diff --git a/backend/app/api/catalog.py b/backend/app/api/catalog.py
index f5a85a6b..8810486c 100644
--- a/backend/app/api/catalog.py
+++ b/backend/app/api/catalog.py
@@ -114,8 +114,9 @@ GUIDANCE_NOT_TEXT = "起草要求要是一段文字（要清除起草要求请
 class ProfileJobIn(BaseModel):
     #: 要剖析的表（范围 selected，按给的顺序去重）；最多 500 张（超了 422，说明是中文）
     tables: list[str] | None = None
-    #: 不指定表时的范围：needed 需要剖析的表（缺省）、all 全部表（都不含视图、没有可剖析的项的表）
-    scope: Literal["needed", "all"] | None = None
+    #: 不指定表时的范围：needed 需要剖析的表（缺省）、all 全部表（都不含视图、没有可剖析的项的表）、observe 取值观察（只观察
+    #: 有数据目录、观察到期的表，不剖析；见 data/catalog_profile_observe）。maintenance 是定时维护用的，这里不收
+    scope: Literal["needed", "all", "observe"] | None = None
 
 
 async def _resolved(session: AsyncSession, source_id: str) -> tuple[DataSource, Any]:
@@ -226,14 +227,21 @@ async def _detail(session: AsyncSession, source: Any, table: str, entry: catalog
 
 
 async def _table_key(session: AsyncSession, source: Any, table: str) -> tuple[str, catalog.CatalogEntry | None]:
-    """路径里的表名 → 目录的键。表结构里有的按表结构的写法；表结构里没有、但还留着目录的也认（便于清理）。"""
+    """路径里的表名 → 目录的键。表结构里有的按表结构的写法；表结构里没有、但还留着目录的也认（便于清理）。都没有时 404，
+    说清是哪一种「不在」（_missing）。"""
     key = catalog.resolve_table_name(source.schema_cache, table)
     entry = await catalog.read_entry(session, source.id, key or table)
     if key is None and entry is None:
-        raise HTTPException(404, f"数据源中没有表 {table}，可能已被删除或尚未探查结构")
+        raise await _missing(session, source, table)
     return key or table, entry
 
 
+async def _missing(session: AsyncSession, source: Any, table: str) -> HTTPException:
+    """表结构里没有这张表的 404：表已不在、数据源里没有、在探查范围外、读不到结构、说不清，和 /schema、/catalog/search、
+    目录清单的 absent_reason 同一套（source_objects.absent_message）。"""
+    return HTTPException(404, await source_objects.absent_message(session, source, table))
+
+
 def _conflict(e: catalog.CatalogConflict) -> HTTPException:
     return HTTPException(409, str(e))
 
@@ -906,7 +914,7 @@ async def catalog_table_assistant(source_id: str, table: str, payload: Assistant
     row, source = await _resolved(session, source_id)
     key = catalog.resolve_table_name(source.schema_cache, table)
     if key is None:
-        raise HTTPException(404, f"数据源中没有表 {table}，可能已被删除或尚未探查结构")
+        raise await _missing(session, source, table)
     messages = [m.model_dump() for m in payload.messages]
     if problem := catalog_assistant.conversation_problem(messages):
         raise HTTPException(422, problem)
diff --git a/backend/app/api/evals.py b/backend/app/api/evals.py
index 3c722aaa..37da3987 100644
--- a/backend/app/api/evals.py
+++ b/backend/app/api/evals.py
@@ -1019,21 +1019,56 @@ async def _actions(session: AsyncSession, row: EvalRun) -> dict[str, Any]:
 @router.get("/evaluations/{evaluation_id}/results")
 async def evaluation_results(evaluation_id: str, session: AsyncSession = Depends(get_session)) -> dict[str, Any]:
     """逐题结果：结论、各期望的结论（期望值和答案里对上的数、有没有引用、维度）、作答路径、复核、出具档位、SQL 差在哪、look、
-    耗时、按角色的用量，点 run_id 去运行详情。明细清理了的照实写 cleaned。"""
+    耗时、按角色的用量，点 run_id 去运行详情。明细清理了的照实写 cleaned。question 是这一题实际问的问题（变量渲染好了，
+    见 _asked_questions；说不准时为 null），question_template 是题目的原文（带 {{ schedule.… }} 变量的模板）。"""
     from app.eval import runner
 
     row = await _evaluation(session, evaluation_id)
     results = [runner.result_dict(r) for r in (await session.execute(
         select(runner.EvalResult).where(runner.EvalResult.eval_run_id == row.id)
         .order_by(runner.EvalResult.seq))).scalars()]
-    questions = dict((await session.execute(select(EvalCase.id, EvalCase.question).where(
-        EvalCase.id.in_({r["case_id"] for r in results if r["case_id"]})))).all()) if results else {}
+    cases = {c.id: (c.question, c.case_hash) for c in (await session.execute(select(EvalCase).where(
+        EvalCase.id.in_({r["case_id"] for r in results if r["case_id"]})))).scalars()} if results else {}
+    asked = await _asked_questions(session, row, results, cases)
     for r in results:
-        r["question"] = questions.get(r["case_id"])
+        template = (cases.get(r["case_id"]) or (None, None))[0]
+        r["question"], r["question_template"] = asked.get(r["id"]), template
         r["cost"] = _result_cost(r.get("usage") or {})
     return {"evaluation_id": row.id, "cleaned": row.cleaned_at is not None, "results": results}
 
 
+async def _asked_questions(session: AsyncSession, row: EvalRun, results: list[dict[str, Any]],
+                           cases: dict[str, tuple[str, str]]) -> dict[str, str | None]:
+    """每一题实际问的问题 {结果 id: 问题}：题目的模板按这次评测的截至时间（as_of）、题集的时区、定时的 cron 渲染——和执行时
+    runner._unit 交给作答的是同一个算法、同一组输入，所以一字不差（as_of 在建评测时就定下、重试也不变）。题目在这次评测之后
+    改过（结果记的 case_hash 对不上现在的）：现在的模板不一定是当时问的，改取作答运行的输入里的 question；也没有（反问、
+    建图失败没有运行，工作流题的输入里没有问题）就是 None，不拿现在的模板编一个。"""
+    from app.db.models import Run
+    from app.eval import runner
+    from app.eval.cases import render, variable_values
+
+    if not results:
+        return {}
+    leader = row if runner.leader_of(row) == row.id else await session.get(EvalRun, runner.leader_of(row))
+    suite = await session.get(EvalSuite, row.suite_id)
+    as_of = (leader.as_of if leader is not None else None) or row.as_of
+    values = variable_values(as_of, suite.timezone, ((leader or row).fingerprint or {}).get("cron")) \
+        if as_of is not None and suite is not None else None
+    stale = {r["id"] for r in results
+             if r["case_id"] not in cases or cases[r["case_id"]][1] != r["case_hash"] or values is None}
+    runs = {r["run_id"] for r in results if r["id"] in stale and r["run_id"]}
+    inputs = dict((await session.execute(select(Run.id, Run.input).where(Run.id.in_(runs))))
+                  .all()) if runs else {}
+    out: dict[str, str | None] = {}
+    for r in results:
+        if r["id"] in stale:
+            given = (inputs.get(r["run_id"]) or {}).get("question") if r["run_id"] else None
+            out[r["id"]] = given if isinstance(given, str) and given else None
+        else:
+            out[r["id"]] = render(cases[r["case_id"]][0] or "", values or {})
+    return out
+
+
 def _result_cost(usage: dict[str, Any]) -> dict[str, Any]:
     """一题的花费：助手（runner 记的 assistant 已含复核）+ 写作 + 结论裁判，和这次评测记在账上的同一份；估不出金额的调用另计
     次数（界面写「≥」或「金额无法估算」）。加在服务端：界面不算数。"""
diff --git a/backend/app/api/runs.py b/backend/app/api/runs.py
index f06d8b4a..606f82ad 100644
--- a/backend/app/api/runs.py
+++ b/backend/app/api/runs.py
@@ -418,6 +418,8 @@ async def resume_run(
             approval = await session.get(Approval, target)
         if approval is not None and approval.run_id == run_id and approval.status == "pending":
             return await _respond_to(target, payload.response, x_actor)
+    from app.engine.runner import ContinueRefused
+
     try:
         # approval_id 以前定义了却从没传下去——一次运行里有多条待审批时，
         # 引擎无从知道这次回复的是哪一条
@@ -428,6 +430,9 @@ async def resume_run(
         )
     except KeyError as e:
         raise _missing(e) from e
+    except ContinueRefused as e:
+        # 被驳回的（run_rejected）：和 /continue 同一个机器码
+        raise CodedHTTPException(409, str(e), e.code) from e
     except ValueError as e:
         raise HTTPException(409, str(e)) from e
 
diff --git a/backend/app/api/shares.py b/backend/app/api/shares.py
index 4bcc1b94..27006a5a 100644
--- a/backend/app/api/shares.py
+++ b/backend/app/api/shares.py
@@ -216,7 +216,9 @@ async def list_shares(app_id: str, session: AsyncSession = Depends(get_session),
 @_manage.delete("/{app_id}/shares/{share_id}")
 async def revoke_share(app_id: str, share_id: str, session: AsyncSession = Depends(get_session),
                        x_actor: str | None = Header(default=None)) -> dict[str, Any]:
-    """撤销：之后打开回 410「链接已撤销」。撤销过的再撤销照旧交回它（不重复记审计）。管理员关了开关也能撤销。"""
+    """撤销：之后打开回 410「链接已撤销」。撤销过的再撤销照旧交回它（不重复记审计）。管理员关了开关也能撤销。交回的形状和
+    建、列的同一个（page_title 是页面标题，取法同列表：应用当前的版本）。"""
+    from app.api.apps import _previous
     from app.apps.audit import record
 
     app, _ = await _managed(session, app_id, x_actor)
@@ -228,7 +230,8 @@ async def revoke_share(app_id: str, share_id: str, session: AsyncSession = Depen
         await record(session, action="app.share_revoke", subject_kind="view_share", subject_id=row.id,
                      actor=actor_of(x_actor), data={"app_id": app.id, "page_id": row.page_id, "prefix": row.prefix})
         await session.commit()
-    return share_out(row)
+    version = await _previous(session, app)
+    return share_out(row, _titles(version.spec if version is not None else None))
 
 
 # --------------------------------------------------------------------------
diff --git a/backend/app/core/job_runtime.py b/backend/app/core/job_runtime.py
index 98b641a9..f1f55b10 100644
--- a/backend/app/core/job_runtime.py
+++ b/backend/app/core/job_runtime.py
@@ -22,8 +22,9 @@
   说明源上正有别的事，接口回 409；占到了，执行期间新建的任务排在它后面。**队里的事是开放的**：运行时不认具体哪几种，
   每种在自己的模块里 register_lane_kind 登记一次（kind → 叫法），进队时只核对登记过——第 2 波的后台探查（introspect）
   登记一行就排进同一条队，不用改这里；拼错的 kind 当场报错，不会悄悄排成一种谁都不认识的事。
-- **close_job / interrupt_on_startup**：收尾（还没结束的表记成某个状态，任务记成某个状态），和服务启动时的整理：排队、
-  进行中的标 interrupted，没结束的表记 cancelled。**不自动续跑**：复制出来的数据目录（检查脚本、验收副本）启动时不能
+- **close_job / interrupt_on_startup**：收尾（还没结束的表记成某个状态，任务记成某个状态；可选的 finish 在同一个事务里
+  补一步，评测拿它按改好的逐项状态重算计数、记停下的原因），和服务启动时的整理：排队、进行中的标 interrupted，没结束的
+  表记 cancelled。**不自动续跑**：复制出来的数据目录（检查脚本、验收副本）启动时不能
   自己去调模型、查业务库，界面让人点「重试」接着做。
 
 进程内登记、进程内的队：服务端是单进程部署，多进程部署时要换成库里的锁。
@@ -243,9 +244,16 @@ lanes = SourceLanes()
 # ==========================================================================
 
 
+#: 收尾时在同一个事务里补写的那一步（close_job、cancel_queued 的 finish）：表和任务的状态已经改好、还没提交
+Finish = Callable[[AsyncSession], Awaitable[None]]
+
+
 async def close_job(session: AsyncSession, job_model: Any, task_model: Any, job_id: str, job_status: str,
-                    job_error: str | None, task_status: str, task_error: str | None) -> None:
-    """收尾：还没结束的表记 task_status（原因 task_error），任务（还没结束的话）记 job_status。提交。"""
+                    job_error: str | None, task_status: str, task_error: str | None, *,
+                    finish: Finish | None = None) -> None:
+    """收尾：还没结束的表记 task_status（原因 task_error），任务（还没结束的话）记 job_status。提交。
+
+    finish：提交前在同一个事务里再做的一步（评测按改好的逐项状态重算计数、记停下的原因）。起草、剖析、探查不传，照旧。"""
     now = utcnow()
     await session.execute(
         update(task_model).where(task_model.job_id == job_id, task_model.status.in_(OPEN))
@@ -253,12 +261,15 @@ async def close_job(session: AsyncSession, job_model: Any, task_model: Any, job_
     await session.execute(
         update(job_model).where(job_model.id == job_id, job_model.status.in_(ACTIVE))
         .values(status=job_status, error=job_error, finished_at=now).execution_options(synchronize_session=False))
+    if finish is not None:
+        await finish(session)
     await session.commit()
 
 
-async def cancel_queued(session: AsyncSession, job_model: Any, task_model: Any, job_id: str) -> bool:
+async def cancel_queued(session: AsyncSession, job_model: Any, task_model: Any, job_id: str, *,
+                        finish: Finish | None = None) -> bool:
     """停止一个排队中的任务：任务和没结束的表直接记 cancelled。已经被后台认领（开始了）返回 False，交给调用方按进行中
-    的停。"""
+    的停。finish 同 close_job（只在真停了时做）。"""
     now = utcnow()
     claimed = await session.execute(
         update(job_model).where(job_model.id == job_id, job_model.status == "queued")
@@ -269,6 +280,8 @@ async def cancel_queued(session: AsyncSession, job_model: Any, task_model: Any,
     await session.execute(
         update(task_model).where(task_model.job_id == job_id, task_model.status.in_(OPEN))
         .values(status="cancelled", finished_at=now).execution_options(synchronize_session=False))
+    if finish is not None:
+        await finish(session)
     await session.commit()
     return True
 
diff --git a/backend/app/data/catalog_profile.py b/backend/app/data/catalog_profile.py
index ed63b2ed..7ea86279 100644
--- a/backend/app/data/catalog_profile.py
+++ b/backend/app/data/catalog_profile.py
@@ -45,7 +45,7 @@ import re
 import time
 from collections.abc import Awaitable, Callable
 from dataclasses import asdict, dataclass, field
-from datetime import date as _date, datetime, timezone
+from datetime import date as _date, datetime, timedelta, timezone
 from decimal import Decimal
 from typing import Any, NamedTuple
 
@@ -493,17 +493,23 @@ class SqlDialect:
         hi = self.date_literal(VALID_DATE_BEFORE, text=text)
         return f"SELECT MAX({column}) FROM {table} WHERE {column} >= {lo} AND {column} < {hi}"
 
-    def _window(self, column: str, since: Any, text: bool) -> str:
-        """时间窗口的条件：since 起（含），并且早于占位日期 9999-12-31——PostgreSQL 的 infinity、Oracle 的 9999-12-31
-        都大于任何一天，不截掉上端就会在「最近 35 天」里多出一天。"""
+    def _window(self, column: str, since: Any, text: bool, before: Any = None) -> str:
+        """时间窗口的条件：since 起（含），并且早于 before（不含；缺省是占位日期 9999-12-31——PostgreSQL 的 infinity、
+        Oracle 的 9999-12-31 都大于任何一天，不截掉上端就会在「最近 35 天」里多出一天）。"""
         return (f"{column} >= {self.date_literal(since, text=text)} "
-                f"AND {column} < {self.date_literal(VALID_DATE_BEFORE, text=text)}")
+                f"AND {column} < {self.date_literal(before or VALID_DATE_BEFORE, text=text)}")
 
-    def day_volume_sql(self, table: str, column: str, since: Any, *, text: bool = False) -> str:
-        """最近若干天按天计数：since 起（含）每天几行。只在日期列打头有索引时对大表发（走索引的范围扫描）。"""
+    def day_volume_sql(self, table: str, column: str, since: Any, until: Any, *, text: bool = False) -> str:
+        """最近若干天按天计数：since 到 until（都含；until 是观察当天）每天几行。只在日期列打头有索引时对大表发（走索引的
+        范围扫描）。上界是 until 的第二天（不含）：业务日期有未来值（预约、计划类的日期，或者写错的）时，今天以后的日子
+        不进窗口——它们不是「最近的整天」，也不该被当成缺天（data/value_drift）。时间列上「< 第二天」把 until 当天的
+        时分秒都算进来；文字存的日期按字符串比，同样成立。"""
+        last = until if isinstance(until, _date) else _date.fromisoformat(str(until))
+        if isinstance(last, datetime):
+            last = last.date()
         day = self.day_trunc(column)
         return (f"SELECT {day} AS d, COUNT(*) AS n FROM {table} "
-                f"WHERE {self._window(column, since, text)} GROUP BY {day} ORDER BY d")
+                f"WHERE {self._window(column, since, text, last + timedelta(days=1))} GROUP BY {day} ORDER BY d")
 
     def window_value_counts_sql(self, table: str, column: str, date_column: str, since: Any, n: int, *,
                                 text: bool = False) -> str:
diff --git a/backend/app/data/catalog_profile_jobs.py b/backend/app/data/catalog_profile_jobs.py
index e24d2c19..37a4b097 100644
--- a/backend/app/data/catalog_profile_jobs.py
+++ b/backend/app/data/catalog_profile_jobs.py
@@ -692,6 +692,18 @@ async def _behind_draft(session: AsyncSession, source_id: str) -> bool:
     return catalog_jobs.sync_drafting(source_id) or await catalog_jobs.active_job(session, source_id) is not None
 
 
+async def nothing_reason(session: AsyncSession, row: Any, source: Any, scope: str) -> str | None:
+    """按此刻的表结构和目录，这个范围一张表都挑不出来时为什么（ProfileNothing 的原话）；挑得出、或者排在起草后面（开始时按
+    起草后的目录再挑，现在说不准）返回 None。定时维护建出 0 张表的任务时用它把原因记进触发记录。"""
+    if await _behind_draft(session, row.id):
+        return None
+    try:
+        await _pick(session, row, source, tables=None, scope=scope)
+    except ProfileNothing as e:
+        return str(e)
+    return None
+
+
 async def create_job(session: AsyncSession, row: Any, source: Any, *, tables: list[str] | None, scope: str,
                      settings: ProfileSettings, actor: str | None, retry_of: str | None = None,
                      trigger: str = "manual", fire_at: str | None = None, schedule_id: str | None = None,
@@ -1161,5 +1173,6 @@ __all__ = ["INTERRUPTED_ERROR", "KEEP_JOBS", "MAX_TABLES", "NOTHING_AT_ALL", "NO
            "NOTHING_TO_OBSERVE", "NothingToRetry", "OBSERVING_SCOPES", "SCOPES", "task_modes", "trigger_of",
            "ProfileBusy", "ProfileFinished", "ProfileJobs", "ProfileNothing", "THEN_PROFILE", "active_job",
            "busy_message", "cancel", "create_job", "due_reason", "ensure_free", "job_brief", "job_dict", "job_out",
-           "last_detail", "last_facts", "last_profiles", "outside_of", "latest_job", "load", "load_task", "manager", "prune", "public_last",
-           "retry_tables", "startup", "task_detail", "task_row", "task_rows", "unsaved_error", "waiting_for"]
+           "last_detail", "last_facts", "last_profiles", "outside_of", "latest_job", "load", "load_task", "manager",
+           "nothing_reason", "prune", "public_last", "retry_tables", "startup", "task_detail", "task_row", "task_rows",
+           "unsaved_error", "waiting_for"]
diff --git a/backend/app/data/catalog_profile_observe.py b/backend/app/data/catalog_profile_observe.py
index 78b07a31..0663312e 100644
--- a/backend/app/data/catalog_profile_observe.py
+++ b/backend/app/data/catalog_profile_observe.py
@@ -8,7 +8,11 @@
 - rows：这张表的行数（剖析开头估行数那一条的结果，不另发查询）；口径 stats（数据库统计的估算值）/ count（数到上限）；
 - codes_observe：码值列的取值分布——**所有**码值列都观察，包括人工确认过、标了已列全的（目录-D6）；只读，不改含义；
 - freshness：业务日期列（已确认或已验证）的最新日期；
-- daily_volume：业务日期最近 35 天每天的行数。
+- daily_volume：业务日期最近 35 天每天的行数，数到观察当天为止（今天以后的日子不进窗口，见 SqlDialect.day_volume_sql）。
+
+**「今天」按服务端所在时区**（local_today，和待办汇总的 notify.digest_at 同一个口径）：窗口从哪天起、数到哪天、比较器的
+「今天」（最近一个整天是哪天、断档离今天几天）都用它。业务日期是本地日期，按 UTC 的日期算，东八区凌晨 0–8 点的「今天」
+还是前一天，这几样都会差一天。服务端和业务库不在一个时区时，差的那几个小时照实差着（业务库的时区探查不到）。
 
