fix(ai): subject 里塞了行数，每轮询一次就新建一个任务

生产上 trends 队列里积了 36 个任务，subject 是 2026-09-01:1033、:1039、
:1044……一路涨。这台账号当时正在补历史，get_summary 的行数每隔几分钟就变，
而我把 len(rows) 写进了 subject——subject 同时是缓存键和任务队列的键，一变
就是一条全新的任务，轮询几次就刷出十几条。

subject 该回答的是「这条解读是关于什么的」，不是「当时有多少行数据」。
数据变化本来就由 fingerprint 负责。

- trends 的 subject 改成快照日期；sleep 用配置的窗口常量而不是实际夜数
  （缺一晚也不该换键）；challenges 用固定键
- 加了不变量测试：补一天历史数据后 subject 不许变；任何 subject 段都不许
  长得像行数

顺带加一层兜底 jobs.supersede()：单实例 scope 只该有一个在跑的 subject，
队列里同 kind 的其它 pending 任务是关于已经不存在的快照的，跑完也没人看。
per_item 的 daily / activity 不受影响——它们本来就一天一条、一次运动一条。
兜底不是机制，机制是 subject 稳定；它存在只是因为这次 subject 不稳定，而
36 条任务堆在那里之前没人发现。

顺带按要求把 AiPanel 改成默认精简：只显示标题、来源和一句话结论，点「展开
详细」才出要点/建议/依据，可再收起——和今日晨报卡片一致。这些面板压在本来
就很密的图表页上面，全部默认展开会把真正的数据一次性挤到屏幕外。

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

backend/services/analysis.py

@@ -499,6 +499,8 @@ def get_scope_insight(user_id, scope, subject=None, refresh=False):
             insight, meta = cached
             return {"insight": insight, "context": context, "meta": meta}
 
+    if not scopes.SCOPES[scope].per_item:
+        jobs.supersede(user_id, scope, resolved)
     state = jobs.enqueue(user_id, scope, resolved, fingerprint,
                          jobs.PRIORITY_INTERACTIVE)
     return {
@@ -540,12 +542,15 @@ def prefetch_insights(user_id):
         if not built:
             continue
         resolved, context = built
+        if not scopes.SCOPES[scope].per_item:
+            jobs.supersede(user_id, scope, resolved)
         jobs.enqueue(user_id, scope, resolved, _context_fingerprint(context),
                      jobs.PRIORITY_PREFETCH)
         queued.append(scope)
 
     context = insights.build_context(user_id)
     if context:
+        jobs.supersede(user_id, "briefing", context["snapshotDate"])
         jobs.enqueue(user_id, "briefing", context["snapshotDate"],
                      _context_fingerprint(context), jobs.PRIORITY_PREFETCH)
         queued.append("briefing")

backend/services/jobs.py

@@ -150,6 +150,25 @@ def enqueue(user_id, kind, subject, fingerprint=None,
     return "pending"
 
 
+def supersede(user_id, kind, subject):
+    """Drop queued work of the same kind for a different subject.
+
+    A single-entry scope has one live subject; anything else queued under that
+    kind is about a snapshot that no longer exists, and running it would spend
+    a gateway call on an answer nothing will read.
+
+    This is a safety net, not the mechanism: subjects are supposed to be stable
+    (see scopes.py). It exists because they were not — a row count in the key
+    made every poll mint a new `trends` job, and production had 36 of them
+    queued before anyone noticed.
+    """
+    execute(
+        "DELETE FROM ai_jobs WHERE user_id = ? AND kind = ? AND subject != ? "
+        "AND status = 'pending'",
+        [user_id, kind, subject],
+    )
+
+
 def status_of(user_id, kind, subject):
     row = query_one("SELECT * FROM ai_jobs WHERE id = ?",
                     [job_id(user_id, kind, subject)])

backend/services/scopes.py

@@ -164,7 +164,9 @@ def build_sleep(user_id, subject=None):
                       + ("，偏低。" if deep < deep_low else "，达标。"),
         })
 
-    return f"{window[-1]['date']}:{len(window)}", {
+    # The configured window, not `len(window)`: a night missing from the record
+    # would otherwise change the key and queue a second job for the same screen.
+    return f"{window[-1]['date']}:{NIGHTS}", {
         "scope": "sleep",
         "label": "睡眠",
         "windowNights": len(window),
@@ -286,7 +288,7 @@ def build_trends(user_id, subject=None):
     if not highlights:
         highlights = [{"title": "整体", "detail": "各项指标长期走势平稳，无明显方向性变化。"}]
 
-    return f"{rows[-1]['date']}:{len(rows)}", {
+    return rows[-1]["date"], {
         "scope": "trends",
         "label": "长期趋势",
         "highlights": highlights,
@@ -492,7 +494,9 @@ def build_challenges(user_id, subject=None):
             "detail": f"{c.get('startDate')} ~ {c.get('endDate')}，状态 {c.get('status')}。",
         })
 
-    return f"{rows[0].get('startDate')}:{len(rows)}", {
+    # One entry per account: which challenges exist is what the fingerprint
+    # tracks, so the key does not need to encode how many there are.
+    return "all", {
         "scope": "challenges",
         "label": "挑战赛",
         "highlights": highlights,
@@ -559,13 +563,17 @@ def build_activity(user_id, subject=None):
 
 # --- registry ----------------------------------------------------------------
 class Scope:
-    __slots__ = ("name", "label", "build", "needs_subject")
+    __slots__ = ("name", "label", "build", "needs_subject", "per_item")
 
-    def __init__(self, name, label, build, needs_subject=False):
+    def __init__(self, name, label, build, needs_subject=False, per_item=False):
         self.name = name
         self.label = label
         self.build = build
         self.needs_subject = needs_subject
+        # `per_item` scopes legitimately have one entry per date or per
+        # session. Every other scope has exactly one live entry, so an older
+        # subject sitting in the queue is stale work — see `supersede`.
+        self.per_item = per_item
 
 
 SCOPES = {
@@ -574,12 +582,13 @@ SCOPES = {
         Scope("sleep", "睡眠", build_sleep),
         Scope("exercise", "运动", build_exercise),
         Scope("trends", "长期趋势", build_trends),
-        Scope("daily", "每日数据", build_daily),
+        Scope("daily", "每日数据", build_daily, per_item=True),
         Scope("body", "身体成分", build_body),
         Scope("race", "成绩预测", build_race),
         Scope("bodyAge", "身体年龄", build_body_age),
         Scope("challenges", "挑战赛", build_challenges),
-        Scope("activity", "运动详情", build_activity, needs_subject=True),
+        Scope("activity", "运动详情", build_activity, needs_subject=True,
+              per_item=True),
     )
 }