fix(sync): 「全部历史」真的拉全部历史，自动同步不再每次静默失败

四个独立的 bug 叠在一起，表现为「只同步两天、没有进度」：

* 前端 `...(days ? { days } : {})` 把 days=0 当成未传。「全部历史」
  存的就是 0，请求体里根本没有 days，后端退回 7 天默认值。
* scheduler 用 `s[0]` 读 query_one 返回的 dict，抛 KeyError 后被
  per-account 的 except 吞掉。只要用户存过一次设置，每 30 分钟的
  自动同步就一次都没成功过——库里那 2 天全是手动点出来的。
* 增量同步查 `health_daily`（表其实叫 health_data），后台线程直接
  死掉，状态永远卡在 syncing，进度条不动。
* UI 完全不看 /sync 的返回值，rate_limited 时按钮点了没反应；轮询
  结束时又把 rate_limited 归进 else 分支报「同步完成」。

顺带：
* 日循环遇到 429 立即退避并保留已拉到的天数，而不是当成「跳过一天」
  继续往下捶 700 天——这正是之前限流死循环的来源之一。
* 定时循环显式传 SYNC_DAYS。历史范围按 UI 文案只描述手动全量同步，
  让半小时一次的 tick 重拉 730 天必然把限流撞得更深。
* 短同步逐天上报进度（原来每 5 天一次，7 天的同步全程停在 0）。
* /sync 路由重复解析 body，空 body 会 None.get 崩。
* 4 个 StubGarth 缺 configure()，7 个测试在此之前一直是红的。

新增 deploy/push.sh：NAS 只认密码，脚本开一个 ssh 复用连接，密码只
输一次，后面推送 / 重启 / 健康检查全走它。不碰 .env、.venv 和数据库。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/services/garmin.py

@@ -1021,8 +1021,14 @@ def sync_data(user_id, creds, days=None, client=None):
     # -1 means "incremental sync": pick up from the latest date already in the
     # local database rather than pulling a fixed window.
     if days == -1:
-        row = query_one("SELECT MAX(date) FROM health_daily WHERE user_id = ?", (user_id,))
-        latest = row[0] if row and row[0] else None
+        # The daily rows live in `health_data`; `health_daily` is the name of
+        # the *writer* (health.upsert_health_daily), not of any table. Reading
+        # it raised inside the background thread, so an incremental sync died
+        # silently and left the status stuck on "syncing".
+        row = query_one(
+            "SELECT MAX(date) AS latest FROM health_data WHERE user_id = ?", (user_id,)
+        )
+        latest = row["latest"] if row and row.get("latest") else None
         if latest is None:
             days = DEFAULT_SYNC_DAYS  # first sync → fall back to default window
         else:
@@ -1038,12 +1044,21 @@ def sync_data(user_id, creds, days=None, client=None):
 
     days_synced = 0
     day_errors = []
+    rate_limited_mid_run = None
     for i in range(days):
         date_str = (today - datetime.timedelta(days=i)).isoformat()
         try:
             record = _extract_daily(client, date_str)
             record.update(extras.daily_extras(client, date_str))
         except Exception as e:
+            # A 429 inside the day loop used to be filed as one more skipped
+            # day, so a 730-day backfill kept hammering Garmin for another
+            # 700 days and drove the throttle deeper. Stand down at the first
+            # one and keep whatever was already stored.
+            if isinstance(e, RateLimited) or _is_rate_limited(e):
+                _note_rate_limit(user_id)
+                rate_limited_mid_run = describe(e)
+                break
             day_errors.append(f"{date_str}: {describe(e)}")
             continue
         # A day Garmin has no data for comes back all-None; storing it would
@@ -1060,15 +1075,34 @@ def sync_data(user_id, creds, days=None, client=None):
                 except Exception as e:  # noqa: BLE001
                     day_errors.append(f"{date_str} series: {describe(e)}")
 
-        # Reported every few days rather than every day: the write is cheap
-        # but not free, and the UI polls on a 2s cadence anyway.
-        if (i + 1) % 5 == 0 or i + 1 == days:
+        # A long backfill reports every fifth day — the write is cheap but not
+        # free. A short one reports every day: at 7 days a "every 5th" cadence
+        # meant the bar sat at 0 for most of the run and then vanished.
+        if days <= 30 or (i + 1) % 5 == 0 or i + 1 == days:
             _set_sync_status(
                 user_id, "syncing", now,
                 records_synced=days_synced, progress_current=i + 1,
                 progress_total=days, stage=f"每日数据 {date_str}",
             )
 
+    if rate_limited_mid_run:
+        until, msg = _rate_limit_block(user_id)
+        message = msg or "Garmin 限制了请求频率，已自动退避。"
+        _set_sync_status(
+            user_id, "rate_limited", now, records_synced=days_synced,
+            progress_current=days_synced, progress_total=days, stage=None,
+            last_error=message,
+        )
+        return {
+            "status": "rate_limited",
+            "recordsSynced": days_synced,
+            "message": f"已同步 {days_synced} 天后被 Garmin 限流：{message}",
+            "retryAfterSeconds": (
+                int((until - datetime.datetime.utcnow()).total_seconds()) if until else None
+            ),
+            "lastSyncTime": now,
+        }
+
     activities_synced = 0
     _set_sync_status(user_id, "syncing", now, records_synced=days_synced,
                      progress_current=days, progress_total=days,

backend/services/scheduler.py

@@ -130,7 +130,8 @@ def sync_all_accounts(days=None, respect_schedule=False):
 
     When `days` is not provided, each account's saved `history_days` from
     `user_settings` is used (the user's 历史范围 picker), falling back to
-    `SYNC_DAYS`.
+    `SYNC_DAYS`. The scheduled loop always passes `days` explicitly — 历史范围
+    is the window for the manual 全量同步, not for a half-hourly tick.
     """
     rows = query_all("SELECT user_id FROM garmin_tokens")
     results = []
@@ -150,8 +151,14 @@ def sync_all_accounts(days=None, respect_schedule=False):
             # Resolve the sync window: prefer the user's saved history_days,
             # then the caller override, then the global default.
             if days is None:
-                s = query_one("SELECT history_days FROM user_settings WHERE user_id = ?", (uid,))
-                user_days = s[0] if s and s[0] is not None else None
+                # query_one returns a dict, so `s[0]` raised KeyError — caught
+                # by the per-account handler below, which meant every single
+                # scheduled tick failed for every account and nothing was ever
+                # synced automatically.
+                s = query_one(
+                    "SELECT history_days FROM user_settings WHERE user_id = ?", (uid,)
+                )
+                user_days = s.get("history_days") if s else None
                 if user_days == 0:
                     user_days = 730  # 全部历史 → 最大范围
                 d = user_days if user_days is not None else SYNC_DAYS
@@ -180,7 +187,12 @@ def _loop():
         try:
             if claim(interval=TICK_SECONDS):
                 try:
-                    sync_all_accounts(respect_schedule=True)
+                    # Explicitly the recent window, never the user's 历史范围:
+                    # that setting describes the manual 全量同步. Letting a
+                    # half-hourly tick re-pull "全部历史" meant 730 days x ~7
+                    # Garmin calls every 30 minutes, which is precisely what
+                    # kept the account in a 429 loop.
+                    sync_all_accounts(days=SYNC_DAYS, respect_schedule=True)
                 finally:
                     release()
         except Exception as e:  # noqa: BLE001 - the loop must outlive any single failure