 **大表的口径**（目录-D7）：小表（行数不超过整表统计的上限）看整表（full）。大表：业务日期已确认或已验证、而且这一列打头
 有索引（探查记下的 index_first_columns）时看最近 35 天的窗口（window，走索引的范围扫描）；PostgreSQL / Oracle 没有就块采样
@@ -139,12 +143,13 @@ def sample_percent(rows: int | None, target: int) -> float | None:
     return next((p for p in SAMPLE_LADDER if p <= want), SAMPLE_LADDER[-1])
 
 
-def _today() -> date:
-    return datetime.now(timezone.utc).date()
+def local_today(now: datetime | None = None) -> date:
+    """观察用的「今天」：服务端所在时区的日期（和待办汇总的 notify.digest_at 同一个口径，见模块说明）。now 缺省是现在。"""
+    return (now or utcnow()).astimezone().date()
 
 
 def _since() -> date:
-    return _today() - timedelta(days=WINDOW_DAYS)
+    return local_today() - timedelta(days=WINDOW_DAYS)
 
 
 def _window_ok(cx: Any, meta: Mapping[str, Any], day: str | None) -> bool:
@@ -271,8 +276,9 @@ async def _dates(cx: Any, table: str, kind: str, column: str, meta: Mapping[str,
         result.observations.append(_observation("freshness", column, {"method": "full"},
                                                 {"max": str(value)[:19]}, []))
         return
-    since = _since()
-    res = await cx.run(d.day_volume_sql(t, c, since, text=text), max_rows=WINDOW_DAYS + 2)
+    today = local_today()
+    since = today - timedelta(days=WINDOW_DAYS)
+    res = await cx.run(d.day_volume_sql(t, c, since, today, text=text), max_rows=WINDOW_DAYS + 2)
     days: dict[str, int] = {}
     for row in res.rows:
         day = str(row[0])[:10] if row and row[0] is not None else ""
@@ -280,7 +286,7 @@ async def _dates(cx: Any, table: str, kind: str, column: str, meta: Mapping[str,
             days[day] = days.get(day, 0) + (cp._as_int(row[1]) or 0)
     basis = {"method": "window", "column": column, "days": WINDOW_DAYS, "since": since.isoformat()}
     result.observations.append(_observation("daily_volume", column, basis,
-                                            {"since": since.isoformat(), "until": _today().isoformat(),
+                                            {"since": since.isoformat(), "until": today.isoformat(),
                                              "days": days}, []))
 
 
@@ -396,7 +402,7 @@ async def _save_one(session: AsyncSession, source: Any, table: str, obs: Mapping
     current = vd.Obs(kind, obs["basis_key"], obs["data"], now)
     past = [vd.Obs(h.kind, h.basis_key, h.data or {}, h.at) for h in history]
     fresh = 0
-    for f in vd.compare(current, past, today=now.date(), known=known, complete=catalog.codes_complete(item)):
+    for f in vd.compare(current, past, today=local_today(now), known=known, complete=catalog.codes_complete(item)):
         after = dict(f.after) if isinstance(f.after, Mapping) else {"value": f.after}
         if f.revoke_complete:
             version = await _revoke_complete(session, source.id, table, column)
diff --git a/backend/app/data/catalog_review.py b/backend/app/data/catalog_review.py
index 43e27594..78eb3438 100644
--- a/backend/app/data/catalog_review.py
+++ b/backend/app/data/catalog_review.py
@@ -41,7 +41,7 @@ from typing import Any
 from sqlalchemy import func, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from app.data import catalog
+from app.data import catalog, source_objects
 from app.db.models import Run
 
 #: 队列的筛选：全部、影响数字的项、名称和说明
@@ -346,7 +346,7 @@ async def review_batch(session: AsyncSession, source: Any, items: Iterable[Batch
         key = known[item.table_name]
         if key is None:
             results[i] = _result(item, ok=False, code=CODE_TABLE_MISSING,
-                                 error=f"数据源中没有表 {item.table_name}，可能已被删除或尚未探查结构")
+                                 error=await source_objects.absent_message(session, source, item.table_name))
             continue
         groups.setdefault(key, []).append(i)
     at = catalog.now_iso()
diff --git a/backend/app/data/source_maintenance.py b/backend/app/data/source_maintenance.py
index a2bba31e..aa8597bf 100644
--- a/backend/app/data/source_maintenance.py
+++ b/backend/app/data/source_maintenance.py
@@ -13,7 +13,10 @@ observe 取值观察，可多选，没设过是三步都做。
 
 **到点**（fire，持着 scheduler 的 launch_lock）：按顺序建探查任务 → 剖析任务（范围 needed / observe / maintenance，按选的
 步骤），任务上记 trigger=schedule 和 fire_at，靠 SourceLanes 串行：剖析排在同一次触发的探查后面（等探查进了同源队才建剖析
-任务），开始时按那时的表结构和目录挑表（挑出 0 张也照建，任务以 0 张表结束、提醒里写明为什么）。预算、止损照剖析设置。
+任务），开始时按那时的表结构和目录挑表（挑出 0 张也照建，任务以 0 张表结束、提醒里写明为什么）。建任务时就判断得出会是 0 张
+的（这次没有探查、也没排在起草后面，开始时表结构和目录都不会变——比如有数据目录的表都在 12 小时内观察过），原因另记进触发
+记录的 inputs.skipped 和 reason；同一次触发里有探查、或者排在起草后面时说不准（探查完、起草完可能就有了），保持现状，只在
+任务的提醒里写。预算、止损照剖析设置。
 照实记没执行的：数据源不存在、已停用、是上传表格或记录表时整次不执行（failed_to_start）；剖析没开时剖析、观察两步不执行、
 探查照做；上一次的维护（或别的探查、剖析）还没做完时这次跳过（skipped_overlap）。每次触发记下建了哪几个任务、哪几步没执行、
 为什么（触发记录的 inputs：{steps, jobs, skipped}）。
@@ -221,6 +224,11 @@ async def _start(source_id: str, steps: list[str], *, actor: str, fire_at: str,
                 table_versions.SnapshotMissing) as e:
             skipped += [{"step": s, "reason": str(e)} for s in wanted]
             return jobs, skipped
+        if not tasks and introspect_id is None and (why := await catalog_profile_jobs.nothing_reason(
+                session, row, view, scope)):
+            # 建任务时就挑不出表、这次又没有探查（开始时表结构不会变）、也没排在起草后面（目录不会变）：任务会以 0 张表
+            # 结束，原因照实记进触发记录，不只在剖析任务的提醒里。任务照建（触发照旧是 started）
+            skipped += [{"step": s, "reason": why} for s in wanted]
         catalog_profile_jobs.manager.start(job.id)
         jobs.append({"step": "+".join(wanted), "kind": "profile", "id": job.id, "scope": scope, "tables": len(tasks)})
     return jobs, skipped
diff --git a/backend/app/data/source_objects.py b/backend/app/data/source_objects.py
index 058fc989..b938a7d4 100644
--- a/backend/app/data/source_objects.py
+++ b/backend/app/data/source_objects.py
@@ -825,6 +825,33 @@ async def absent_reasons(session: AsyncSession, row: Any, names: Iterable[str])
     return out
 
 
+#: 表结构里没有的表给人看的一句（接口的 404、批量审阅逐条的错误）：和界面的 CATALOG_ABSENT_TEXT、/schema 的说法
+#: （schema_view.absent_detail，给模型看）同一套「不在」。{table} 是点名的表
+ABSENT_MESSAGE = {
+    ABSENT_REMOVED: "数据源中已经没有表 {table}（已删除或改名）",
+    ABSENT_MISSING: "数据源中没有表 {table}（全库的对象清单里也没有）",
+    ABSENT_OUT_OF_SCOPE: ("表 {table} 在探查范围外：对象索引里有它，但没有读取结构。在数据源卡片的「探查范围」里加入它，"
+                          "或在数据目录的搜索框搜索它并「加入范围」"),
+    ABSENT_REFLECT_FAILED: "表 {table} 在探查范围内，但上次探查读不到它的结构（权限不够或对象有问题），修好后重新探查即可",
+    ABSENT_UNKNOWN: ("表 {table} 不在已探查的表里：升级前的探查只取了一部分表，无法判断它是否还在，请回到数据源卡片重新"
+                     "「探查结构」"),
+}
+#: 说不出是哪一种「不在」时（上传表格、记录表按版本走；没有对象索引、又没截断过的老缓存认不出名字是不是从来就没有）
+ABSENT_PLAIN = "数据源中没有表 {table}"
+
+
+async def absent_message(session: AsyncSession, row: Any, table: str) -> str:
+    """点名的表不在表结构里时给人看的一句（ABSENT_MESSAGE）。手工接入的源按 absent_reasons 说是哪一种；上传表格、记录表
+    的结构随版本走，只说没有（目录清单把它们记成 removed，是说「那一版里没有」，点名一个从来没有过的名字时说不得「已删除」）；
+    没有对象索引的老缓存同理：absent_reasons 对没截断的一律记 removed，名字写错了也会这样，只说没有。"""
+    if (getattr(row, "origin", None) or "manual") != "manual":
+        return ABSENT_PLAIN.format(table=table)
+    reason = (await absent_reasons(session, row, [table])).get(table)
+    if reason == ABSENT_REMOVED and not await has_index(session, row.id, index_schema(row)):
+        reason = None
+    return ABSENT_MESSAGE.get(reason or "", ABSENT_PLAIN).format(table=table)
+
+
 @dataclass(frozen=True)
 class OutsideObject:
     """对象索引里在、范围外的一个对象（助手检索、目录搜索用）。"""
@@ -1039,7 +1066,8 @@ async def delete_index(session: AsyncSession, source_id: str) -> None:
 
 
 __all__ = [
-    "ABSENT_MISSING", "ABSENT_OUT_OF_SCOPE", "ABSENT_REFLECT_FAILED", "ABSENT_REMOVED", "ABSENT_UNKNOWN", "ADD_LIMIT", "AddOutcome",
+    "ABSENT_MESSAGE", "ABSENT_MISSING", "ABSENT_OUT_OF_SCOPE", "ABSENT_PLAIN", "ABSENT_REFLECT_FAILED",
+    "ABSENT_REMOVED", "ABSENT_UNKNOWN", "ADD_LIMIT", "AddOutcome", "absent_message",
     "INDEX_LIMIT", "IndexResult", "IndexedObject", "OutsideObject", "REFLECT_BATCH", "RULES_MAX", "Reflected",
     "RefreshStats", "SCOPE_DEFAULT", "SCOPE_MAX", "SCOPE_OPTION", "SCOPE_REASONS", "ScopePlan", "ScopeRules", "Trail",
     "VIA_CATALOG", "VIA_LOOK", "VIA_SCHEMA_TOOL", "absent_reasons", "add_to_scope", "apply_scope",
diff --git a/backend/app/data/sql_semantics.py b/backend/app/data/sql_semantics.py
index 1ab1848e..ffb7ae58 100644
--- a/backend/app/data/sql_semantics.py
+++ b/backend/app/data/sql_semantics.py
@@ -111,7 +111,7 @@ PIPES_AS_CONCAT 的库这两处的意思不同，sqlglot 的 MySQL 读法和这
 **具体说法（见 _Explainer）。** 只在两边的规范形式都是「简单形状」时说（_simple_block：整条语句就是一段 SELECT，没有
 CTE、子查询、UNION 这类合并、窗口函数、DISTINCT ON；FROM 里只有实表，连接只有内连接、CROSS JOIN、逗号和 LEFT JOIN；
 LIMIT 是至少为 1 的常数；没有认不准的名字、SELECT 列表里没有 sqlglot 不认识的函数；不认列的那一级里原文没有在
-WHERE / GROUP BY 里用列别名的），而且只说四类：
+WHERE / GROUP BY 里用列别名的），而且只说五类：
 
 - 时间列（time）：两边各在一张底层表的一列上按时间筛、列不同（另一边哪儿都没用这一列筛选，两边的时间条件都不是恒为
   真的，时间列不在 LEFT JOIN 的可空一侧）；一边按时间筛了、另一边哪儿都没有可能是时间窗口的条件的（见下），也说成
@@ -129,6 +129,13 @@ WHERE / GROUP BY 里用列别名的），而且只说四类：
   不用到 LEFT JOIN 可空一侧的列（WHERE r.id IS NULL 是反连接，说的是「没关联上」；r.order_id IS NULL 和 ON 里的
   r.order_id = o.id 一起已经蕴含 r.id IS NULL）；不是可能的时间窗口（_window_like：date_key BETWEEN 20260901 AND
   20260930、stat_month = 202609——认不准是不是时间列，说成有效行过滤会被写进数据目录）。detail 给条件涉及的实表列。
+- 同一列上的条件改写（filter / time_range，a、b 两边都有）：两边 WHERE 里对同一张实表的同一列各写了一套条件、留下的
+  取值不同——等号对 IN（口径 status = 'paid'，这次 status IN ('paid', 'refunded')）、取值集合不同、范围的端点不同，说
+  「A 对 status 的条件是 …，B 是 …」，两边都只是取值集合、又有共同的取值时补「（B 多了 'refunded'）」。要两边用到这一列
+  的条件都只涉及这一列、比的就是那一列本身，按取值逐段算过两边留下的取值确实不同（_Range.compare：表结构里确定不为空
+  的列不拿空值比；x >= 100 AND x >= 50 对 x >= 100 这类冗余写法算出来一样，不说），这一列不在 LEFT JOIN 的可空一侧，
+  两边在 WHERE 之外（ON、HAVING、CASE / FILTER、聚合的参数）都没用到这一列，不是认不准的时间窗口，没换表。两边都是
+  时间条件时 aspect 是 time_range。只改说明：走到这里时已经判了不等价（见 _Explainer._rewrites）。
 - 聚合与去重（aggregate / group_by）：选出来的列只差在顶层的聚合函数上（同一列做参数，SUM 对 AVG；COUNT(*) 对 SUM(x)）
   或者去重上（COUNT / SUM / AVG 去不去重、COUNT(*) 对 COUNT(DISTINCT x)、SELECT DISTINCT 对不去重），分组一样。参数
   得是光一列、在每组里不是定死的（不被 WHERE 定成常量、不等于分组键、表结构里它那张表的键没有在组里定死——按 u.id 分组
@@ -144,7 +151,7 @@ WHERE / GROUP BY 里用列别名的），而且只说四类：
 每一句都先确认这处不同真会影响结果：两边的 WHERE、HAVING、ON 里有可能恒为假的条件（常量条件，同一个对象上取不到值的
 几个条件）时什么都不说；结果看不见行的（SELECT COUNT(*) > -1 这类只出一行、值又和行无关的）不说筛选、关联上的不同；
 表结构里的约束（主键、唯一约束、外键、NOT NULL）说明不起作用的不说——约束只拿来少说，不拿来判等价。被 WHERE 用常量定死
-的列出现在排序、分组里的差异，恒为真的 HAVING，都不在四类里，不说。两边给人看的写法一样、和已经说过的一句正好反过来的
+的列出现在排序、分组里的差异，恒为真的 HAVING，都不在五类里，不说。两边给人看的写法一样、和已经说过的一句正好反过来的
 不说。
 
 **兜底。** 规范形式不一致、一条具体说法都没有时，给一条 aspect="other" 的「两条 SQL 的写法不同，未能归纳成具体差异，
@@ -164,8 +171,9 @@ WHERE / GROUP BY 里用列别名的），而且只说四类：
   关联先后、按 WHERE 把外连接改成内连接、换成相等的列；没给表结构（或者类型不一样）时按关联条件相等的两列换着分组、
   排序；PostgreSQL、MySQL 没给表结构时去重、分组、求和的一段调换关联先后、并子查询、去掉没用到的 LEFT JOIN、把只取
   可空一侧聚合的 LEFT JOIN 写成内连接。
-- 只给兜底的：不是简单形状的（CTE、子查询、UNION、窗口函数，规范化没把它们并掉的）；四类以外的不同（同一列上的时间
-  范围、同一列上换了筛选的值、关联条件、分组、HAVING、排序、取几行、选出来的列、聚合的参数）；几张表、几列交错的条件；
+- 只给兜底的：不是简单形状的（CTE、子查询、UNION、窗口函数，规范化没把它们并掉的）；五类以外的不同（算出来的值上的
+  条件改写——LOWER(status) = 'paid' 对 LOWER(status) IN (…)、和别的列交错的同一列条件、关联条件、分组、HAVING、排序、
+  取几行、选出来的列、聚合的参数）；几张表、几列交错的条件；
   自己关联自己、同名的表关联了几次的；换了表又有关联条件不同的；条件写在聚合里的一边对条件写在 WHERE 里的一边（SUM(status
   = 'paid') 对 COUNT(*) … WHERE status = 'paid'：真正的不同是计数为 0 的组、空集时是 0 还是 NULL，说不清）；反连接；
   可能是时间窗口、又认不准是不是时间列的条件（业务日期键、统计月份、分区，第几天这类范围）。
@@ -5902,8 +5910,9 @@ class FacetDiff:
     set_op（几支查询怎么合并）/ group_by（分组键、去重）/ having / limit（取哪几行，含取行前的排序）/ order（只是最终
     结果的排序不同，不影响取到哪些行）/ parse（解析不了）/ other（兜底：未能归纳成具体差异，见 explain_sql_diff）。a、b
     是两边各自的写法（给人看，没有就是 None），text 是一句中文说明。detail 只在 time 上有：两边的时间列
-    [{table, column}]（各是各的表），调用方据此对照数据目录（业务日期）；在 filter 上有 {"columns": [{table, column}]}：
-    条件只涉及的实表列，对照有效行过滤用。"""
+    [{table, column}]（各是各的表），调用方据此对照数据目录（业务日期）；在 filter、time_range 上有 {"columns":
+    [{table, column}]}：条件只涉及的实表列，对照有效行过滤用。filter 一边是 None 的是「一边多出来的条件」；两边都有的是
+    同一列上的条件两边各写了一套（「A 对 X 的条件是 …，B 是 …」），不是一边多出来的有效行条件。"""
 
     aspect: str
     a: Any
@@ -6360,8 +6369,10 @@ class _Range:
             subject = n.this
         return self._truth(n, assign[_fingerprint(subject)])
 
-    def compare(self, conds_a: list[Any], conds_b: list[Any], grouped: bool = False) -> bool | None:
-        """两组条件（各自 AND 连，可以一边为空 = 不筛）留下的取值是不是一样：一样 True，不一样 False，认不出 None。"""
+    def compare(self, conds_a: list[Any], conds_b: list[Any], grouped: bool = False,
+                never_null: bool = False) -> bool | None:
+        """两组条件（各自 AND 连，可以一边为空 = 不筛）留下的取值是不是一样：一样 True，不一样 False，认不出 None。
+        never_null：对象一定不为空（表结构里不为空的列），不拿空值比。"""
         parsed = [self.subject(c) for c in conds_a + conds_b]
         if any(p is None for p in parsed) or len({p[0] for p in parsed}) != 1:
             return None
@@ -6373,7 +6384,7 @@ class _Range:
         from sqlglot import exp
 
         values = self._pieces(subject, literals, ranged, grouped)
-        if not isinstance(_unparen(subject), exp.Count):
+        if not isinstance(_unparen(subject), exp.Count) and not never_null:
             values.append(None)
         for v in values:
             ta = all(self._truth(c, v) == "T" for c in conds_a)
@@ -6745,7 +6756,7 @@ def _alias_named(root: Any) -> bool:
 class _Explainer:
     """在两边的规范形式上找有把握、又确实影响结果的不同，写成具体的中文句子，同时记下是不是所有不同都说到了。
 
