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/tests/test_coach.py

@@ -981,3 +981,75 @@ class TestOutageRecovery:
         assert out["meta"]["pending"] is False
         assert out["meta"]["reason"], "the card should be able to say why"
         assert out["insight"]["headline"], "and still show the computed facts"
+
+
+class TestSubjectStability:
+    """The subject keys the cache *and* the job queue, so it must identify what
+    the insight is about — never how much data happened to be there.
+
+    A count in the key made every poll mint a fresh job: production
+    accumulated a dozen `trends` jobs in minutes, all for the same screen."""
+
+    def test_the_subject_does_not_move_when_the_data_grows(
+        self, db, user, seed_health
+    ):
+        seed_health([
+            {"date": f"2026-08-{d:02d}", "steps": 8000, "heart_rate": 60,
+             "hrv": 45, "sleep_duration": 7, "sleep_quality": 80, "stress": 30}
+            for d in range(1, 29)
+        ])
+        before = {
+            name: scopes.build(user["id"], name)[0]
+            for name in ("health", "sleep", "exercise", "trends")
+            if scopes.build(user["id"], name)
+        }
+
+        # A backfilled day, as a history sync would add: older than the newest,
+        # so what the screen is about has not changed.
+        seed_health([{"date": "2026-07-31", "steps": 7000, "heart_rate": 61,
+                      "hrv": 44, "sleep_duration": 7, "sleep_quality": 79,
+                      "stress": 31}])
+
+        after = {name: scopes.build(user["id"], name)[0] for name in before}
+        assert after == before
+
+    def test_no_subject_encodes_a_row_count(self, month):
+        for name in ("health", "sleep", "exercise", "trends", "challenges"):
+            built = scopes.build(month["id"], name)
+            if not built:
+                continue
+            subject = built[0]
+            # A count would grow without bound; a date or a window constant
+            # will not. This catches the shape of the mistake, not one instance.
+            for part in str(subject).split(":"):
+                assert not (part.isdigit() and int(part) > 400), \
+                    f"{name} subject {subject!r} looks like a row count"
+
+
+class TestSupersede:
+    def test_opening_a_screen_clears_queued_work_for_an_older_snapshot(
+        self, month, gateway
+    ):
+        jobs.enqueue(month["id"], "trends", "2026-08-01")
+        jobs.enqueue(month["id"], "trends", "2026-08-15")
+        analysis_svc.get_scope_insight(month["id"], "trends")
+        rows = analysis_svc.query_all(
+            "SELECT subject FROM ai_jobs WHERE kind = 'trends'")
+        assert len(rows) == 1, "only the current snapshot should be queued"
+
+    def test_per_item_screens_keep_one_job_each(self, month, gateway):
+        """每日 and 运动详情 legitimately have one entry per date / session."""
+        jobs.enqueue(month["id"], "daily", "2026-08-01")
+        jobs.enqueue(month["id"], "daily", "2026-08-02")
+        analysis_svc.get_scope_insight(month["id"], "daily", subject="2026-08-03")
+        rows = analysis_svc.query_all(
+            "SELECT subject FROM ai_jobs WHERE kind = 'daily'")
+        assert len(rows) == 3
+
+    def test_running_work_is_not_dropped_from_under_the_worker(self, month, gateway):
+        jobs.enqueue(month["id"], "trends", "2026-08-01")
+        jobs._claim_next()
+        analysis_svc.get_scope_insight(month["id"], "trends")
+        statuses = {r["subject"]: r["status"] for r in analysis_svc.query_all(
+            "SELECT subject, status FROM ai_jobs WHERE kind = 'trends'")}
+        assert statuses.get("2026-08-01") == "running"