempfile.TemporaryDirectory() as tmp:
         db, q = setup(tmp)
         try:
             repo = Repository(db)
-            finished = FIXED_NOW - 10
+            # 午后完成：间隔后仍在静默窗外，应严格 finished+3600
+            finished = int(datetime(2026, 10, 31, 15, 0, 0, tzinfo=SHANGHAI_TZ).timestamp())
             repo.kv_set(
                 "task:last:snapshot_recent",
                 json.dumps({
@@ -627,6 +630,128 @@ def test_snapshot_recent_next_run_follows_completion():
             DASHBOARD_CACHE.clear()
             system = q.system()
             assert system["next_runs"]["snapshot_recent"] == finished + 3600
+            assert system["next_runs"]["snapshot_recent"] == int(
+                interval_next_run(float(finished), 3600)
+            )
+        finally:
+            db.close()
+
+
+def test_snapshot_recent_next_run_defers_quiet_window_and_retry_on_skip():
+    import json
+    from datetime import datetime
+    from reaction_board.timeutil import SHANGHAI_TZ, next_after_snapshot_quiet_window
+
+    with tempfile.TemporaryDirectory() as tmp:
+        db, q = setup(tmp)
+        try:
+            repo = Repository(db)
+            # 22:50 成功 → 下一档 23:50 落入静默窗 → 推到 01:00
+            finished = int(datetime(2026, 10, 31, 22, 50, 0, tzinfo=SHANGHAI_TZ).timestamp())
+            repo.kv_set(
+                "task:last:snapshot_recent",
+                json.dumps({
+                    "name": "snapshot_recent", "ok": True,
+                    "started_at": finished - 5, "finished_at": finished,
+                    "error": None, "summary": {},
+                }),
+            )
+            DASHBOARD_CACHE.clear()
+            system = q.system()
+            expected = int(datetime(2026, 11, 1, 1, 0, 0, tzinfo=SHANGHAI_TZ).timestamp())
+            assert system["next_runs"]["snapshot_recent"] == expected
+
+            # skip 记录：补跑到窗后 01:00，不得当成功推进
+            quiet = int(datetime(2026, 10, 31, 23, 20, 0, tzinfo=SHANGHAI_TZ).timestamp())
+            repo.kv_set(
+                "task:last:snapshot_recent",
+                json.dumps({
+                    "name": "snapshot_recent", "ok": False, "skipped": True,
+                    "started_at": quiet, "finished_at": quiet,
+                    "error": "snapshot_quiet_window",
+                    "summary": {"skipped": True, "reason": "snapshot_quiet_window"},
+                }),
+            )
+            DASHBOARD_CACHE.clear()
+            system = q.system()
+            assert system["next_runs"]["snapshot_recent"] == int(
+                next_after_snapshot_quiet_window(float(quiet))
+            )
+            assert not any(e["task"] == "snapshot_recent" for e in system["errors"])
+        finally:
+            db.close()
+
+
+def test_snapshot_failed_next_run_retries_soon_not_next_day():
+    import json
+    from datetime import datetime
+    from reaction_board.timeutil import SHANGHAI_TZ, next_daily_run
+
+    with tempfile.TemporaryDirectory() as tmp:
+        db, q = setup(tmp)
+        try:
+            repo = Repository(db)
+            finished = int(datetime(2026, 10, 31, 23, 15, 0, tzinfo=SHANGHAI_TZ).timestamp())
+            repo.kv_set(
+                "task:last:snapshot",
+                json.dumps({
+                    "name": "snapshot", "ok": False,
+                    "started_at": finished - 1, "finished_at": finished,
+                    "error": "boom", "summary": {"completed": False},
+                }),
+            )
+            DASHBOARD_CACHE.clear()
+            system = q.system()
+            assert system["next_runs"]["snapshot"] == finished + 300
+            assert system["next_runs"]["snapshot"] != int(
+                next_daily_run(finished, "23:15")
+            )
+            assert any(e["task"] == "snapshot" for e in system["errors"])
+        finally:
+            db.close()
+
+
+def test_snapshot_seed_skip_next_run_retries_ten_minutes():
+    import json
+    from datetime import datetime
+    from reaction_board.timeutil import SHANGHAI_TZ, next_daily_run
+
+    with tempfile.TemporaryDirectory() as tmp:
+        db, q = setup(tmp)
+        try:
+            repo = Repository(db)
+            finished = int(datetime(2026, 11, 1, 0, 0, 0, tzinfo=SHANGHAI_TZ).timestamp())
+            repo.kv_set(
+                "task:last:snapshot_seed",
+                json.dumps({
+                    "name": "snapshot_seed", "ok": False, "skipped": True,
+                    "started_at": finished - 1, "finished_at": finished,
+                    "error": "snapshot_running",
+                    "summary": {"skipped": True, "reason": "snapshot_running"},
+                }),
+            )
+            DASHBOARD_CACHE.clear()
+            system = q.system()
+            assert system["next_runs"]["snapshot_seed"] == finished + 600
+            assert system["next_runs"]["snapshot_seed"] != int(
+                next_daily_run(finished, "00:00")
+            )
+            # skip 不算 errors
+            assert not any(e["task"] == "snapshot_seed" for e in system["errors"])
+
+            # no_source 同样 10 分钟
+            repo.kv_set(
+                "task:last:snapshot_seed",
+                json.dumps({
+                    "name": "snapshot_seed", "ok": False, "skipped": True,
+                    "started_at": finished - 1, "finished_at": finished,
+                    "error": "no_source",
+                    "summary": {"skipped": True, "reason": "no_source"},
+                }),
+            )
+            DASHBOARD_CACHE.clear()
+            system = q.system()
+            assert system["next_runs"]["snapshot_seed"] == finished + 600
         finally:
             db.close()
 
@@ -882,6 +1007,9 @@ def main_test():
     test_system_tasks_sorted_by_started_at_desc()
     test_growth_auto_falls_back_to_one_day()
     test_snapshot_recent_next_run_follows_completion()
+    test_snapshot_recent_next_run_defers_quiet_window_and_retry_on_skip()
+    test_snapshot_failed_next_run_retries_soon_not_next_day()
+    test_snapshot_seed_skip_next_run_retries_ten_minutes()
     test_overview_adjacent_previous_with_seed_and_dual_fields()
     test_same_day_growth_baselines()
     test_growth_rate_sort_uses_prior()
@@ -948,13 +1076,15 @@ def test_system_status_uses_completed_date_and_running_flags():
 
 def test_system_next_runs_follow_last_completion_not_wall_clock():
     import json
+    from datetime import datetime
+    from reaction_board.timeutil import SHANGHAI_TZ
 
     with tempfile.TemporaryDirectory() as tmp:
         db, q = setup(tmp)
         try:
             repo = Repository(db)
-            # 发现任务刚完成不久：下次 = finished + 60min，即使 now 已超过也不跳档
-            finished = FIXED_NOW - 30
+            # 午后完成：下次 = finished + 60min，不跳档、不误入静默窗推算
+            finished = int(datetime(2026, 10, 31, 15, 0, 0, tzinfo=SHANGHAI_TZ).timestamp())
             repo.kv_set(
                 "task:last:discovery",
                 json.dumps({
@@ -966,7 +1096,6 @@ def test_system_next_runs_follow_last_completion_not_wall_clock():
             DASHBOARD_CACHE.clear()
             system = q.system()
             assert system["next_runs"]["discover"] == finished + 3600
-            # 仍小于/等于 now 也可以（表示本轮后的下一档已到期或未到），但绝不能跳到 +7200
             assert system["next_runs"]["discover"] != finished + 7200
         finally:
             db.close()