-    只在两边都是简单形状时说（见 _simple_block），也只说下面四类；别的不同（同一列上的时间范围、同一列上换了筛选的值、
+    只在两边都是简单形状时说（见 _simple_block），也只说下面五类；别的不同（算出来的值上换了筛选的值、
     关联条件、分组、HAVING、排序、取几行、选出来的列、聚合的参数……）一律算没说到，交给兜底说明：
 
     - time：底层表上按哪一列筛时间不同（两边各在一列上按时间筛，列不同，另一边哪儿都没用这一列筛选，时间列不在 LEFT
@@ -6757,6 +6768,8 @@ class _Explainer:
       那一列（LOWER(x) <> 'PAID' 这类认不出取值范围、可能恒为真）；按取值逐段算过不是恒为真的（见 _Range.always，一定
       不为空的列不算空值，见 _never_null）；不用到 LEFT JOIN 可空一侧的列（反连接，见 _on_nullable_side）；不是认不准
       的时间窗口（见 _window_like）。
+    - filter / time_range（a、b 两边都有）：同一列上的条件两边各写了一套、留下的取值确实不同（等号对 IN、取值集合不同、
+      范围的端点不同），说「A 对 X 的条件是 …，B 是 …」，见 _rewrites。
     - aggregate / group_by：选出来的列只差在顶层的聚合函数（同一列做参数，SUM 对 AVG；COUNT(*) 对 SUM(x)）或者去重上
       （COUNT / SUM / AVG 去不去重，COUNT(*) 对 COUNT(DISTINCT x)，SELECT DISTINCT 对不去重），分组一样。参数在组里
       定死的不说（被 WHERE / ON 定成常量、等于分组键、本身是常量、它那张表的键在组里定死了，见 _fixed_in_group）；去重
@@ -7312,8 +7325,8 @@ class _Explainer:
 
     def conditions(self, A: _Block, B: _Block, mapping: Mapping[str, str], skip_a: set[str],
                    skip_b: set[str]) -> bool:
-        """WHERE：两边没配上的条件里说得准的两类——按哪一列筛时间不同（_times）、一边多出来的筛选条件（_extra_filters）；
-        剩下的算没说到。"""
+        """WHERE：两边没配上的条件里说得准的三类——按哪一列筛时间不同（_times）、同一列上的条件两边各写了一套（_rewrites）、
+        一边多出来的筛选条件（_extra_filters）；剩下的算没说到。"""
         xa = [c for c in A.where if self.fp(c) not in skip_a]
         xb = [c for c in B.where if self.fp(c) not in skip_b]
         only_a, only_b = self._unmatched(xa, xb, mapping)
@@ -7322,6 +7335,7 @@ class _Explainer:
         if not self.visible:
             return False
         only_a, only_b, covered = self._times(A, B, only_a, only_b)
+        only_a, only_b = self._rewrites(A, B, xa, xb, only_a, only_b, mapping)
         covered &= self._extra_filters(A, B, xa, only_a, 0)
         covered &= self._extra_filters(B, A, xb, only_b, 1)
         return covered
@@ -7361,6 +7375,101 @@ class _Explainer:
         return [c for c in only_a if all(c is not x for x in conds_a)], \
             [c for c in only_b if all(c is not x for x in conds_b)], True
 
+    def _rewrites(self, A: _Block, B: _Block, xa: list[Any], xb: list[Any], only_a: list[Any], only_b: list[Any],
+                  mapping: Mapping[str, str]) -> tuple[list[Any], list[Any]]:
+        """同一列上的条件两边各写了一套、留下的取值确实不同：说「A 对 X 的条件是 …，B 是 …」——等号对 IN（口径
+        status = 'paid'，这次 status IN ('paid', 'refunded')）、取值集合不同、范围的端点不同。两边都只是「等于常量 / 在
+        常量列表里」、又有共同的取值时补一句各多了哪几个取值。说到了的从两边拿掉，说不准的原样还回去（交给
+        _extra_filters，多半是兜底）。
+
+        说得准要（每一条都是为了不说出不存在的不同）：两边对这一列都有条件、至少一边有没配上的；两边 WHERE 里用到这一列的
+        条件都只涉及这一列（不和别的列交错）、比的就是那一列本身（不是 LOWER(x)），按取值逐段算过两边留下的取值不一样
+        （_Range.compare：表结构里确定不为空的列不拿空值比；认不出、算出来一样的不说——x >= 100 AND x >= 50 对 x >= 100
+        只是冗余的写法）；这一列不在 LEFT JOIN 的可空一侧（那是关联上的事）；两边在 WHERE 之外（ON、HAVING、CASE / FILTER、
+        条件里套着的聚合参数）都没用到这一列（SUM(CASE WHEN status = 'paid' …) 配上 WHERE status IN (…) 可能和只筛
+        'paid' 一样）；不是认不准的时间窗口（_window_like：date_key BETWEEN 20260901 AND 20260930）；没换表。两边的条件
+        都是时间条件（timed）时 aspect 是 time_range，否则是 filter（a、b 两边都有，dual_check_facts 不会把它当成「一边
+        多出来的有效行条件」写进数据目录）。
+
+        只改说明，不碰等价判定：走到这里时两边已经判了不等价（见 explain_sql_diff）。"""
+        if self.swapped or self.usage_a is None or self.usage_b is None:
+            return only_a, only_b
+        refs_a = [self.refs(c, A) for c in xa]
+        refs_b = [self.refs(c, B) for c in xb]
+        if any(r is None for r in refs_a + refs_b):
+            # 有认不出用到哪些列的条件：它可能也用到了这一列，说「A 对 X 的条件是 …」就不全
+            return only_a, only_b
+        keys: list[tuple[str, str]] = []
+        for c, block in [(c, A) for c in only_a] + [(c, B) for c in only_b]:
+            found = self.refs(c, block)
+            if found is not None and len(found) == 1 and next(iter(found)) not in keys:
+                keys.append(next(iter(found)))
+        for key in keys:
+            on_a = [c for c, r in zip(xa, refs_a) if key in r]
+            on_b = [c for c, r in zip(xb, refs_b) if key in r]
+            pending_a = [c for c in only_a if any(c is x for x in on_a)]
+            pending_b = [c for c in only_b if any(c is x for x in on_b)]
+            if not on_a or not on_b or not (pending_a or pending_b):
+                continue
+            if any(r != {key} for r in refs_a if r & {key}) or any(r != {key} for r in refs_b if r & {key}):
+                continue
+            if self._on_nullable_side(on_a, A) or self._on_nullable_side(on_b, B):
+                continue
+            mapped_b = [self.mapped(c, mapping) for c in on_b]
+            if not self._plain_subjects(on_a + mapped_b):
+                continue
+            times = {self.timed(c, A) for c in on_a} | {self.timed(c, B) for c in on_b}
+            if times != {key} and any(_window_like(c) for c in on_a + on_b):
+                continue
+            kinds = _kinds_for(on_a + on_b)
+            if self.usage_a.touches({key}, skip=on_a, kinds=kinds) or \
+                    self.usage_b.touches({key}, skip=on_b, kinds=kinds):
+                continue
+            never_null = _not_null_column(self.info, key[0], key[1], self.read)
+            if self.range.compare(on_a, mapped_b, never_null=never_null) is not False:
+                continue
+            shown_a, shown_b = self.show_all(_by_bound(on_a), A), self.show_all(_by_bound(on_b), B)
+            text = _say(self.la, "对", self.named(*key, qualified=len(A.all_rels()) > 1), "的条件是", shown_a, "，",
+                        self.lb, "是", shown_b)
+            more = self._members(on_a), self._members(mapped_b)
+            # 两边都只是取值集合、又有共同的取值时补一句各多了哪几个（两边的取值毫不相干时前一句已经说全了）
+            if more[0] is not None and more[1] is not None and set(more[0]) & set(more[1]):
+                extra_a = [shown for value, shown in more[0].items() if value not in more[1]]
+                extra_b = [shown for value, shown in more[1].items() if value not in more[0]]
+                parts = [_say(label, "多了", "、".join(extra)) for label, extra in ((self.la, extra_a),
+                                                                                  (self.lb, extra_b)) if extra]
+                if parts:
+                    text = _say(text, "（", "，".join(parts), "）")
+            self.add("time_range" if times == {key} else "filter", shown_a, shown_b, text,
+                     {"columns": [{"table": key[0], "column": key[1]}]})
+            only_a = [c for c in only_a if all(c is not x for x in on_a)]
+            only_b = [c for c in only_b if all(c is not x for x in on_b)]
+        return only_a, only_b
+
+    def _members(self, conds: list[Any]) -> dict[Any, str] | None:
+        """几个条件（AND 连）都是「一列等于常量 / 在常量列表里」时，留下的取值 {比较用的值（见 _Range._norm）: 给人看的
+        写法}；有别的写法（比大小、不等于、NOT、OR、列表里的 NULL）返回 None。"""
+        from sqlglot import exp
+
+        kept: dict[Any, str] | None = None
+        for cond in conds:
+            cond = _unparen(cond)
+            if isinstance(cond, exp.EQ):
+                items = [cond.expression] if _literal_value(cond.expression) is not None else [cond.this]
+            elif isinstance(cond, exp.In) and cond.expressions and not cond.args.get("query") and \
+                    not cond.args.get("unnest") and not cond.args.get("field"):
+                items = list(cond.expressions)
+            else:
+                return None
+            values: dict[Any, str] = {}
+            for item in items:
+                value = _literal_value(item)
+                if value is None:
+                    return None
+                values.setdefault(self.range._norm(value), self.sql(item))
+            kept = values if kept is None else {v: shown for v, shown in kept.items() if v in values}
+        return kept
+
     def _extra_filters(self, block: _Block, other: _Block, where: list[Any], rest: list[Any], side: int) -> bool:
         """block 这一边多出来的筛选条件（rest 是它没配上的 WHERE 项）。按用到的列连成一组（x > 5 和 x IS NOT NULL
         是一组）一组一组看，说得准的说「A 只保留 … 的行，B 没有这个条件」「A 排除了 … 的行，B 没有」「A 按 X 筛时间，
@@ -7943,9 +8052,9 @@ def explain_sql_diff(a: str, b: str, dialect: str | None = None, *, schema: Mapp
     一元加号、0x10、换了名字的类型名，不知道是哪种库时套着的块注释、紧跟着字的 --、双引号括起来的名字……，见
     _lost_in_parsing、_misread）不在语法树上判等价，只给兜底和原文的逐行对照。
 
-    不等价时怎么说：两边的规范形式都是简单形状时（一段 SELECT、实表、普通的连接，见 _simple_block），只把四类有把握、
-    真会影响结果的不同写成具体的中文句子（按哪一列筛时间、一边多出来的筛选条件、顶层的聚合函数和去重、关联的表和关联
-    类型，见 _Explainer）；具体说法一条都没有时给一条 aspect="other" 的兜底说明（「两条 SQL 的写法不同，未能归纳成具体
+    不等价时怎么说：两边的规范形式都是简单形状时（一段 SELECT、实表、普通的连接，见 _simple_block），只把五类有把握、
+    真会影响结果的不同写成具体的中文句子（按哪一列筛时间、一边多出来的筛选条件、同一列上的条件改写、顶层的聚合函数和
+    去重、关联的表和关联类型，见 _Explainer）；具体说法一条都没有时给一条 aspect="other" 的兜底说明（「两条 SQL 的写法不同，未能归纳成具体
     差异，请对照原文」），只覆盖了一部分时追加一条「另有未能归纳的差异」，fallback 为真；text_diff 给出规范化之后不同
     的行（没有规范形式可比时给原文的逐行对照）。
 
diff --git a/backend/app/data/value_drift.py b/backend/app/data/value_drift.py
index a165f995..432d947a 100644
--- a/backend/app/data/value_drift.py
+++ b/backend/app/data/value_drift.py
@@ -13,9 +13,17 @@
   记 info，列出变化最大的几个取值。
 - **断档**（date_stalled）：最新业务日期没前进，且这个日期离今天超过 max(2 × 以前前进的中位间隔, 3 天)；要见过至少一次
   前进才知道它多久前进一次。
-- **缺天**（date_gap）：窗口里（有数据的第一天到最后一天之间）有日子没数据，而以前同口径的窗口都没有过。
+- **缺天**（date_gap）：窗口里（有数据的第一天到最后一天之间）有日子没数据，而以前同口径的窗口都没有过。只看到观察当天
+  为止：业务日期有未来值（预约、计划类的日期，或者写错的）时，今天以后的空日子不是「没数据」，那几天本来就还没到；以前
+  那次观察也只看到它自己那天（data.until）为止，不然它窗口里未来的空日子会被当成「以前就有过缺天」，挡住这一次真缺的天。
 - **量跌**（volume_drop）：最近一个整天的行数不到它前面 28 天中位数的 30%（前 28 天至少 14 天有数据、中位数至少 20 行）。
-  最后一天和观察当天同一天（还在写入）时往前退一天。只看这一次的窗口，不需要历史。
+  最近一个整天是「有数据的最后一天」和「观察当天的前一天」里早的那个：观察当天还在写入，不是整天；业务日期有未来值时
+  有数据的最后一天在今天以后，也不能拿它退一天（那样会退到今天这个不完整的日子上）。只看这一次的窗口，不需要历史。
+- **断档**遇到未来值：最新业务日期在今天以后时离今天是负数，不会报断档（报不出、也不误报）；观察查询本身不截未来值
+  （最新业务日期照实记，目录里也照实给人看）。
+
+「今天」由调用方传进来（today）：取值观察按服务端所在时区算（data/catalog_profile_observe.local_today），业务日期是本地
+日期，用 UTC 的日期在本地凌晨会差一天。
 - **行数**（rows_drop / rows_jump）：比上一次同口径的少了 30% 以上 / 涨到 3 倍以上；依据是数据库统计的估算值时，要连续
   两次都这样（和跌、涨之前那一次比）才报。数到上限、只知道「至少多少行」的不比。
 """
@@ -177,7 +185,10 @@ def _series(data: Mapping[str, Any]) -> dict[date, int]:
     return out
 
 
-def _holes(days: Mapping[date, int]) -> list[date]:
+def _holes(days: Mapping[date, int], until: date | None = None) -> list[date]:
+    """有数据的第一天到最后一天之间没数据的日子；until 给了时只看到这一天为止（今天以后的空日子不算缺天）。"""
+    if until is not None:
+        days = {d: n for d, n in days.items() if d <= until}
     if not days:
         return []
     first, last = min(days), max(days)
@@ -190,15 +201,14 @@ def daily(current: Obs, history: Iterable[Obs], *, today: date) -> list[Finding]
     same, _note = comparable(current, history)
     days = _series(current.data)
     out: list[Finding] = []
-    holes = _holes(days)
-    if holes and same and all(not _holes(_series(h.data)) for h in same):
+    holes = _holes(days, today)
+    if holes and same and all(not _holes(_series(h.data), _day(h.data.get("until"))) for h in same):
         out.append(Finding("date_gap", "warning", before={"gaps": 0},
                            after={"gaps": len(holes), "days": [d.isoformat() for d in holes[:GAPS_SHOWN]]},
                            key=tuple(d.isoformat() for d in holes)))
     if days:
-        last = max(days)
-        if last >= today:
-            last -= timedelta(days=1)
+        # 最近一个整天：有数据的最后一天和昨天里早的那个（今天还在写入；未来的日子不是整天）
+        last = min(max(days), today - timedelta(days=1))
         base = [days.get(last - timedelta(days=i), 0) for i in range(1, VOLUME_BASELINE_DAYS + 1)]
         present = sum(1 for n in base if n > 0)
         median = statistics.median(base) if base else 0
diff --git a/backend/app/engine/runner.py b/backend/app/engine/runner.py
index ac31b861..c3acfd8c 100644
--- a/backend/app/engine/runner.py
+++ b/backend/app/engine/runner.py
@@ -52,6 +52,8 @@ END_REJECTED = "rejected"
 END_STOPPED = "stopped"
 #: 被驳回的运行不能继续运行：从断点重跑会读到同一个驳回答复、再停在同一处
 REJECTED_CONTINUE = "被驳回的运行不能继续运行：改好内容后重新发起"
+#: 被驳回的运行也没有可恢复的审批（恢复、继续运行同一个错误码 RUN_REJECTED）
+REJECTED_RESUME = "这次运行已被驳回，没有待处理的审批，无法恢复：改好内容后重新发起"
 #: 服务重启时还在排队的运行改成失败时写的话（run.error 和补发的 run.failed 是同一句）
 QUEUED_AT_RESTART = "服务重启时这次运行仍在排队，没有可恢复的进度，请重新发起运行"
 
@@ -355,7 +357,7 @@ class RunManager:
             if not run:
                 raise KeyError("运行记录不存在")
             if run.status not in ("interrupted",):
-                raise ValueError(_cannot_resume(run.status))
+                raise await _resume_refused(session, run)
             spec = GraphSpec.model_validate(run.graph)
             workflow_id = run.workflow_id
             carried = _carried(run)
@@ -458,8 +460,8 @@ class RunManager:
                 await session.commit()
                 run = await session.get(Run, run_id)
                 if run.status != "interrupted":
-                    # 审批没了是因为运行在这期间被放弃了：不能回一个「已记下」
-                    raise ValueError(_cannot_resume(run.status))
+                    # 审批没了是因为运行在这期间被放弃了（或被别人驳回了）：不能回一个「已记下」
+                    raise await _resume_refused(session, run)
                 return run
 
             # 齐了：正式落 resolved，然后驱动引擎。谁批的跟着交给节点，
@@ -1186,6 +1188,16 @@ _STATUS_TEXT = {
 }
 
 
+async def _resume_refused(session: Any, run: Run) -> ValueError:
+    """恢复被拒时抛什么：被驳回的（cancelled + end_reason=rejected，以及加 end_reason 之前记成 failed 的驳回）说「被驳回」、
+    带 RUN_REJECTED，和继续运行同一个错误码（以前照状态说「已取消的运行请重新发起」、对存量的还叫人点「继续运行」，
+    可继续运行也会被拒）；别的照状态说（_cannot_resume）。"""
+    if (run.status == "cancelled" and run.end_reason == END_REJECTED) or (
+            run.status == "failed" and await _rejected_before(session, run)):
+        return ContinueRefused(REJECTED_RESUME, code=RUN_REJECTED)
+    return ValueError(_cannot_resume(run.status))
+
+
 def _cannot_resume(status: str) -> str:
     hint = {
         "running": "运行仍在进行，无需恢复。",
diff --git a/backend/app/eval/runner.py b/backend/app/eval/runner.py
index 03224dcd..3412c005 100644
--- a/backend/app/eval/runner.py
+++ b/backend/app/eval/runner.py
@@ -708,20 +708,26 @@ class EvalExecution(runtime.JobRun):
                                   .execution_options(synchronize_session=False))
             await session.commit()
 
+        # 停止、中断是「停下」不是「出错」：原因记 stop_reason（error 留空）；整体出错记 error
+        stopped = status in ("cancelled", "interrupted")
         try:
             await self._write(write)
             if not self.arms:
                 async def bare(session: AsyncSession) -> None:
-                    await session.execute(update(EvalRun).where(
+                    ids = list((await session.execute(select(EvalRun.id).where(
                         (EvalRun.id == self.job_id) | (EvalRun.pair_id == self.job_id),
-                        EvalRun.status.in_(runtime.ACTIVE))
-                        .values(status=status, error=error, finished_at=utcnow())
-                        .execution_options(synchronize_session=False))
+                        EvalRun.status.in_(runtime.ACTIVE)))).scalars())
+                    await session.execute(update(EvalRun).where(EvalRun.id.in_(ids))
+                                          .values(status=status, error=None if stopped else error,
+                                                  stop_reason=error if stopped else None, finished_at=utcnow())
+                                          .execution_options(synchronize_session=False))
+                    for run_id in ids:
+                        await _recount(session, run_id)
                     await session.commit()
 
                 await self._write(bare)
                 return
-            await self._conclude(status, None, error)
+            await self._conclude(status, error if stopped else None, None if stopped else error)
         except Exception:  # noqa: BLE001
             logger.exception("评测 %s 收尾时写库出错", self.job_id)
 
@@ -793,16 +799,31 @@ def _sql_shapes(answer: Any) -> list[dict[str, Any]]:
     return out
 
 
+#: 逐项结束了、却没有跑出结果的状态：cancelled 停止、中断时还没评完的，skipped 到了金额上限、止损之后没开评的
+NOT_RUN_STATUSES = ("cancelled", "skipped")
+
+
 def _counts(rows: Sequence[Mapping[str, Any]]) -> dict[str, int]:
-    out = {"total": len(rows), "done": 0, "running": 0, "pending": 0, "passed": 0, "failed": 0, "ungradable": 0,
-           "skipped": 0, "errors": 0}
+    """一臂的计数（EvalRun.counts）。几样各管各的，加起来不重：
+
+    - done：评完了的（ok、partial）和评的时候出错的（failed）——真的跑过的；running、pending 是还在评、还没轮到的；
+    - not_run：停止、中断时还没评完，到了上限、止损之后没开评的（NOT_RUN_STATUSES）。以前这些算进 done 和 errors，停下的
+      评测就成了 done = total、一堆「出错」，读起来像全评完了、出了很多错；
+    - errors：出错的（failed：题目没了、评的时候出了意外），不含没跑的；
+    - passed / failed / ungradable / skipped：评完了的按结论分（skipped 是评了、但期望算不出、不评分，不含没开评的）。
+
+    done + not_run + running + pending = total。"""
+    out = {"total": len(rows), "done": 0, "running": 0, "pending": 0, "not_run": 0, "passed": 0, "failed": 0,
+           "ungradable": 0, "skipped": 0, "errors": 0}
     for r in rows:
         status = r.get("status")
-        if status in runtime.TASK_DONE:
+        if status in NOT_RUN_STATUSES:
+            out["not_run"] += 1
+        elif status in runtime.TASK_DONE:
             out["done"] += 1
         elif status in ("running", "pending"):
             out[status] += 1
-        if status in ("failed", "cancelled"):
+        if status == "failed":
             out["errors"] += 1
         verdict = r.get("verdict")
         if verdict == "pass":
@@ -811,11 +832,21 @@ def _counts(rows: Sequence[Mapping[str, Any]]) -> dict[str, int]:
             out["failed"] += 1
         elif verdict == "ungradable":
             out["ungradable"] += 1
-        elif verdict == "skipped" or status == "skipped":
+        elif verdict == "skipped":
             out["skipped"] += 1
     return out
 
 
+def _closing(run_id: str, reason: str) -> runtime.Finish:
+    """runtime.close_job / cancel_queued 的 finish：同一个事务里记停下的原因（停止、中断），按改好的逐项状态重算计数。"""
+    async def finish(session: AsyncSession) -> None:
+        await session.execute(update(EvalRun).where(EvalRun.id == run_id).values(stop_reason=reason)
+                              .execution_options(synchronize_session=False))
+        await _recount(session, run_id)
+
+    return finish
+
+
 async def _recount(session: AsyncSession, run_id: str) -> None:
     rows = (await session.execute(select(EvalResult.status, EvalResult.verdict).where(
         EvalResult.eval_run_id == run_id))).all()
@@ -910,16 +941,16 @@ async def reconcile_stale(session: AsyncSession, rows: Sequence[EvalRun]) -> Non
     """读接口当场整理：记着排队、进行中，这个进程里却没在跑（服务重启过）的，照实标中断。"""
     stale = [r for r in rows if r.status in runtime.ACTIVE and not running(r)]
     for row in stale:
-        await runtime.close_job(session, EvalRun, EvalResult, row.id, "interrupted", INTERRUPTED, "cancelled",
-                                INTERRUPTED)
+        await runtime.close_job(session, EvalRun, EvalResult, row.id, "interrupted", None, "cancelled",
+                                INTERRUPTED, finish=_closing(row.id, INTERRUPTED))
         await session.refresh(row)
 
 
 async def startup() -> int:
     """服务启动时：没结束的评测标中断（没结束的项记停止），不自动续跑。"""
     async def close(session: AsyncSession, job_id: str) -> None:
-        await runtime.close_job(session, EvalRun, EvalResult, job_id, "interrupted", INTERRUPTED, "cancelled",
-                                INTERRUPTED)
+        await runtime.close_job(session, EvalRun, EvalResult, job_id, "interrupted", None, "cancelled",
+                                INTERRUPTED, finish=_closing(job_id, INTERRUPTED))
 
     return await runtime.interrupt_on_startup(EvalRun, close, running=manager.running)
 
@@ -938,7 +969,8 @@ async def stop(run_id: str) -> bool:
             (EvalRun.id == leader) | (EvalRun.pair_id == leader)))).scalars())
         done = False
         for arm in arms:
-            done = await runtime.cancel_queued(session, EvalRun, EvalResult, arm) or done
+            done = await runtime.cancel_queued(session, EvalRun, EvalResult, arm,
+                                               finish=_closing(arm, STOPPED)) or done
         return done
 
 
diff --git a/backend/app/eval/stats.py b/backend/app/eval/stats.py
index 395eace6..7e4f6a8e 100644
--- a/backend/app/eval/stats.py
+++ b/backend/app/eval/stats.py
@@ -1,8 +1,8 @@
 """评测的汇总与统计（细案 01 做法 11、12；评测-D10）。所有汇总都在后端算好，界面直接显示（界面层不放逻辑：同一个数不能
 有两套算法）。纯计算，输入是一次评测的逐题结果（EvalResult 的列，result_row 的形状）。
 
-- 准确率和可追溯准确率：分母是能判的题（通过 + 不通过）；不能评分、跳过的单列，既不算对也不算错，没开评、停下、出错的
-  另列 not_run。都带 Wilson 95% 区间和样本数——题集小、模型有随机性，差两三题就可能被读成「变好了」「变差了」（评测-D10：只展示、不卡门）。
+- 准确率和可追溯准确率：分母是能判的题（通过 + 不通过）；不能评分、跳过的单列，既不算对也不算错，没跑的（停止、中断、
+  到了上限、止损没开评）另列 not_run，出错的另列 errors。都带 Wilson 95% 区间和样本数——题集小、模型有随机性，差两三题就可能被读成「变好了」「变差了」（评测-D10：只展示、不卡门）。
 - 作答路径、出具档位（声明了出具的按 formal / degraded / withheld，没声明的单列「探索问答」）、复核结论（verdict × severity）
   和自动重来的次数、复核用时超过问数据页等待（25 秒）的题数。
 - 结论裁判（打开了才有）：supported / partial / contradicted / insufficient / 未判的句数；文字要点的判断单列。
@@ -79,7 +79,7 @@ def summarize(results: Sequence[Mapping[str, Any]], *, judge_on: bool = False) -
     from app.eval.score import SCORER_VERSION
 
     done = [r for r in results if r.get("status") in ("ok", "skipped", "failed", "cancelled", "partial")]
-    #: 评完了的（status=ok）；没开评、停下、出错的单列成 not_run
+    #: 评完了的（status=ok）；没跑的（停止、中断、到了上限、止损没开评）单列成 not_run，出错的单列成 errors（同 runner._counts）
     evaluated = [r for r in done if r.get("status") == "ok"]
     verdicts = Counter(r.get("verdict") or "none" for r in done)
     graded = [r for r in done if r.get("verdict") in ("pass", "fail")]
@@ -118,7 +118,8 @@ def summarize(results: Sequence[Mapping[str, Any]], *, judge_on: bool = False) -
         "items": len(results), "done": len(done),
         "cases": len({r.get("case_id") for r in results}),
         "verdicts": {k: verdicts.get(k, 0) for k in ("pass", "fail", "ungradable", "skipped")},
-        "not_run": len(done) - len(evaluated),
+        "not_run": sum(1 for r in done if r.get("status") in ("cancelled", "skipped")),
+        "errors": sum(1 for r in done if r.get("status") == "failed"),
         "accuracy": rate(len(passed), len(graded)),
         "traced_accuracy": rate(len(traced), len(graded)),
         "outcomes": dict(outcomes),
diff --git a/backend/app/ledger/certify.py b/backend/app/ledger/certify.py
index cf57526b..f4384c2b 100644
--- a/backend/app/ledger/certify.py
+++ b/backend/app/ledger/certify.py
@@ -11,13 +11,16 @@
   个问题的类；
 - 被反复用的答案（reused）：导出过 ≥ 2 次（report.exported 审计，第 2 波）、做成过应用（蓝图的来源是这次运行）、在画布里
   打开过或提取过（工作流的 origin.run_id 是这次运行）的轮次。
-每项给代表问题、轮次链接、（shape 的）SQL 形状和参数候选、有没有相关口径（可以对账的那份）、是不是已经固化过。点「不是口径」
+每项给代表问题、轮次链接、（shape 的）SQL 形状和参数候选、有没有相关口径（可以对账的那份）、是不是已经固化过（frozen_into，
+同一个工作流只列一次）。点「不是口径」
 写一条审计（ledger.certify.dismiss，subject 是这一项的 id），之后不再进队。门槛同 ledger/stats（E3 的重复度）。
 
 **固化成口径草稿**（freeze，from-run 加 freeze 选项后转调它）：代表运行里报告引用了的查询，各变成一个调用工具节点——SQL 用
 规范化后的文本，跨轮变化的字面量换成 '{{ input.p1 }}'（同组的别的运行 peers 用来认哪些位置变了），入口字段的缺省值取
-代表运行里的值；只有一行的结果每个数值列建一项口径卡指标（cell(nodes.取数, 0, '列')）；报告节点保留，出口契约写
-report_from + metrics_from（有多行结果的另加 cells: true），过 validate_graph。新工作流记 origin.sources：
+代表运行里的值；只有一行的结果每个数值列建一项口径卡指标（cell(nodes.取数, 0, '列')，指标名能从数据目录取到中文名的
+用中文名，取不到用列名）；口径卡的名字用传入的 name，没传用代表问题；报告节点保留，指令在有入口参数时不写代表那一轮的问题
+原文（里面的日期参数化之后就不对了），改写运行时渲染的参数，没有参数时照旧写问题原文；出口契约写 report_from +
+metrics_from（有多行结果的另加 cells: true），过 validate_graph。新工作流记 origin.sources：
 [{run_id, conversation_id, turn_id, question}]，这就是「口径 → 来源问题」的记录。之后由人在画布上审改、按受管级别发布——
 提取器是作者，不是发布者。
 
@@ -64,6 +67,9 @@ DRAFT_NAME = "{question}（固化的口径草稿）"
 DRAFT_DESCRIPTION = ("从问数据的回答固化：报告引用了的 {n} 条查询变成固定的取数，跨轮变化的 {p} 个字面量成为入口参数。"
                      "经人工审核、按受管级别发布后，再问同样的问题就能认出等价、按口径出具")
 REPORT_INSTRUCTIONS = "回答：{question}"
+#: 有入口参数时的报告指令：不写代表那一轮的问题原文——里面的日期（「上周（10月5日到11日）」「九月第一周」）参数化之后
+#: 就不对了；改写这次运行时渲染出来的参数。{params} 是「p1 = {{ input.p1 }}，…」
+REPORT_INSTRUCTIONS_PARAMS = "按取到的数回答这个口径的问题（这次的参数：{params}；时间范围、筛选的取值以参数为准）"
 
 
 class CertifyRefused(Exception):
@@ -228,6 +234,16 @@ async def _frozen(session: Any) -> dict[str, list[dict[str, Any]]]:
     return out
 
 
+def _frozen_of(run_ids: Sequence[str | None], frozen: dict[str, list[dict[str, Any]]]) -> list[dict[str, Any]]:
+    """这几次运行固化进了哪些工作流：同一个工作流只列一次（固化时同组的运行都记进 origin.sources，一项里的几次运行常常
+    指着同一个工作流），按先后。"""
+    out: dict[str, dict[str, Any]] = {}
+    for run_id in run_ids:
+        for w in frozen.get(run_id or "", []):
+            out.setdefault(str(w["workflow_id"]), w)
+    return list(out.values())
+
+
 def _turn_view(member: Any) -> dict[str, Any]:
     return {"run_id": member.run_id, "conversation_id": member.conversation_id, "turn_id": member.turn_id,
             "question": member.question}
@@ -308,7 +324,7 @@ async def queue(session: Any, *, principal: Any) -> dict[str, Any]:
             "constants": len(rep.literals) - len(params), "related": related,
             "representative": {"run_id": rep.run_id, "query": rep.artifact},
             "peers": [m.run_id for m in once[1:]],
-            "frozen_into": [w for m in once for w in frozen.get(m.run_id, [])],
+            "frozen_into": _frozen_of([m.run_id for m in once], frozen),
         })
     # 问题聚类
     questions = await recent_questions(session, principal, limit=QUESTION_LIMIT)
@@ -322,7 +338,7 @@ async def queue(session: Any, *, principal: Any) -> dict[str, Any]:
         items.append({"id": iid, "kind": "cluster", "question": seed.text, "count": cluster.count,
                       "turns": [{"run_id": q.run_id, "turn_id": q.turn_id, "question": q.text}
                                 for q in cluster.members[:TURNS_LIMIT]],
-                      "frozen_into": [w for q in cluster.members if q.run_id for w in frozen.get(q.run_id, [])]})
+                      "frozen_into": _frozen_of([q.run_id for q in cluster.members], frozen)})
     items += await _reused(session, runs, turns, dismissed, frozen)
     return {"items": items, "considered": {"runs": len(runs), "matched": len(latest), "questions": len(questions)},
             "threshold": {"runs": REPEAT_RUNS, "conversations": REPEAT_CONVERSATIONS, "exports": EXPORTS},
@@ -367,7 +383,7 @@ async def _reused(session: Any, runs: list[Run], turns: dict[str, dict[str, Any]
         out.append({"id": iid, "kind": "reused", "question": info.get("question") or "", "uses": uses,
                     "turns": [{"run_id": run.id, "conversation_id": info.get("conversation_id"),
                                "turn_id": info.get("turn_id"), "question": info.get("question") or ""}],
-                    "representative": {"run_id": run.id}, "frozen_into": frozen.get(run.id, [])})
+                    "representative": {"run_id": run.id}, "frozen_into": _frozen_of([run.id], frozen)})
     return out
 
 
@@ -385,6 +401,24 @@ async def dismiss(session: Any, iid: str, *, kind: str | None = None, note: str
 # ==========================================================================
 
 
+async def _labels(session: Any, snap: dict[str, Any], source: str) -> Any:
+    """结果列的中文名：先看查询当时冻结的数据目录（快照的 schema_artifact，同口径比对），没冻结的（老快照）看数据源现在
+    的目录。只认列名本身就是目录里的一列（SELECT amount、SUM(amount) AS amount）；算出来起的别名（gmv）取不到，调用方
+    照旧用列名。返回 _labeler：(表, 列) → 中文名或 None。"""
+    from app.data import catalog
+    from app.data.sql_semantics import _labeler, catalog_labels
+    from app.db.models import DataSource
+    from app.ledger.match import _frozen as frozen_schema
+
+    _, names = frozen_schema(snap.get("schema_artifact"))
+    if not names:
+        source_id = (await session.execute(select(DataSource.id).where(DataSource.name == source))).scalar()
+        if source_id:
+            entries = await catalog.read_catalog(session, source_id)
+            names = catalog_labels({table: entry.notes for table, entry in entries.items()})
+    return _labeler(names)
+
+
 async def freeze(session: Any, run_id: str, *, peers: Sequence[str] = (), name: str = "",
                  conversation_id: str | None = None, turn_id: str | None = None) -> dict[str, Any]:
     """把一次问数据运行固化成口径草稿（见模块说明），落库成新的草稿工作流，返回 {workflow_id, name, nodes, edges, params,
@@ -460,9 +494,10 @@ async def freeze(session: Any, run_id: str, *, peers: Sequence[str] = (), name:
         rows = snap.get("rows") if isinstance(snap.get("rows"), list) else []
         columns = [str(c) for c in snap.get("columns") or []]
         if len(rows) == 1 and isinstance(rows[0], (list, tuple)):
+            label = await _labels(session, snap, shape.source)
             for col, value in zip(columns, rows[0]):
                 if isinstance(value, (int, float)) and not isinstance(value, bool):
-                    metrics.append({"id": f"m{len(metrics) + 1}", "name": col,
+                    metrics.append({"id": f"m{len(metrics) + 1}", "name": label((None, col)) or col,
                                     "expression": f"cell(nodes.{nid(i)}, 0, {col!r})"})
         else:
             multi = True
@@ -473,15 +508,20 @@ async def freeze(session: Any, run_id: str, *, peers: Sequence[str] = (), name:
         found = next((n for n in (run.graph or {}).get("nodes") or [] if n.get("id") == reports[0].node_id), None)
         cfg = ((found or {}).get("data") or {}).get("config") or {}
         report_cfg = {k: cfg[k] for k in ("provider", "model", "entities") if k in cfg}
-    report_cfg["instructions"] = REPORT_INSTRUCTIONS.format(question=question or "（问题）")
+    if fields:
+        report_cfg["instructions"] = REPORT_INSTRUCTIONS_PARAMS.format(
+            params="，".join(f"{f['name']} = {{{{ input.{f['name']} }}}}" for f in fields))
+    else:
+        # 没有入口参数：取数的字面量都是口径的常量，问题原文里的日期和它们对得上
+        report_cfg["instructions"] = REPORT_INSTRUCTIONS.format(question=question or "（问题）")
     # 报告的核对规则一律按受管级别的要求写（engine/governance.REPORT_POLICY，和发布门禁同一份）：草稿要能按受管发布
     report_cfg.update({key: value for key, value, _ in REPORT_POLICY})
     contract: dict[str, Any] = {"report_from": "write"}
     tail: list[dict[str, Any]] = []
     if metrics:
         tail.append({"id": "card", "type": "metrics", "position": {"x": 0, "y": 0},
-                     "data": {"label": "口径卡", "config": {"caliber": (question or "口径")[:40], "caliber_version": "v1",
-                                                         "metrics": metrics}}})
+                     "data": {"label": "口径卡", "config": {"caliber": (name.strip() or question or "口径")[:40],
+                                                         "caliber_version": "v1", "metrics": metrics}}})
         report_cfg["metrics_from"] = ["card"]
         # 受管出具要写必需指标（不然「不予出具」这一档触发不了）：固化出的每个指标都是这个口径要给的数
         contract.update({"metrics_from": "card", "required": [m["id"] for m in metrics]})
diff --git a/backend/app/ledger/stats.py b/backend/app/ledger/stats.py
index 180d984a..64f37605 100644
--- a/backend/app/ledger/stats.py
+++ b/backend/app/ledger/stats.py
@@ -17,7 +17,8 @@
   段对话的算重复的形状，给出这样的形状几个、有重复形状的运行占几成；问题聚类（app/chat/questions：全站唯一的聚类，只取问题
   原文、不含 history、排除系统派生的运行）里 ≥ REPEAT_RUNS 个问题的类占几成。认证队列的转化（第 3 波，ledger/certify）：固化成
   口径草稿几份（origin.freeze 的工作流）、其中发布了几份、点了「不是口径」几项。
-- E4 召回：每个口径版本被用了多少次（比对里等价已证明的运行数、这一版的正式运行数）；每次发布（口径升到某一版）召回了几组、
+- E4 召回：每个口径版本被用了多少次（比对里等价已证明的运行数、这一版的正式运行数——只算图上有源头口径卡的版本，没有口径卡
+  的普通发布工作流的正式运行不算）；每次发布（口径升到某一版）召回了几组、
   收过尾的几组、其中数变了的几组和比例、重算失败的几组（ledger/recall：召回运行和它收尾时记的 ledger.recalled，第 4 波）。
   按召回运行的发起时间筛窗口；表结构变了之后的口径试算不算召回。
 """
@@ -289,8 +290,10 @@ async def _e4(session: Any, contents: list[dict[str, Any]], **scope: Any) -> dic
         stmt = stmt.where(Run.created_at >= scope["since"])
     if scope.get("until") is not None:
         stmt = stmt.where(Run.created_at < scope["until"])
-    formal = {ledger_labels.caliber_key(w, v): int(n)
-              for w, v, n in (await session.execute(stmt.group_by(Run.workflow_id, Run.version))).all()}
+    counted = {(w, v): int(n) for w, v, n in (await session.execute(stmt.group_by(Run.workflow_id, Run.version))).all()}
+    # 只算有口径卡的那几版：没有口径卡的普通发布工作流的正式运行不是「这个口径版本被用了」
+    calibers = await _caliber_versions(session, counted)
+    formal = {ledger_labels.caliber_key(w, v): n for (w, v), n in counted.items() if (w, v) in calibers}
     recalls = int(await session.scalar(select(func.count()).select_from(Run).where(Run.origin_kind == "recall")) or 0)
     usage = [{"caliber": key, "workflow_id": key.rsplit("@", 1)[0], "version": int(key.rsplit("@", 1)[1]),
               "proven_runs": proven.get(key, 0), "formal_runs": formal.get(key, 0)}
@@ -298,6 +301,24 @@ async def _e4(session: Any, contents: list[dict[str, Any]], **scope: Any) -> dic
     return {"usage": usage, "recall_runs": recalls, "per_publish": await _per_publish(session, **scope)}
 
 
+async def _caliber_versions(session: Any, keys: Collection[tuple[str, int]]) -> set[tuple[str, int]]:
+    """这几个（工作流, 版本）里图上有源头口径卡的那些（recall.source_cards：metrics 节点、不是转抄别处的 caliber_from）。"""
+    from sqlalchemy import or_
+
+    from app.db.models import WorkflowVersion
+    from app.ledger.recall import source_cards
+
+    keys = list(keys)
+    out: set[tuple[str, int]] = set()
+    for i in range(0, len(keys), 200):
+        chunk = keys[i:i + 200]
+        rows = (await session.execute(select(WorkflowVersion.workflow_id, WorkflowVersion.version, WorkflowVersion.graph)
+                                      .where(or_(*[(WorkflowVersion.workflow_id == w) & (WorkflowVersion.version == v)
+                                                   for w, v in chunk])))).all()
+        out |= {(w, v) for w, v, graph in rows if source_cards(graph)}
+    return out
+
+
 async def _per_publish(session: Any, **scope: Any) -> dict[str, Any]:
     """每次发布召回了几组、其中数变了的比例（E4 下半）：按（工作流、新版本）分，读召回运行收尾时记的结论。"""
     from app.core.events import EventType
diff --git a/backend/app/notify/service.py b/backend/app/notify/service.py
index 6ec395f5..462f4643 100644
--- a/backend/app/notify/service.py
+++ b/backend/app/notify/service.py
@@ -342,6 +342,18 @@ async def notify_task(task_id: str, event: str, *, users: list[str], exclude: fr
                        task_id=task_id)
 
 
+def _answers_text(uses: list[dict[str, Any]]) -> str:
+    """召回消息里「答案」那一行：按种类分开写条数（「问数回答 1 个、正式运行 1 个」，种类的叫法同版本用量 ledger.uses.KIND_TEXT），
+    只写条数、不放数。以前合成一个「2 个」，一个问数回答加一次正式运行读起来像两个问数回答。一个都没有写「0 个」。"""
+    from collections import Counter
+
+    from app.ledger.uses import KIND_TEXT, KINDS
+
+    counts = Counter(str(u.get("kind") or "") for u in uses)
+    order = [k for k in KINDS if counts.get(k)] + sorted(k for k in counts if k not in KINDS)
+    return "、".join(f"{KIND_TEXT.get(k, k or '答案')} {counts[k]} 个" for k in order) or "0 个"
+
+
 async def notify_recall(task_id: str, users: list[str]) -> int:
     """口径召回的待办开出来了（app/ledger/recall）：口径升版、数真变了的那一种推 ledger.recall——订阅了它的全局渠道、答案主人
     自己的个人渠道；然后照别的待办推 task.opened（已经推过 ledger.recall 的渠道不再推，同一个渠道只推一次）。目录、表结构变了
@@ -352,8 +364,9 @@ async def notify_recall(task_id: str, users: list[str]) -> int:
     答案、待办按这一批汇总，链到待办箱（各人在那里看到派给自己的那条）；别的几条连 task.opened 也不再推全局渠道。个人渠道
     照旧每个主人各一条，只数他的答案、链到他的那条待办。
 
-    消息照模块说明的规矩（X6）：口径名（搭建时起的工作流名）、新旧版本号、几个答案（几条待办）、召回编号、去处理的链接；不放
-    旧口径的值、新口径的值、差值和原答案的值。返回推了几个渠道。"""
+    消息照模块说明的规矩（X6）：口径名（搭建时起的工作流名）、新旧版本号、几个答案（按种类分开写条数：问数回答几个、正式运行
+    几个……，见 _answers_text；几条待办）、召回编号、去处理的链接；不放旧口径的值、新口径的值、差值和原答案的值。返回推了几个
+    渠道。"""
     from app.ledger import recall
 
     try:
@@ -387,13 +400,13 @@ async def notify_recall(task_id: str, users: list[str]) -> int:
         head = [("状态", "待处理"), ("口径", f"「{name[:80]}」"),
                 ("版本", f"v{origin.get('from_version')} → v{origin.get('to_version')}")]
         tail = [("召回", f"#{task.run_id[:6]}")]
-        mine = len(recall._uses_of(origin, recall._owner_of_key(key)))
-        everyone = len(recall._uses_of(origin, ("anyone", "")))
+        mine = _answers_text(recall._uses_of(origin, recall._owner_of_key(key)))
+        everyone = _answers_text(recall._uses_of(origin, ("anyone", "")))
 
-        def message(answers: int, tasks: int, link: tuple[str, str] | None) -> dict[str, Any]:
+        def message(answers: str, tasks: int, link: tuple[str, str] | None) -> dict[str, Any]:
             return {"event": event, "status": "pending", "ok": False,
                     "title": recall.CHANGED_TITLE.format(name=name[:80] or "口径", version=origin.get("to_version")),
-                    "lines": [*head, ("答案", f"{answers} 个"), *([("待办", f"{tasks} 条")] if tasks > 1 else []), *tail],
+                    "lines": [*head, ("答案", answers), *([("待办", f"{tasks} 条")] if tasks > 1 else []), *tail],
                     "links": [link] if base and link else [],
                     "subject": {"task_id": task_id, "run_id": task.run_id, "workflow_id": origin.get("workflow_id")}}
 
diff --git a/backend/app/providers/mock_model.py b/backend/app/providers/mock_model.py
index 0ad18dd2..77870282 100644
--- a/backend/app/providers/mock_model.py
+++ b/backend/app/providers/mock_model.py
@@ -126,6 +126,12 @@ class MockChatModel(BaseChatModel):
     def _llm_type(self) -> str:
         return "mock"
 
+    @property
+    def _identifying_params(self) -> dict[str, Any]:
+        """调用参数里带上模型名（同 OpenAI 兼容、Anthropic 的 model）：用量计量（providers/usage）按它记在哪个模型名下，
+        不报的话评测汇总里助手用量的模型就成了 unknown。"""
+        return {"model_name": self.model_name}
+
     def bind_tools(self, tools: Sequence[Any], **kwargs: Any) -> "MockChatModel":
         normalized: list[dict[str, Any]] = []
         for tool in tools:
diff --git a/backend/app/providers/usage.py b/backend/app/providers/usage.py
index 9404039a..56298a2c 100644
--- a/backend/app/providers/usage.py
+++ b/backend/app/providers/usage.py
@@ -18,8 +18,10 @@ async generator：with 块跨着 yield 的话，变量在两次 yield 之间留
 流末尾那一段用量收得到；只有模型在那一行之后还往下写超过 DRAIN_CHARS 字的闲话时才当场关流（不为闲话付钱），连接断了也一样，
 这一次只拿得到已经报了的部分：照样计次数、计已报的 token，记成估不出，不悄悄少算。
 
-模型名按发起调用时的参数取（invocation_params 的 model / model_name，就是接入里配的模型 id，价格目录按它查），取不到再
-看回复里的 model_name。
+模型名按发起调用时的参数取（invocation_params 的 model / model_name，就是接入里配的模型 id，价格目录按它查）；调用参数里
+没有的（以前的演示模型就没报，评测汇总里助手用量的模型因此记成 unknown），看 LangChain 照模型对象给回调的 metadata 里的
+ls_model_name；再取不到看回复里的 model_name，都没有才记 UNKNOWN_MODEL。OpenAI 兼容、Anthropic 的模型在调用参数里就带着
+model，演示模型（providers/mock_model）也把 model_name 报进调用参数。
 """
 from __future__ import annotations
 
@@ -54,6 +56,10 @@ class Meter(BaseCallbackHandler):
     def _started(self, run_id: UUID, serialized: dict[str, Any] | None, kwargs: dict[str, Any]) -> None:
         params = kwargs.get("invocation_params") or {}
         model = params.get("model") or params.get("model_name") or params.get("model_id")
+        if not model:
+            # LangChain 照模型对象给回调的 metadata 里带着 ls_model_name（调用时传的 model，否则对象上的 model /
+            # model_name）：调用参数里不报模型名的模型类也认得出
+            model = (kwargs.get("metadata") or {}).get("ls_model_name")
         if not model and isinstance(serialized, dict):
             model = (serialized.get("kwargs") or {}).get("model") or (serialized.get("kwargs") or {}).get("model_name")
         with self._lock:
diff --git a/backend/app/tools/datasource.py b/backend/app/tools/datasource.py
index ca79a969..fe249549 100644
--- a/backend/app/tools/datasource.py
+++ b/backend/app/tools/datasource.py
@@ -968,12 +968,18 @@ async def _on_demand(source: Any, missing: list[str], ctx: ToolContext | None) -
     from app.db.base import SessionLocal
     from app.look.base import ON_DEMAND_PER_LOOK, ON_DEMAND_PER_RUN
 
+    from app.db.models import Run
+
     out: dict[str, str] = {}
     lines: list[str] = []
+    starter: str | None = None
     try:
         async with SessionLocal() as session:
             reasons = await so.absent_reasons(session, source, missing)
             used = await so.count_added_in_run(session, ctx.run_id) if ctx is not None and ctx.run_id else 0
+            # 留痕记上运行的发起人（同 look 那条路记对话里的发起人）：不然只有 run_id，要回头去翻运行才知道是谁
+            if ctx is not None and ctx.run_id:
+                starter = await session.scalar(select(Run.started_by).where(Run.id == ctx.run_id))
     except Exception:  # noqa: BLE001 - 查不了原因就照旧说
         logger.exception("db_schema 查表不在的原因失败（数据源 %s）", getattr(source, "name", "?"))
         return out, lines
@@ -992,7 +998,7 @@ async def _on_demand(source: Any, missing: list[str], ctx: ToolContext | None) -
             out[t] = text
     if take:
         got = await so.add_to_scope(getattr(source, "id", ""), take, trail=so.Trail(
-            via=so.VIA_SCHEMA_TOOL, run_id=(ctx.run_id or None) if ctx else None,
+            via=so.VIA_SCHEMA_TOOL, actor=starter, run_id=(ctx.run_id or None) if ctx else None,
             ref=(ctx.node_id or None) if ctx else None,
             need=f"运行时 {SCHEMA_PREFIX}{getattr(source, 'name', '')} 点名：{'、'.join(take)}"))
         if got.cache is not None and got.added:
diff --git a/backend/tests/fixtures/dialect/mariadb.sql b/backend/tests/fixtures/dialect/mariadb.sql
index b837c0b0..75bdc227 100644
--- a/backend/tests/fixtures/dialect/mariadb.sql
+++ b/backend/tests/fixtures/dialect/mariadb.sql
@@ -22,7 +22,7 @@ SELECT MIN(CASE WHEN paid_at >= DATE '0001-01-02' AND paid_at < DATE '9999-12-31
 -- date_window
 SELECT status, COUNT(*) AS cnt FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' AND status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status LIMIT 21;
 -- day_volume
-SELECT DATE(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' GROUP BY DATE(paid_at) ORDER BY d;
+SELECT DATE(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '2026-10-02' GROUP BY DATE(paid_at) ORDER BY d;
 -- sample
 -- （这种方言没有这类查询）
 -- max_date
diff --git a/backend/tests/fixtures/dialect/mysql.sql b/backend/tests/fixtures/dialect/mysql.sql
index 17f0f5da..57141703 100644
--- a/backend/tests/fixtures/dialect/mysql.sql
+++ b/backend/tests/fixtures/dialect/mysql.sql
@@ -22,7 +22,7 @@ SELECT MIN(CASE WHEN paid_at >= DATE '0001-01-02' AND paid_at < DATE '9999-12-31
 -- date_window
 SELECT status, COUNT(*) AS cnt FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' AND status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status LIMIT 21;
 -- day_volume
-SELECT DATE(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' GROUP BY DATE(paid_at) ORDER BY d;
+SELECT DATE(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '2026-10-02' GROUP BY DATE(paid_at) ORDER BY d;
 -- sample
 -- （这种方言没有这类查询）
 -- max_date
diff --git a/backend/tests/fixtures/dialect/oracle.sql b/backend/tests/fixtures/dialect/oracle.sql
index ba09c3d6..d8c1de40 100644
--- a/backend/tests/fixtures/dialect/oracle.sql
+++ b/backend/tests/fixtures/dialect/oracle.sql
@@ -22,7 +22,7 @@ SELECT MIN(CASE WHEN paid_at >= DATE '0001-01-02' AND paid_at < DATE '9999-12-31
 -- date_window
 SELECT * FROM (SELECT status, COUNT(*) AS cnt FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' AND status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status) WHERE ROWNUM <= 21;
 -- day_volume
-SELECT TRUNC(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' GROUP BY TRUNC(paid_at) ORDER BY d;
+SELECT TRUNC(paid_at) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '2026-10-02' GROUP BY TRUNC(paid_at) ORDER BY d;
 -- sample
 SELECT * FROM (SELECT status, COUNT(*) AS cnt FROM sales.orders SAMPLE BLOCK (1) WHERE status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status) WHERE ROWNUM <= 21;
 -- max_date
diff --git a/backend/tests/fixtures/dialect/postgresql.sql b/backend/tests/fixtures/dialect/postgresql.sql
index 995eb3f3..0a9b9ca4 100644
--- a/backend/tests/fixtures/dialect/postgresql.sql
+++ b/backend/tests/fixtures/dialect/postgresql.sql
@@ -22,7 +22,7 @@ SELECT MIN(CASE WHEN paid_at >= DATE '0001-01-02' AND paid_at < DATE '9999-12-31
 -- date_window
 SELECT status, COUNT(*) AS cnt FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' AND status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status LIMIT 21;
 -- day_volume
-SELECT CAST(paid_at AS DATE) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '9999-12-31' GROUP BY CAST(paid_at AS DATE) ORDER BY d;
+SELECT CAST(paid_at AS DATE) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= DATE '2026-08-27' AND paid_at < DATE '2026-10-02' GROUP BY CAST(paid_at AS DATE) ORDER BY d;
 -- sample
 SELECT status, COUNT(*) AS cnt FROM sales.orders TABLESAMPLE SYSTEM (1) WHERE status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status LIMIT 21;
 -- max_date
diff --git a/backend/tests/fixtures/dialect/sqlite.sql b/backend/tests/fixtures/dialect/sqlite.sql
index 44e0a904..1947d6d2 100644
--- a/backend/tests/fixtures/dialect/sqlite.sql
+++ b/backend/tests/fixtures/dialect/sqlite.sql
@@ -22,7 +22,7 @@ SELECT MIN(CASE WHEN paid_at >= '0001-01-02' AND paid_at < '9999-12-31' THEN pai
 -- date_window
 SELECT status, COUNT(*) AS cnt FROM sales.orders WHERE paid_at >= '2026-08-27' AND paid_at < '9999-12-31' AND status IS NOT NULL GROUP BY status ORDER BY cnt DESC, status LIMIT 21;
 -- day_volume
-SELECT substr(paid_at, 1, 10) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= '2026-08-27' AND paid_at < '9999-12-31' GROUP BY substr(paid_at, 1, 10) ORDER BY d;
+SELECT substr(paid_at, 1, 10) AS d, COUNT(*) AS n FROM sales.orders WHERE paid_at >= '2026-08-27' AND paid_at < '2026-10-02' GROUP BY substr(paid_at, 1, 10) ORDER BY d;
 -- sample
 -- （这种方言没有这类查询）
 -- max_date
diff --git a/backend/tests/test_catalog_observe.py b/backend/tests/test_catalog_observe.py
index 68d07455..689cfe74 100644
--- a/backend/tests/test_catalog_observe.py
+++ b/backend/tests/test_catalog_observe.py
@@ -228,6 +228,70 @@ async def test_large_table_basis_window_or_first_rows_and_no_cross_basis_compare
         await drop_source(b)
 
 
+async def test_future_business_dates_stay_out_of_the_daily_window(client, tmp_path):
+    """业务日期最晚到明天、今天只写了几行：按日行数的查询只数到今天（上界，今天以后的不算），不记量跌。以前窗口只有下界，
+    明天那一天也进了窗口，比较器再退一天就退到今天这个不完整的日子上，记出一条量跌（下一轮验收）。"""
+    path = tmp_path / "ahead.db"
+    _build(path, rows=0)
+    db = sqlite3.connect(path)
+    rows = [(TODAY - timedelta(days=d)).isoformat() for d in range(1, 31) for _ in range(30)]
+    rows += [TODAY.isoformat()] * 3 + [(TODAY + timedelta(days=1)).isoformat()] * 30
+    db.executemany("INSERT INTO orders (status, paid_on, amount, secret) VALUES ('paid', ?, 1, 'x')", [(r,) for r in rows])
+    db.commit()
+    db.close()
+    sid = await _make(client, path)
+    try:
+        job_id = await _observe(sid)
+        got = {(o.kind, o.column_name): o for o in await _obs(sid)}
+        days = got[("daily_volume", "paid_on")].data["days"]
+        assert (TODAY + timedelta(days=1)).isoformat() not in days and days[TODAY.isoformat()] == 3
+        assert got[("daily_volume", "paid_on")].data["until"] == TODAY.isoformat()
+        async with SessionLocal() as session:
+            drops = list((await session.execute(select(SourceChange).where(
+                SourceChange.source_id == sid, SourceChange.kind == "volume_drop"))).scalars())
+        assert drops == [], "今天还没写完、明天的日子不是整天：不算量跌"
+        assert all(o.job_id == job_id for o in got.values())
+    finally:
+        await drop_source(sid)
+
+
+def test_observation_today_is_the_server_local_date(monkeypatch):
+    """观察用的「今天」按服务端所在时区（和待办汇总 notify.digest_at 同一个口径）：业务日期是本地日期，UTC 20 点在东八区
+    已经是第二天。"""
+    import time
+    from datetime import datetime, timezone
+
+    monkeypatch.setenv("TZ", "Asia/Shanghai")
+    time.tzset()
+    try:
+        assert observe.local_today(datetime(2026, 10, 17, 20, 0, tzinfo=timezone.utc)) == date(2026, 10, 18)
+        assert observe.local_today(datetime(2026, 10, 17, 15, 0, tzinfo=timezone.utc)) == date(2026, 10, 17)
+    finally:
+        monkeypatch.undo()
+        time.tzset()
+
+
+async def test_the_profile_jobs_api_takes_the_observe_scope(client, shop):
+    """取值观察以前只能经定时维护触发：POST /catalog/profile-jobs 的 scope 只收 needed、all（下一轮验收）。现在也收 observe：
+    只观察、不剖析；12 小时内都观察过时 409 profile_nothing，说清为什么。"""
+    sid, _ = shop
+    r = await client.post(f"/api/datasources/{sid}/catalog/profile-jobs", json={"scope": "observe"})
+    assert r.status_code == 202, r.text
+    job = r.json()["job"]
+    assert job["scope"] == "observe"
+    await catalog_profile_jobs.manager.wait(job["id"], timeout=60)
+    async with SessionLocal() as session:
+        [task] = (await session.execute(select(CatalogProfileTask).where(
+            CatalogProfileTask.job_id == job["id"]))).scalars()
+    assert {c["kind"] for c in task.checks} == set(observe.OBSERVE_KINDS), "只观察：不做关系、码值、日期的剖析"
+    assert {o.kind for o in await _obs(sid)} == set(observe.OBSERVE_KINDS)
+    again = await client.post(f"/api/datasources/{sid}/catalog/profile-jobs", json={"scope": "observe"})
+    assert again.status_code == 409 and again.json()["code"] == "profile_nothing"
+    assert f"{observe.OBSERVE_EVERY_H} 小时内都观察过" in again.json()["detail"]
+    bad = await client.post(f"/api/datasources/{sid}/catalog/profile-jobs", json={"scope": "maintenance"})
+    assert bad.status_code == 422, "maintenance 是定时维护用的，接口不收"
+
+
 class _FakeCx:
     """只看拼出来的 SQL：大表、业务日期没有索引。"""
 
diff --git a/backend/tests/test_catalog_profile_units.py b/backend/tests/test_catalog_profile_units.py
index 22933754..7ce3c114 100644
--- a/backend/tests/test_catalog_profile_units.py
+++ b/backend/tests/test_catalog_profile_units.py
@@ -145,7 +145,7 @@ def test_profile_sql_passes_readonly_guard_on_every_dialect(kind):
             d.value_counts_sql(tbl, col, 21, scan_cap=100_000),
             d.min_max_sql(tbl, [(col, False)]),
             d.min_max_sql(tbl, [(col, True)]),
-            d.day_volume_sql(tbl, col, "2026-08-01"),
+            d.day_volume_sql(tbl, col, "2026-08-01", "2026-09-04"),
             d.window_value_counts_sql(tbl, col, col, "2026-08-01", 21),
         ]
         sampled = d.sample_value_counts_sql(tbl, col, 1, 21)
diff --git a/backend/tests/test_catalog_schema_truncated.py b/backend/tests/test_catalog_schema_truncated.py
index f0c99435..beb9b500 100644
--- a/backend/tests/test_catalog_schema_truncated.py
+++ b/backend/tests/test_catalog_schema_truncated.py
@@ -147,6 +147,35 @@ async def test_schema_tool_adds_an_out_of_scope_table_instead_of_denying_it(wide
     assert "没有 nope" in await tool.coroutine(table="nope")
 
 
+async def test_schema_tool_on_demand_trail_records_who_started_the_run(wide):
+    """运行时 db_schema 按需加入范围的留痕：记上运行的发起人（by），不只是 run_id（下一轮验收：以前 by 是 null）。"""
+    from sqlalchemy import select as _select
+
+    from app.data import source_objects
+    from app.db.models import Run
+    from app.db.models_sources import SourceObject
+    from app.tools.datasource import build_datasource_tools
+    from app.tools.registry import ToolContext, Versions
+
+    sid, name = wide
+    async with SessionLocal() as session:
+        run = Run(workflow_name="按需加入", status="running", graph={"nodes": [], "edges": []}, input={},
+                  started_by="分析员甲")
+        session.add(run)
+        await session.commit()
+        run_id = run.id
+    async with SessionLocal() as session:
+        (tool,) = await build_datasource_tools(
+            [f"db_schema__{name}"], ToolContext(run_id=run_id, node_id="ask", data_versions=Versions.CURRENT), session)
+    assert "按需加入了探查范围" in await tool.coroutine(table="t204")
+    async with SessionLocal() as session:
+        obj = (await session.execute(_select(SourceObject).where(SourceObject.source_id == sid,
+                                                                 SourceObject.name == "t204"))).scalar_one()
+    trail = source_objects.trail_dict(obj)
+    assert trail["via"] == source_objects.VIA_SCHEMA_TOOL and trail["run_id"] == run_id and trail["ref"] == "ask"
+    assert trail["by"] == "分析员甲", trail
+
+
 def test_legacy_truncated_cache_still_does_not_deny():
     """升级前只探查了前 200 张的老缓存（没有对象索引）：查一张没探查到的表照旧说「只探查了」、请查数据字典。"""
     from types import SimpleNamespace
@@ -157,3 +186,24 @@ def test_legacy_truncated_cache_still_does_not_deny():
                                                        "truncated": True, "total": 205})
     text = describe_table(row, "t300")
     assert "中没有" not in text and "只探查了" in text and "information_schema" in text
+
+
+async def test_catalog_detail_of_an_out_of_scope_table_says_so(client, wide):
+    """对象索引里有、在探查范围外的表：GET /catalog/{表} 的 404 说「在探查范围外」，和 /schema、/catalog/search 同一套「不在」
+    （下一轮验收：以前说「可能已被删除或尚未探查结构」）；表助手、批量审阅的同一个错误也一样。对象索引里也没有的说「没有
+    这张表」。"""
+    sid, _ = wide
+    r = await client.get(f"/api/datasources/{sid}/catalog/t200")
+    assert r.status_code == 404 and "探查范围外" in r.json()["detail"], r.text
+    assert "可能已被删除" not in r.json()["detail"]
+    found = (await client.get(f"/api/datasources/{sid}/catalog/search", params={"q": "t200"})).json()["items"]
+    assert [i["absent_reason"] for i in found if i["name"] == "t200"] in (["out_of_scope"], [None]), found
+    r = await client.post(f"/api/datasources/{sid}/catalog/t201/assistant", json={"messages": [
+        {"role": "user", "content": "这张表是干什么的"}]})
+    assert r.status_code == 404 and "探查范围外" in r.json()["detail"], r.text
+    r = await client.post(f"/api/datasources/{sid}/catalog/review/batch", json={"items": [
+        {"table_name": "t202", "path": "label", "action": "confirm", "if_version": 0}]})
+    [result] = r.json()["results"]
+    assert result["ok"] is False and "探查范围外" in result["error"], result
+    r = await client.get(f"/api/datasources/{sid}/catalog/no_such_table")
+    assert r.status_code == 404 and "没有表 no_such_table" in r.json()["detail"] and "探查范围外" not in r.json()["detail"]
diff --git a/backend/tests/test_dual_check_facts.py b/backend/tests/test_dual_check_facts.py
index 66fdf79b..c6bff52d 100644
--- a/backend/tests/test_dual_check_facts.py
+++ b/backend/tests/test_dual_check_facts.py
@@ -8,8 +8,9 @@
   配方里认作日期的列（候选的 upload_dates）：是的算日期列，不是的、读不出配方的说清日期是按文字保存的；数据源还没有
   可用的表结构时说表结构不可用
 - 差了不止一处、有兜底说明、两条 SQL 规范化后一致、差在别的方面、被采纳的一边没有可记的写法、只有一边按时间筛：不写，
-  照实说原因。同一列上的时间范围不同、跨两张表的条件，差异说明只给兜底，同样不写；说明以后真把跨两张表的条件说成一条
-  筛选条件，derive 自己那道「只涉及一张表」的检查也拦得下
+  照实说原因。同一列上的时间范围不同（差异说明说成 time_range）、同一列上换了筛选的值（filter，两边都有条件）不是业务日期、
+  也不是一边多出来的有效记录条件，同样不写；跨两张表的条件差异说明只给兜底，说明以后真把它说成一条筛选条件，derive 自己
+  那道「只涉及一张表」的检查也拦得下
 - 指标自己的筛选范围（只看华东、只看线上）不是有效记录条件：取值出现在指标名称、口径说明、所在分组、运行输入里的，
   条件用到分组的列的，不是在排除退款、作废、测试这类无效记录的，都不写
 - 写的是人工填写、已确认，备注写明运行、节点、指标；同样的说法已确认过就不再写（不升版本），确认过的不同说法不覆盖；
@@ -234,10 +235,15 @@ def test_equivalent_sql_or_fallback_or_other_aspects_are_not_recorded():
     assert [i["aspect"] for i in joined["diff"]["items"]] == ["aggregate"]
     assert "不只是去不去重" in derive(joined, SCHEMA)[1]
 
-    # 同一列上的时间范围不同：差异说明只给兜底（说不清差的是哪一段），不写
+    # 同一列上的时间范围不同：差异说明说成 time_range（「原流程对 paid_date 的条件是 …，盲算是 …」），不是换了时间列，不写
     ranged = candidate(RIGHT, RIGHT.replace("'2026-09-07'", "'2026-09-06'"), "b")
-    assert ranged["diff"]["fallback"] is True and [i["aspect"] for i in ranged["diff"]["items"]] == ["other"]
-    assert "未能归纳" in derive(ranged, SCHEMA)[1]
+    assert ranged["diff"]["fallback"] is False and [i["aspect"] for i in ranged["diff"]["items"]] == ["time_range"]
+    assert "两边差在时间范围" in derive(ranged, SCHEMA)[1]
+    # 同一列上换了筛选的值：两边都有条件（filter 两边都有），不是一边多出来的有效记录条件，不写
+    valued = candidate(RIGHT, RIGHT.replace("status <> 'refunded'", "status NOT IN ('refunded', 'void')"), "b")
+    assert valued["diff"]["fallback"] is False and [(i["aspect"], i["a"] is not None, i["b"] is not None)
+                                                    for i in valued["diff"]["items"]] == [("filter", True, True)]
+    assert "两边各有不同的筛选条件" in derive(valued, SCHEMA)[1]
 
     # 只差一个方面、可那个方面对不上目录里的三项（关联方式）：不写，原因说清差在哪
     left = (f"SELECT SUM(o.amount) AS gmv FROM orders o LEFT JOIN customers c ON c.id = o.customer_id "
diff --git a/backend/tests/test_dual_check_fixture.py b/backend/tests/test_dual_check_fixture.py
index cbdea294..f10f100e 100644
--- a/backend/tests/test_dual_check_fixture.py
+++ b/backend/tests/test_dual_check_fixture.py
@@ -9,12 +9,13 @@ check-verify 在夹具上能过，真实数据上却一处都不高亮（2026-10
 差异都没有」，delta_pct 是 (盲算值 − 原流程值) ÷ |原流程值| × 100 保留两位。夹具的数据源当 SQLite 比（方言 sqlite），
 两边的叫法是「原流程」「盲算」，同节点调用时一样。
 
-**变体。** variants 换掉「销售额」的那一对 SQL，挑的是收尾版差异说明（只说时间列、多出来的筛选、聚合与去重、关联
-四类，别的只给兜底）下各有代表性的几种：按哪一列筛时间不同（time_column，两列都声明为日期类型，预告写业务日期）、
+**变体。** variants 换掉「销售额」的那一对 SQL，挑的是收尾版差异说明（只说时间列、多出来的筛选、同一列上的条件改写、
+聚合与去重、关联五类，别的只给兜底）下各有代表性的几种：按哪一列筛时间不同（time_column，两列都声明为日期类型，预告写业务日期）、
 只有一边按时间筛、一边多出来的筛选（排除无效记录的 filter_extra 预告写有效记录条件，指标自己范围的 filter_scope 不写）、
 差两个筛选条件、关联方式、只差一个关联条件（join_condition：LEFT JOIN 的 ON 里多一个条件，只给兜底，逐行对照里那一行
-条件前面带着它所在的 JOIN）、聚合只差去重（dedup，预告写去重规则）、说不清的（fallback：同一列上的时间范围不同，只给
-兜底，审批卡默认展开逐行对照）、一边解析失败。
+条件前面带着它所在的 JOIN）、聚合只差去重（dedup，预告写去重规则）、说不清的（fallback：套着 DATE() 的时间范围不同——
+光一列上的范围不同如今说得出「对 paid_at 的条件是 …」，套着函数的认不出取值范围，只给兜底，审批卡默认展开逐行对照）、
+一边解析失败。
 
 **写目录的预告。** 服务端在摘要里另带每个指标的 catalog（采用原流程、采用盲算各会往数据目录里写什么，
 dual_check_facts.preview，约定之外另加的），审批卡照它说；老服务端没有时卡上按差异的形状有条件地说，两种说法都按
diff --git a/backend/tests/test_eval_runner.py b/backend/tests/test_eval_runner.py
index 8e348397..c6e0ee79 100644
--- a/backend/tests/test_eval_runner.py
+++ b/backend/tests/test_eval_runner.py
@@ -146,10 +146,14 @@ async def test_stop_then_retry_only_the_unfinished(monkeypatch):
     await asyncio.sleep(0.1)
     assert await runner.stop(run.id)
     row = await get(run.id)
-    assert row.status == "cancelled" and row.error == runner.STOPPED
+    # 停止是「停下」不是「出错」：原因记在 stop_reason；没跑的题单独计 not_run，不算进 errors，done 不再等于 total
+    assert row.status == "cancelled" and row.stop_reason == runner.STOPPED and row.error is None
     before = {r.id: (r.status, r.finished_at) for r in await results(run.id)}
     done = [rid for rid, (status, _) in before.items() if status == "ok"]
     assert done and any(status == "cancelled" for status, _ in before.values())
+    assert row.counts["errors"] == 0 and row.counts["done"] == len(done)
+    assert row.counts["not_run"] == 4 - len(done) and row.counts["done"] + row.counts["not_run"] == row.counts["total"]
+    assert row.summary["not_run"] == 4 - len(done) and row.summary["errors"] == 0
     assert not await runner.stop(run.id), "已经结束的停不了"
 
     stub(monkeypatch, delay=0.01)
@@ -175,12 +179,25 @@ async def test_a_restart_marks_unfinished_evaluations_interrupted_and_does_not_r
         run_id = row.id
     assert await runner.startup() >= 1
     row = await get(run_id)
-    assert row.status == "interrupted" and row.error == runner.INTERRUPTED
+    assert row.status == "interrupted" and row.stop_reason == runner.INTERRUPTED and row.error is None
+    assert row.counts["not_run"] == 1 and row.counts["running"] == 0 and row.counts["errors"] == 0, row.counts
     [item] = await results(run_id)
     assert item.status == "cancelled"
     assert not runner.manager.running(run_id), "不自动续跑"
 
 
+async def test_stopping_a_queued_evaluation_counts_everything_as_not_run(monkeypatch):
+    """排队中就停掉（还没开始）：每一题都没跑，stop_reason 写停止，计数跟着重算（以前还是建的时候那份：pending 照旧）。"""
+    stub(monkeypatch)
+    suite_id = await es.make_suite("none", cases=ask_cases(2))
+    monkeypatch.setattr(runner.manager, "start", lambda *_a, **_k: None)  # 不让它开跑：停在排队
+    [run] = await start(suite_id)
+    assert await runner.stop(run.id)
+    row = await get(run.id)
+    assert row.status == "cancelled" and row.stop_reason == runner.STOPPED
+    assert (row.counts["not_run"], row.counts["pending"], row.counts["done"]) == (2, 0, 0), row.counts
+
+
 async def test_the_per_run_cap_stops_new_items(monkeypatch):
     stub(monkeypatch, cost=1.0)
     suite_id = await es.make_suite("none", cases=ask_cases(3), budget={"max_parallel": 1, "per_run_max_usd": 1.5})
@@ -190,6 +207,8 @@ async def test_the_per_run_cap_stops_new_items(monkeypatch):
     items = await results(run.id)
     assert [r.status for r in items] == ["ok", "ok", "skipped"]
     assert row.status == "partial" and row.stop_reason == runner.LIMIT_RUN.format(limit=1.5)
+    assert (row.counts["done"], row.counts["not_run"], row.counts["errors"], row.counts["skipped"]) == (2, 1, 0, 0), \
+        "到了上限没开评的是没跑，不是跳过（跳过是评了、但期望算不出、不评分）"
     assert "金额上限" in items[2].error and row.summary["cost_usd"] == 2.0
 
 
@@ -328,6 +347,44 @@ async def test_the_api_starts_estimates_and_reports_progress(client, monkeypatch
     assert bad.status_code == 422 and bad.json()["code"] == "eval_variant_invalid"
 
 
+async def test_results_give_the_question_actually_asked_and_the_template_apart(client, monkeypatch):
+    """逐题结果的 question 是这一题实际问的（变量按这次评测的截至时间渲染好了，和作答时 Driver 收到的一字不差），模板原文
+    另放 question_template（下一轮验收：以前 question 是带 {{ schedule.last_week_start }} 的原始模板）。没有变量的题两个一样。"""
+    from datetime import datetime, timezone
+
+    from app.db.models_eval import EvalCase
+    from app.eval.cases import render, variable_values
+
+    asked: list[str] = []
+    real_init = answer_mod.Driver.__init__
+
+    def init(self, question, **kw):
+        asked.append(question)
+        real_init(self, question, **kw)
+
+    monkeypatch.setattr(answer_mod.Driver, "__init__", init)
+    stub(monkeypatch)
+    template = "{{ schedule.last_week_start }} 到 {{ schedule.last_week_end }} 那周的问题"
+    suite_id = await es.make_suite("none", cases=[{"question": template, "expectations": [{"kind": "ask"}]},
+                                                  {"question": "没有变量的问题", "expectations": [{"kind": "ask"}]}])
+    as_of = datetime(2026, 10, 14, 2, 0, tzinfo=timezone.utc)
+    [run] = await start(suite_id, as_of=as_of)
+    await es.wait_eval(run.id)
+    rows = (await client.get(f"/api/evals/evaluations/{run.id}/results")).json()["results"]
+    expected = render(template, variable_values(as_of, "Asia/Shanghai", None))
+    assert "{{" not in expected and rows[0]["question"] == expected, rows[0]["question"]
+    assert rows[0]["question"] in asked, "和作答时实际问的一样"
+    assert rows[0]["question_template"] == template
+    assert rows[1]["question"] == rows[1]["question_template"] == "没有变量的问题"
+    # 题目改过（指纹对不上）：现在的模板不一定是当时问的，没有作答运行可取时不编一个，question 为空、模板照给
+    async with SessionLocal() as session:
+        case = (await session.execute(select(EvalCase).where(EvalCase.suite_id == suite_id, EvalCase.seq == 1))).scalar_one()
+        case.question, case.case_hash = "改过的 {{ schedule.today }}", "changed"
+        await session.commit()
+    rows = (await client.get(f"/api/evals/evaluations/{run.id}/results")).json()["results"]
+    assert rows[0]["question"] is None and rows[0]["question_template"] == "改过的 {{ schedule.today }}"
+
+
 async def test_eval_settings_group_and_spend_endpoint(client):
     got = (await client.get("/api/settings")).json()["eval"]
     assert got == {"max_parallel": 2, "per_case_timeout_s": 900, "per_run_max_usd": 2.0, "daily_max_usd": 5.0,
diff --git a/backend/tests/test_ledger_api.py b/backend/tests/test_ledger_api.py
index 2e5eeff2..3814e4e5 100644
--- a/backend/tests/test_ledger_api.py
+++ b/backend/tests/test_ledger_api.py
@@ -295,6 +295,35 @@ async def test_stats_count_what_the_fixtures_did(client, monkeypatch):
     assert body["e4"]["per_publish"]["available"] is True and body["e4"]["per_publish"]["groups"] == 0
 
 
+async def test_e4_usage_counts_formal_runs_only_of_workflows_with_a_caliber_card(client):
+    """E4 的版本用量只算有口径卡的工作流（下一轮验收：没有口径卡的普通发布工作流的正式运行也被当成口径计数）。卡是转抄别处
+    口径（caliber_from）的不算源头口径，同 recall.source_cards。"""
+    from datetime import datetime, timezone
+
+    from app.db.models import Run, Workflow, WorkflowVersion
+
+    since = datetime.now(timezone.utc)
+    card = {"id": "card", "type": "metrics", "position": {"x": 0, "y": 0},
+            "data": {"label": "口径卡", "config": {"caliber": "周销售", "caliber_version": "v1", "metrics": []}}}
+    plain = {"id": "out", "type": "output", "position": {"x": 0, "y": 0}, "data": {"label": "出口", "config": {}}}
+    async with SessionLocal() as session:
+        made = {}
+        for name, nodes in (("有口径卡", [card, plain]), ("普通工作流", [plain])):
+            wf = Workflow(name=name, graph={"nodes": nodes, "edges": []}, status="published", published_version=1)
+            session.add(wf)
+            await session.flush()
+            session.add(WorkflowVersion(workflow_id=wf.id, version=1, graph={"nodes": nodes, "edges": []}))
+            for _ in range(2):
+                session.add(Run(workflow_id=wf.id, version=1, run_class="formal", status="succeeded",
+                                graph={"nodes": nodes, "edges": []}, input={}))
+            made[name] = wf.id
+        await session.commit()
+    body = (await client.get("/api/ledger/stats", params={"since": since.isoformat()})).json()
+    usage = {u["caliber"]: u for u in body["e4"]["usage"]}
+    assert usage[f"{made['有口径卡']}@1"]["formal_runs"] == 2
+    assert f"{made['普通工作流']}@1" not in usage, usage
+
+
 async def test_stats_take_the_run_ids_the_caller_names(monkeypatch):
     """评测（第 3 波）直接调 compute_stats，点名自己的运行：只算这几次。"""
     from app.ledger.stats import compute_stats
diff --git a/backend/tests/test_ledger_certify.py b/backend/tests/test_ledger_certify.py
index e91dee59..4444064c 100644
--- a/backend/tests/test_ledger_certify.py
+++ b/backend/tests/test_ledger_certify.py
@@ -180,7 +180,11 @@ async def test_freezing_makes_a_fixed_sql_draft_that_passes_the_governed_gate(cl
                                                           "required": ["m1"]}
     assert {k: nodes["write"]["data"]["config"].get(k) for k in ("numbers", "on_violation", "claims")} == {
         "numbers": "strict", "on_violation": "fail", "claims": "require_citation"}
-    assert nodes["write"]["data"]["config"]["instructions"] == f"回答：{QUESTION}"
+    # 有入口参数：不写代表那一轮的问题原文（「九月第一周」参数化之后就不对了），写运行时渲染的参数
+    instructions = nodes["write"]["data"]["config"]["instructions"]
+    assert QUESTION not in instructions and "2026-09" not in instructions, instructions
+    assert "{{ input.p1 }}" in instructions and "{{ input.p2 }}" in instructions
+    assert nodes["card"]["data"]["config"]["caliber"] == QUESTION, "没传名字：口径卡照旧用代表问题"
     spec = GraphSpec.model_validate(wf.graph)
     assert not [i for i in validate_graph(spec).issues if i.level == "error"]
     assert not [i for i in publish_issues(spec, level="governed") if i.level == "error"], \
@@ -189,9 +193,12 @@ async def test_freezing_makes_a_fixed_sql_draft_that_passes_the_governed_gate(cl
     assert all(s["question"] == QUESTION for s in wf.origin["sources"])
     assert wf.origin["run_id"] == rep_run.id and wf.origin["freeze"] is True
 
-    # 队列上标着已经固化过
-    [again] = shape_items((await client.get("/api/ledger/certify-queue")).json())
-    assert {w["workflow_id"] for w in again["frozen_into"]} == {wf.id}
+    # 队列上标着已经固化过：同一个工作流只列一次（三个来源运行都指着它）
+    body = (await client.get("/api/ledger/certify-queue")).json()
+    [again] = shape_items(body)
+    assert [w["workflow_id"] for w in again["frozen_into"]] == [wf.id]
+    [cluster] = [i for i in body["items"] if i["kind"] == "cluster" and i["question"] == QUESTION]
+    assert [w["workflow_id"] for w in cluster["frozen_into"]] == [wf.id], cluster["frozen_into"]
 
     # 按受管级别发布：来源运行在后台补比，变成等价已证明、出了队
     res = await client.post(f"/api/workflows/{wf.id}/publish", json={"level": "governed"})
@@ -216,6 +223,47 @@ async def test_freezing_makes_a_fixed_sql_draft_that_passes_the_governed_gate(cl
     assert stats["frozen"] >= 1 and stats["published"] >= 1
 
 
+async def test_freezing_uses_the_given_name_and_catalog_labels_and_no_dates_without_parameters(client, monkeypatch):
+    """传了名字：口径卡用它（下一轮验收：以前用代表问题的原文）。指标名能从数据目录取到中文名的用中文名，取不到照旧用
+    列名。没有入口参数（字面量都没变过）时报告指令照旧写问题原文：那时问题里的日期就是口径的常量。"""
+    from app.data import catalog
+    from app.db.models import DataSource
+    from sqlalchemy import select
+
+    async with SessionLocal() as session:
+        source = (await session.execute(select(DataSource).where(DataSource.name == S.CURRENT["source"]))).scalar_one()
+        await catalog.write_entry(session, source.id, "orders", {"columns": {
+            "amount": {"label": catalog.make_item("金额", "human")}}}, if_version=0, actor="测试员")
+    sql = "SELECT SUM(amount) AS amount, COUNT(*) AS n FROM orders WHERE status = 'paid' AND paid_at >= '2026-09-01'"
+    S.script(monkeypatch, "金额 [[v:Q1.r0.amount]]，[[v:Q1.r0.n]] 单。")
+    conv = await conversation()
+    async with SessionLocal() as session:
+        turn = ConversationTurn(conversation_id=conv, question=QUESTION)
+        session.add(turn)
+        await session.commit()
+        turn_id = turn.id
+    run = await run_manager.start(graph=S.explore_graph(sql), input_payload={"question": QUESTION}, origin_kind="chat",
+                                  origin={"conversation_id": conv, "turn_id": turn_id})
+    row = await S.wait(run.id)
+    assert row.status == "succeeded", row.error
+    async with SessionLocal() as session:
+        (await session.get(ConversationTurn, turn_id)).run_id = row.id
+        await session.commit()
+    await ledger_match.drain()
+    res = await client.post("/api/copilot/from-run", json={"run_id": row.id, "freeze": True, "name": "九月销售额口径"})
+    assert res.status_code == 200, res.text
+    made = res.json()
+    assert made["params"] == [] and made["name"] == "九月销售额口径"
+    async with SessionLocal() as session:
+        wf = await session.get(Workflow, made["workflow_id"])
+    nodes = {n["id"]: n for n in wf.graph["nodes"]}
+    card = nodes["card"]["data"]["config"]
+    assert card["caliber"] == "九月销售额口径"
+    assert [(m["name"], m["expression"]) for m in card["metrics"]] == [
+        ("金额", "cell(nodes.fetch, 0, 'amount')"), ("n", "cell(nodes.fetch, 0, 'n')")]
+    assert nodes["write"]["data"]["config"]["instructions"] == f"回答：{QUESTION}"
+
+
 async def test_freezing_needs_a_matched_successful_run(client, monkeypatch):
     res = await client.post("/api/copilot/from-run", json={"run_id": "nope", "freeze": True})
     assert res.status_code == 404
diff --git a/backend/tests/test_ledger_recall.py b/backend/tests/test_ledger_recall.py
index 0052f306..7e7232c7 100644
--- a/backend/tests/test_ledger_recall.py
+++ b/backend/tests/test_ledger_recall.py
@@ -464,11 +464,57 @@ async def test_recall_notification_has_no_numbers(monkeypatch):
     text = mine[0]["title"] + "".join(f"{k}{v}" for k, v in mine[0]["lines"])
     for number in ("170.5", "208.5", "38.0", "38.00"):
         assert number not in text, f"消息里不放数：{number}"
-    assert dict(mine[0]["lines"]) == {"状态": "待处理", "口径": "「周销售」", "版本": "v1 → v2", "答案": "1 个",
+    assert dict(mine[0]["lines"]) == {"状态": "待处理", "口径": "「周销售」", "版本": "v1 → v2", "答案": "问数回答 1 个",
                                       "召回": dict(mine[0]["lines"])["召回"]}
     assert "周销售" in mine[0]["title"] and "v2" in mine[0]["title"]
 
 
+async def test_recall_message_counts_answers_by_kind(monkeypatch):
+    """召回消息按种类分开写条数：一个问数回答、一次正式运行写「问数回答 1 个、正式运行 1 个」（下一轮验收：以前写「答案：2 个」，
+    读的人以为是两个问数回答）。只写条数，照旧不放数。"""
+    from app.core.crypto import encrypt
+    from app.db.models import NotifyChannel
+    from app.engine.launch import launch_run
+    from app.notify import service
+
+    sent: list[dict] = []
+
+    async def capture(channel, message):
+        sent.append({"channel": channel.name, **message})
+        from app.notify.core import Outcome
+
+        return Outcome(True, 200, 1, None)
+
+    monkeypatch.setattr("app.notify.send", capture)
+    async with SessionLocal() as session:
+        row = NotifyChannel(name="召回-按种类", scope="global", url=encrypt("https://hooks.example.com/recall-kinds"),
+                            events=["ledger.recall"])
+        session.add(row)
+        await session.commit()
+        channel_id = row.id
+    try:
+        wf = await S.publish()
+        await answer(monkeypatch)
+        S.script(monkeypatch, "本周销售额 [[m:gmv]]。")
+        async with SessionLocal() as session:
+            formal = await launch_run(session, workflow_id=wf, run_class="formal", version=1, input_payload=WEEK1,
+                                      actor="analyst")
+        assert (await S.wait(formal.id)).status == "succeeded"
+        await publish_v2(wf)
+        await run_recall(wf)
+        await service.drain()
+    finally:
+        async with SessionLocal() as session:
+            await session.delete(await session.get(NotifyChannel, channel_id))
+            await session.commit()
+    mine = [m for m in sent if m["channel"] == "召回-按种类"]
+    assert mine, sent
+    answers = [v for m in mine for k, v in m["lines"] if k == "答案"]
+    assert "问数回答 1 个、正式运行 1 个" in answers, answers
+    for number in ("170.5", "208.5", "38.0"):
+        assert all(number not in v for m in mine for _, v in m["lines"]), "消息里不放数"
+
+
 async def test_recall_pushes_the_shared_channel_once_per_recall_run(monkeypatch):
     """开登录时一组答案有几个主人就开几条待办，但全局渠道（团队群）按召回运行只收一条（答案、待办都按这次召回汇总，链到
     待办箱）；订阅了 task.opened 的全局渠道也不再按每条待办各推一次。每个主人自己的个人渠道各收他那一条。"""
@@ -522,12 +568,12 @@ async def test_recall_pushes_the_shared_channel_once_per_recall_run(monkeypatch)
     team = [m for m in sent if m["channel"] == "团队-召回"]
     assert [m["event"] for m in team] == ["ledger.recall"], [m["event"] for m in team]
     lines = dict(team[0]["lines"])
-    assert lines["答案"] == "2 个" and lines["待办"] == "2 条", lines
+    assert lines["答案"] == "问数回答 2 个" and lines["待办"] == "2 条", lines
     assert [label for label, _ in team[0]["links"]] == ["去待办箱"] or team[0]["links"] == []
     assert [m["event"] for m in sent if m["channel"] == "团队-待办"] == ["task.opened"], "只订阅了新待办的团队渠道也只收一条"
     for name in ("analyst_a", "analyst_b"):
         mine = [m for m in sent if m["channel"] == f"个人-{name}"]
-        assert [m["event"] for m in mine] == ["ledger.recall"] and dict(mine[0]["lines"])["答案"] == "1 个", mine
+        assert [m["event"] for m in mine] == ["ledger.recall"] and dict(mine[0]["lines"])["答案"] == "问数回答 1 个", mine
     # 账本接口里的召回待办只列看的这个人手上的：第一个答案的主人看得到自己那条，另一个主人看不到
     from app.api.deps import Principal
 
diff --git a/backend/tests/test_profile_dialect_golden.py b/backend/tests/test_profile_dialect_golden.py
index ebdd546e..ffd9b366 100644
--- a/backend/tests/test_profile_dialect_golden.py
+++ b/backend/tests/test_profile_dialect_golden.py
@@ -69,7 +69,7 @@ def _queries(kind: str) -> list[tuple[str, str | None]]:
                                            (d.column(META["orders"], "biz_day"), True)])),
         ("date_window", d.window_value_counts_sql(orders, status, d.column(META["orders"], "paid_at"),
                                                   "2026-08-27", 21)),
-        ("day_volume", d.day_volume_sql(orders, d.column(META["orders"], "paid_at"), "2026-08-27")),
+        ("day_volume", d.day_volume_sql(orders, d.column(META["orders"], "paid_at"), "2026-08-27", "2026-10-01")),
         ("sample", d.sample_value_counts_sql(orders, status, 1, 21)),
         # 第 3 波：取值观察的最新业务日期（文字存的日期按字符串比）
         ("max_date", d.max_date_sql(orders, d.column(META["orders"], "paid_at"))),
diff --git a/backend/tests/test_run_rejected.py b/backend/tests/test_run_rejected.py
index 1da7e538..18f5cf35 100644
--- a/backend/tests/test_run_rejected.py
+++ b/backend/tests/test_run_rejected.py
@@ -186,6 +186,35 @@ async def test_the_api_carries_end_reason_and_refuses_to_continue_a_rejected_run
     assert stop.status_code == 409 and "被驳回" in stop.json()["detail"]
 
 
+async def test_resume_on_a_rejected_run_says_rejected_with_the_same_code(client):
+    """对被驳回的运行调 /resume：照 end_reason 说「被驳回」，和 /continue 同一个错误码 run_rejected（下一轮验收：以前说
+    「这次运行已取消…已取消的运行请重新发起」）。手动停止的照旧说已取消；存量的 failed + 驳回也说被驳回，不再叫人点
+    「继续运行」。"""
+    run_id = await _rejected_run("approve")
+    await _wait(run_id, ("cancelled",))
+    r = await client.post(f"/api/runs/{run_id}/resume", json={"response": {"approved": True}})
+    assert r.status_code == 409 and r.json()["code"] == "run_rejected", r.text
+    assert "被驳回" in r.json()["detail"] and "已取消" not in r.json()["detail"]
+    async with SessionLocal() as session:
+        stopped = Run(workflow_name="停掉的", status="cancelled", end_reason="stopped", graph=chain(START, OUT), input={})
+        old = Run(workflow_name="老的驳回", status="failed", graph=chain(START, GATES["approve"][0], OUT), input={},
+                  error="人工驳回：不对", error_node_id="gate")
+        session.add_all([stopped, old])
+        await session.flush()
+        session.add_all([
+            RunEvent(run_id=old.id, seq=1, type="node.failed", node_id="gate", ts=1.0, data={"error": "人工驳回：不对"}),
+            Task(run_id=old.id, node_id="gate", kind="approve", title="可以吗", status="resolved",
+                 outcome={"decision": "reject"}),
+        ])
+        await session.commit()
+        stopped_id, old_id = stopped.id, old.id
+    other = await client.post(f"/api/runs/{stopped_id}/resume", json={"response": {"approved": True}})
+    assert other.status_code == 409 and other.json().get("code") != "run_rejected" and "已取消" in other.json()["detail"]
+    legacy = await client.post(f"/api/runs/{old_id}/resume", json={"response": {"approved": True}})
+    assert legacy.status_code == 409 and legacy.json()["code"] == "run_rejected", legacy.text
+    assert "继续运行" not in legacy.json()["detail"].replace("不能继续运行", "")
+
+
 async def test_a_legacy_failed_rejection_cannot_be_continued_either(client):
     """加 end_reason 之前被驳回的运行记的是 failed（封存不改写）：失败的那一步上最近一条处理完的待办是驳回，也不许接着跑。
     别的失败照旧能接着跑（这里断点不在，说的是没有断点，不是被驳回）。"""
diff --git a/backend/tests/test_shares.py b/backend/tests/test_shares.py
index 28d110d0..49738185 100644
--- a/backend/tests/test_shares.py
+++ b/backend/tests/test_shares.py
@@ -326,6 +326,10 @@ async def test_revoked_and_expired_links_say_so_with_410(world):  # noqa: F811
     admin = await auth_env.login("root")
     r = await admin.delete(f"/api/apps/{app_id}/shares/{gone['id']}")
     assert r.status_code == 200 and r.json()["state"] == "revoked" and r.json()["revoked_by"] == "root"
+    # 撤销交回的那一条和建、列的同一个形状：page_title 是页面标题，不是页面 id（下一轮验收：以前撤销回的是「home」）
+    assert gone["page_title"] == "首页" and r.json()["page_title"] == gone["page_title"], r.json()["page_title"]
+    again = await admin.delete(f"/api/apps/{app_id}/shares/{gone['id']}")
+    assert again.json()["page_title"] == "首页", "撤销过的再撤销照旧交回它，标题一样"
     async with SessionLocal() as s:
         (await s.get(ViewShare, old["id"])).expires_at = utcnow() - timedelta(minutes=5)
         await s.commit()
diff --git a/backend/tests/test_source_maintenance.py b/backend/tests/test_source_maintenance.py
index 468c580c..73f76413 100644
--- a/backend/tests/test_source_maintenance.py
+++ b/backend/tests/test_source_maintenance.py
@@ -111,6 +111,32 @@ async def test_fire_builds_introspect_then_profile_in_order(client, tmp_path, ma
         await drop_source(sid)
 
 
+async def test_observing_again_within_the_window_records_why_in_the_fire(client, tmp_path, made):
+    """只选取值观察：12 小时内观察过的表不再观察，任务以 0 张表结束。建任务时就判断得出（这次没有探查、也没排在起草后面，
+    开始时表结构、目录不会变）：原因记进触发记录的 inputs.skipped 和 reason（下一轮验收：以前只在剖析任务的提醒里）。
+    任务照建（触发照旧是 started，不当成没能发起）。"""
+    sid = await _source(client, tmp_path / "c.db")
+    try:
+        async with SessionLocal() as session:
+            row = await session.get(DataSource, sid)
+            row.options = sm.with_steps(row.options, ["observe"], actor="测试员")
+            await session.commit()
+        schedule_id = await _schedule(sid)
+        made.append(schedule_id)
+        at = utcnow().replace(microsecond=0)
+        first = await service.fire_once(schedule_id, at, "manual")
+        assert first.status == "started" and not first.inputs.get("skipped"), first.inputs
+        await _wait_jobs(first)
+        second = await service.fire_once(schedule_id, at + timedelta(seconds=1), "manual")
+        await _wait_jobs(second)
+        assert second.status == "started" and [j["tables"] for j in second.inputs["jobs"]] == [0]
+        [skipped] = second.inputs["skipped"]
+        assert skipped["step"] == "observe" and "12 小时内都观察过" in skipped["reason"], skipped
+        assert "取值观察没有执行" in second.reason
+    finally:
+        await drop_source(sid)
+
+
 async def test_profile_off_records_steps_not_run_and_still_introspects(client, tmp_path, made):
     sid = await _source(client, tmp_path / "b.db", profile=False)
     try:
diff --git a/backend/tests/test_sql_semantics.py b/backend/tests/test_sql_semantics.py
index 7a9bbf93..0b9f35cf 100644
--- a/backend/tests/test_sql_semantics.py
+++ b/backend/tests/test_sql_semantics.py
@@ -616,10 +616,10 @@ def test_specific_differences_are_said_in_one_sentence_each(a, b, expected):
 
 
 @pytest.mark.parametrize("a,b", [
-    # 同一列上的时间范围不同、同一列上换了筛选的值：不在能说的四类里
-    (BASE, "SELECT SUM(amount) FROM orders WHERE paid_at BETWEEN '2026-09-01' AND '2026-09-30'"),
-    ("SELECT SUM(amount) FROM orders WHERE status = 'paid'",
-     "SELECT SUM(amount) FROM orders WHERE status IN ('paid', 'shipped')"),
+    # 算出来的值上换了筛选的值：认不出取值范围，不在能说的五类里（光一列上的条件改写见
+    # test_rewritten_conditions_on_one_column_say_both_sides）
+    ("SELECT SUM(amount) FROM orders WHERE LOWER(status) = 'paid'",
+     "SELECT SUM(amount) FROM orders WHERE LOWER(status) IN ('paid', 'shipped')"),
     # 聚合的参数不同、选出来的列换了先后
     ("SELECT SUM(paid_amount) FROM orders", "SELECT SUM(amount) FROM orders"),
     ("SELECT SUM(o.amount) FROM orders o JOIN refunds r ON r.order_id = o.id",
@@ -653,8 +653,8 @@ def test_specific_differences_are_said_in_one_sentence_each(a, b, expected):
      "WITH t AS (SELECT user_id, SUM(amount) AS total FROM orders GROUP BY user_id) SELECT COUNT(*) FROM t WHERE total > 100"),
 ], ids=lambda v: None)
 def test_differences_outside_the_four_kinds_get_only_the_fallback(a, b):
-    """具体说法只说四类（按哪一列筛时间、一边多出来的筛选条件、顶层的聚合函数和去重、关联的表和关联类型），而且只在
-    两边都是简单形状时说。别的不同（同一列上的范围、换了筛选的值、聚合的参数、分组、HAVING、取行、排序、关联条件、
+    """具体说法只说五类（按哪一列筛时间、一边多出来的筛选条件、同一列上的条件改写、顶层的聚合函数和去重、关联的表和
+    关联类型），而且只在两边都是简单形状时说。别的不同（算出来的值上换了筛选的值、聚合的参数、分组、HAVING、取行、排序、关联条件、
     UNION、CTE 和子查询里的不同）以前说得出具体的，复验在这些地方一再抓到不存在的差异（被 WHERE 定死的排序键、恒为真的
     HAVING、外层不在乎的子查询差异……），现在一律只给兜底说明，附规范化后不同的行。"""
     for dialect in DIALECTS:
@@ -663,6 +663,94 @@ def test_differences_outside_the_four_kinds_get_only_the_fallback(a, b):
         assert not diff.equivalent and diff.items == [FacetDiff("other", None, None, FALLBACK)], (dialect, diff.items)
 
 
+@pytest.mark.parametrize("a,b,aspect,text", [
+    # 等号对 IN：口径 status = 'paid'，这次多收了 refunded（账本的叫法：A 是这次、B 是口径）
+    ("SELECT SUM(amount) FROM orders WHERE status IN ('paid', 'refunded')",
+     "SELECT SUM(amount) FROM orders WHERE status = 'paid'", "filter",
+     "这次对 status 的条件是 status IN ('paid', 'refunded')，口径是 status = 'paid'（这次多了 'refunded'）"),
+    # 取值集合不同：各多了几个
+    ("SELECT SUM(o.amount) FROM orders o JOIN stores s ON s.id = o.store_id WHERE o.status IN ('paid', 'open', 'done')",
+     "SELECT SUM(o.amount) FROM orders o JOIN stores s ON s.id = o.store_id WHERE o.status IN ('paid', 'void')", "filter",
+     "这次对 orders.status 的条件是 orders.status IN ('done', 'open', 'paid')，口径是 orders.status IN ('paid', 'void')"
+     "（这次多了 'done'、'open'，口径多了 'void'）"),
+    # 取值毫不相干：前一句已经说全了，不补
+    ("SELECT COUNT(*) FROM orders WHERE status = 'paid'", "SELECT COUNT(*) FROM orders WHERE status = 'done'", "filter",
+     "这次对 status 的条件是 status = 'paid'，口径是 status = 'done'"),
+    # 范围的端点不同
+    ("SELECT COUNT(*) FROM orders WHERE amount >= 100", "SELECT COUNT(*) FROM orders WHERE amount > 100", "filter",
+     "这次对 amount 的条件是 amount >= 100，口径是 amount > 100"),
+    # 同一时间列上的范围不同：time_range
+    (BASE, "SELECT SUM(amount) FROM orders WHERE paid_at >= '2026-09-01' AND paid_at < '2026-11-01'", "time_range",
+     "这次对 paid_at 的条件是 paid_at >= '2026-09-01' AND paid_at < '2026-10-01'，"
+     "口径是 paid_at >= '2026-09-01' AND paid_at < '2026-11-01'"),
+], ids=lambda v: None)
+def test_rewritten_conditions_on_one_column_say_both_sides(a, b, aspect, text):
+    """同一列上的条件两边各写了一套、留下的取值不同（下一轮验收：口径 status = 'paid'，这次 status IN ('paid',
+    'refunded') 以前只给兜底）：说「这次对 X 的条件是 …，口径是 …」，aspect 两边都有（dual_check_facts 不会把它当成一边
+    多出来的有效行条件）。各方言、给不给表结构都一样。"""
+    for dialect in DIALECTS:
+        for schema in (None, sql_pairs.schema_cache()):
+            diff = explain_sql_diff(a, b, dialect, schema=schema, labels=("这次", "口径"))
+            check_shape(diff)
+            assert not diff.equivalent and not diff.fallback, (dialect, diff.items)
+            [item] = diff.items
+            assert (item.aspect, item.text) == (aspect, text), (dialect, item)
+            assert item.a is not None and item.b is not None
+            assert item.detail == {"columns": [{"table": "orders", "column": text.split("对 ")[1].split(" ")[0]
+                                                .split(".")[-1]}]}
+
+
+@pytest.mark.parametrize("a,b,schema", [
+    # 冗余的写法：算出来留下的取值一样，不说
+    ("SELECT COUNT(*) FROM orders WHERE amount >= 100 AND amount >= 50", "SELECT COUNT(*) FROM orders WHERE amount >= 100",
+     None),
+    # 主键不为空：差的只是空值那一段，不说（没给表结构时照说，那时不知道它不为空）
+    ("SELECT COUNT(*) FROM orders WHERE id = 1 OR id IS NULL", "SELECT COUNT(*) FROM orders WHERE id = 1", "schema"),
+    # WHERE 之外还用到了这一列：SUM(CASE WHEN status = 'paid' …) 配上 IN (…) 和只筛 'paid' 结果一样
+    ("SELECT SUM(CASE WHEN status = 'paid' THEN amount END) FROM orders WHERE status IN ('paid', 'refunded')",
+     "SELECT SUM(amount) FROM orders WHERE status = 'paid'", None),
+    # 和别的列交错的条件
+    ("SELECT COUNT(*) FROM orders WHERE status = 'paid' OR amount > 100",
+     "SELECT COUNT(*) FROM orders WHERE status IN ('paid', 'done') OR amount > 100", None),
+], ids=lambda v: None)
+def test_rewritten_conditions_that_may_not_change_the_result_are_not_said(a, b, schema):
+    for dialect in DIALECTS:
+        diff = explain_sql_diff(a, b, dialect, schema=sql_pairs.schema_cache() if schema else None)
+        check_shape(diff)
+        assert not diff.equivalent and not any("的条件是" in d.text for d in diff.items), (dialect, diff.items)
+
+
+def test_explaining_rewritten_conditions_never_changes_the_equivalence_verdict(monkeypatch):
+    """同一列上的条件改写只进说明：等价与否、在哪一级规范形式上判的，和不说这一类时一字不差（语料各对、两种先后、给不给
+    表结构）；规范化本来就认得的改写（IN 只有一个值、IN 列表去重排序）照样判等价、没有条目。"""
+    from app.data import sql_semantics
+
+    pairs = [(p.a, p.b) for p in sql_pairs.PAIRS] + [
+        ("SELECT SUM(amount) FROM orders WHERE status IN ('paid')", "SELECT SUM(amount) FROM orders WHERE status = 'paid'"),
+        ("SELECT COUNT(*) FROM orders WHERE status IN ('void', 'paid', 'paid')",
+         "SELECT COUNT(*) FROM orders WHERE status IN ('paid', 'void')"),
+        ("SELECT SUM(amount) FROM orders WHERE status IN ('paid', 'refunded')",
+         "SELECT SUM(amount) FROM orders WHERE status = 'paid'"),
+    ]
+    schema = sql_pairs.schema_cache()
+
+    def verdicts() -> list[tuple[bool, str | None]]:
+        out = []
+        for a, b in pairs:
+            for x, y in ((a, b), (b, a)):
+                for sc in (None, schema):
+                    diff = explain_sql_diff(x, y, "sqlite", schema=sc)
+                    out.append((diff.equivalent, diff.level))
+        return out
+
+    with_rewrites = verdicts()
+    monkeypatch.setattr(sql_semantics._Explainer, "_rewrites", lambda self, A, B, xa, xb, oa, ob, m: (oa, ob))
+    assert verdicts() == with_rewrites
+    for a, b in pairs[-3:-1]:
+        diff = explain_sql_diff(a, b, "sqlite")
+        assert diff.equivalent and diff.items == [], (a, b)
+
+
 def test_a_difference_inside_a_cte_used_only_in_a_condition_subquery_is_never_missed():
     """第 5 轮回退的例子：只在 IN 子查询里用到的 CTE 里的条件不同，以前整个漏报（返回空列表）。现在条件里有子查询的不是
     简单形状，不说具体的，但一定不判等价、给兜底说明和规范化后不同的行。EXISTS、标量子查询一样。"""
@@ -754,9 +842,9 @@ def test_nothing_specific_gives_one_fallback_item_and_the_lines_that_differ():
 
 
 def test_specific_items_that_cover_only_part_end_with_the_rest_note():
-    """按哪一列筛时间说得准，同一列 status 上换了筛选的值说不了：说时间那一句，补一句「另有未能归纳的差异」。"""
-    a = "SELECT SUM(amount) FROM orders WHERE paid_at >= '2026-09-01' AND status = 'paid'"
-    b = "SELECT SUM(amount) FROM orders WHERE created_at >= '2026-09-01' AND status IN ('paid', 'done')"
+    """按哪一列筛时间说得准，LOWER(status) 上换了筛选的值说不了：说时间那一句，补一句「另有未能归纳的差异」。"""
+    a = "SELECT SUM(amount) FROM orders WHERE paid_at >= '2026-09-01' AND LOWER(status) = 'paid'"
+    b = "SELECT SUM(amount) FROM orders WHERE created_at >= '2026-09-01' AND LOWER(status) IN ('paid', 'done')"
     diff = explain_sql_diff(a, b, labels=LABELS)
     check_shape(diff)
     assert diff.fallback and [d.text for d in diff.items] == ["原流程按 paid_at 筛时间，盲算按 created_at", FALLBACK_REST]
diff --git a/backend/tests/test_sql_text_diff.py b/backend/tests/test_sql_text_diff.py
index f49f670d..87a9680d 100644
--- a/backend/tests/test_sql_text_diff.py
+++ b/backend/tests/test_sql_text_diff.py
@@ -73,11 +73,11 @@ def test_an_extra_condition_sorted_first_does_not_drag_the_shared_ones_along():
 
 
 def test_two_conditions_on_each_side_mark_only_the_one_that_differs():
-    """同一列上的区间只差结束日期（这种只给兜底说明，审核人全靠逐行对照）：开始日期那一行两边一样，不标。"""
+    """同一列上的区间只差结束日期（差异说明说成 time_range，逐行对照照样给）：开始日期那一行两边一样，不标。"""
     a = f"SELECT SUM(fee) FROM parcels WHERE {WEEK}"
     b = "SELECT SUM(fee) FROM parcels WHERE sent_on BETWEEN '2031-03-03' AND '2031-03-10'"
     got = diff(a, b)
-    assert got.fallback
+    assert not got.fallback and [i.aspect for i in got.items] == ["time_range"]
     assert got.text_diff == ("  WHERE", "-   AND parcels.sent_on <= '2031-03-09'", "+   AND parcels.sent_on <= '2031-03-10'")
 
 
diff --git a/backend/tests/test_usage_meter.py b/backend/tests/test_usage_meter.py
index ed0c35b5..828a8130 100644
--- a/backend/tests/test_usage_meter.py
+++ b/backend/tests/test_usage_meter.py
@@ -105,6 +105,85 @@ async def test_metering_counts_by_model_and_prices_from_the_catalog():
     assert got["cost_usd"] == pytest.approx(estimate_cost(PRICED, 2000, 400))
 
 
+class Bare(Scripted):
+    """不报调用参数的模型（同以前的演示模型）：invocation_params 里没有 model，只在对象上有。"""
+
+    @property
+    def _identifying_params(self) -> dict[str, Any]:
+        return {}
+
+
+async def test_the_model_name_comes_from_the_model_object_when_the_call_parameters_lack_it():
+    """评测汇总里助手用量的模型以前记成 unknown（下一轮验收）：演示模型的调用参数里没有模型名，计量也就认不出。现在
+    调用参数里没有时看 LangChain 照模型对象给回调的 ls_model_name；演示模型自己也把 model_name 报进调用参数。"""
+    from app.providers.mock_model import MockChatModel
+
+    m = usage.meter()
+    with usage.metering(m):
+        await Bare(model=PRICED + "-bare").ainvoke("你好")
+        await MockChatModel(model_name="mock-wide", latency=0).ainvoke("你好")
+    got = m.summary()["by_model"]
+    assert set(got) == {PRICED + "-bare", "mock-wide"}, got
+    assert usage.UNKNOWN_MODEL not in got
+
+
+@pytest.mark.parametrize("build", ["openai", "anthropic", "mock"])
+def test_the_call_parameters_of_each_provider_kind_carry_the_configured_model(build):
+    """三种接入（OpenAI 兼容、Anthropic、演示模型）由 providers/factory 建出来的模型：LangChain 回调时给的调用参数和
+    ls_model_name（这里照它的写法直接取，不发请求）计量都认得出，记的就是接入里配的模型 id。"""
+    import uuid as _uuid
+    from types import SimpleNamespace
+
+    from langchain_core.outputs import LLMResult
+
+    from app.providers.factory import ModelSpec, build_chat_model
+
+    provider = SimpleNamespace(kind=build, name="虚构接入", api_key=None, base_url=None, extra={},
+                               default_model=None)
+    model_id = {"openai": "gpt-made-up", "anthropic": "claude-made-up", "mock": "mock-wide"}[build]
+    import os
+    env = {"openai": "OPENAI_API_KEY", "anthropic": "ANTHROPIC_API_KEY"}.get(build)
+    old = os.environ.get(env) if env else None
+    if env:
+        os.environ[env] = "sk-made-up"
+    try:
+        model = build_chat_model(provider, ModelSpec(model=model_id))
+    finally:
+        if env:
+            if old is None:
+                os.environ.pop(env, None)
+            else:
+                os.environ[env] = old
+    m = usage.meter()
+    run_id = _uuid.uuid4()
+    m.on_chat_model_start({}, [], run_id=run_id, invocation_params=model._get_invocation_params(),
+                          metadata=dict(model._get_ls_params()))
+    m.on_llm_end(LLMResult(generations=[[ChatGeneration(message=AIMessage(content="好"))]]), run_id=run_id)
+    assert list(m.summary()["by_model"]) == [model_id]
+    # 调用参数里本身就带着（不靠 ls_model_name 兜着）
+    params = model._get_invocation_params()
+    assert model_id in (params.get("model"), params.get("model_name")), params
+
+
+async def test_the_eval_driver_on_the_demo_model_records_its_name(monkeypatch):
+    """评测作答走演示模型时，助手用量记在演示模型的名下，不是 unknown（下一轮验收：写作用量的模型对、助手的记成 unknown）。"""
+    import app.api.copilot as copilot
+    from app.eval import answer
+    from app.providers.mock_model import MockChatModel
+
+    async def _model(*_a, **_k):
+        return MockChatModel(model_name="mock-fast", latency=0), "mock-fast"
+
+    monkeypatch.setattr(copilot, "get_chat_model", _model)
+    driver = answer.Driver("按天看销售额", datasource_ids=[], ctx=answer.TurnContext(
+        suite_id="s", suite_name="虚构题集", eval_run_id="e", case_id=None, case_seq=1, actor=None))
+    ops, refused = await driver._generate("按天看销售额", None)
+    assert refused is None
+    driver._take(ops)
+    got = driver.turn.usage["assistant"]["by_model"]
+    assert set(got) == {"mock-fast"} and got["mock-fast"]["calls"] >= 1, got
+
+
 async def test_outside_the_meter_nothing_is_counted():
     m = usage.meter()
     with usage.metering(m):
diff --git a/backend/tests/test_value_drift.py b/backend/tests/test_value_drift.py
index 6fa80ef9..9d45e446 100644
--- a/backend/tests/test_value_drift.py
+++ b/backend/tests/test_value_drift.py
@@ -98,6 +98,33 @@ def test_volume_drop_against_the_previous_28_days():
     assert vd.daily(partial, [], today=TODAY) == []
 
 
+def test_future_business_dates_do_not_turn_today_into_the_last_full_day():
+    """业务日期有未来值（最晚到明天）：最近一个整天仍是昨天，今天这个还没写完的日子不算量跌；昨天真跌了照样报。以前取
+    max(days) 再退一天，max 在明天时就落到了今天（只有几行）上，记出一条量跌。"""
+    start = TODAY - timedelta(days=30)                      # start + 30 是今天
+    ahead = _days(start, [100] * 30 + [5, 40])              # 昨天 100、今天 5 行、明天 40 行
+    assert not [f for f in vd.daily(ahead, [], today=TODAY) if f.kind == "volume_drop"]
+    [f] = [f for f in vd.daily(_days(start, [100] * 29 + [20, 5, 40]), [], today=TODAY) if f.kind == "volume_drop"]
+    assert f.after["day"] == (TODAY - timedelta(days=1)).isoformat() and f.after["rows"] == 20
+
+
+def test_days_after_today_are_not_gaps_now_or_in_history():
+    """缺天只看到今天为止：今天以后两天没数据、第三天又有（预约、计划类的日期）不算缺天；以前那次观察窗口里它自己的「今天」
+    以后的空日子也不算以前有过缺天，挡不住这一次真缺的天。"""
+    start = TODAY - timedelta(days=30)
+    full = _days(start - timedelta(days=1), [100] * 31)
+    ahead = _days(start, [100] * 31 + [0, 0, 3])            # 今天之后空两天、第三天 3 行
+    assert not [f for f in vd.daily(ahead, [full], today=TODAY) if f.kind == "date_gap"]
+    then = TODAY - timedelta(days=7)
+    old = vd.Obs("daily_volume", "window:paid_on:35", {
+        "until": then.isoformat(),
+        "days": {**{(then - timedelta(days=i)).isoformat(): 100 for i in range(30)},
+                 (then + timedelta(days=5)).isoformat(): 2}}, AT)  # 那次的「今天」以后空了几天、又有一行
+    holed = _days(start, [100] * 10 + [0, 0] + [100] * 19)
+    [gap] = [f for f in vd.daily(holed, [old], today=TODAY) if f.kind == "date_gap"]
+    assert gap.after["gaps"] == 2
+
+
 def _rows(n: int, method: str = "count") -> vd.Obs:
     return vd.Obs("rows", method, {"rows": n, "method": method}, AT)
 
diff --git a/frontend/src/api/client.ts b/frontend/src/api/client.ts
index 63dc10ed..13c722d2 100644
--- a/frontend/src/api/client.ts
+++ b/frontend/src/api/client.ts
@@ -530,7 +530,7 @@ export const api = {
        * 409 带机读码：profile_disabled、datasource_inactive、schema_missing、snapshot_tampered、profile_busy（这个源已有排队、
        * 进行中的剖析任务，detail 写明进行到第几张）、profile_nothing（按范围挑出 0 张表，detail 写明原因和建议）
        */
-      create: (sourceId: string, body: { tables?: string[]; scope?: 'needed' | 'all' }) =>
+      create: (sourceId: string, body: { tables?: string[]; scope?: 'needed' | 'all' | 'observe' }) =>
         post<{ job: CatalogProfileJob }>(`${catalogBase(sourceId)}/profile-jobs`, body),
       /** 排队或进行中的那个剖析任务；没有为 null */
       active: (sourceId: string, opts?: RequestOptions) =>
diff --git a/frontend/src/dev/fixtures/dual_check.json b/frontend/src/dev/fixtures/dual_check.json
index 6024462e..e2003b81 100644
--- a/frontend/src/dev/fixtures/dual_check.json
+++ b/frontend/src/dev/fixtures/dual_check.json
@@ -843,8 +843,8 @@
     },
     "fallback": {
       "sql": [
-        "SELECT SUM(amount) AS gmv\nFROM orders\nWHERE paid_at >= '2026-09-01'\n  AND paid_at < '2026-09-08'",
-        "SELECT SUM(amount) AS gmv\nFROM orders\nWHERE paid_at >= '2026-09-01'\n  AND paid_at < '2026-09-09'"
+        "SELECT SUM(amount) AS gmv\nFROM orders\nWHERE paid_at >= '2026-09-01'\n  AND DATE(paid_at) < '2026-09-08'",
+        "SELECT SUM(amount) AS gmv\nFROM orders\nWHERE paid_at >= '2026-09-01'\n  AND DATE(paid_at) < '2026-09-09'"
       ],
       "diff": {
         "equivalent": false,
@@ -857,8 +857,8 @@
         "fallback": true,
         "text_diff": [
           "  WHERE",
-          "-   AND orders.paid_at < '2026-09-08'",
-          "+   AND orders.paid_at < '2026-09-09'"
+          "-   AND DATE(orders.paid_at) < '2026-09-08'",
+          "+   AND DATE(orders.paid_at) < '2026-09-09'"
         ]
       },
       "catalog": {
diff --git a/frontend/src/lib/terms.ts b/frontend/src/lib/terms.ts
index a1143961..cd426a7a 100644
--- a/frontend/src/lib/terms.ts
+++ b/frontend/src/lib/terms.ts
@@ -3035,6 +3035,11 @@ export const PROFILE_TEXT = {
   scopeNeededNone: '没有需要剖析的表：都剖析过、上次都做完了，也没有新的待核实关联关系和新的码值、日期列',
   scopeAll: (n: number) => `全部 ${formatNumber(n)} 张表`,
   scopeAllHint: '有可剖析项的表，不含视图',
+  /** 取值观察（范围 observe）：只观察、不剖析。张数是有数据目录的表，其中最近观察过的服务端这次不观察 */
+  scopeObserve: (n: number) => `取值观察：有数据目录的 ${formatNumber(n)} 张表`,
+  scopeObserveHint: (hours: number) => '只观察、不剖析：码值分布、最新业务日期、按日行数和行数，和同口径的历史比出取值变化，'
+    + `列在「变化」里。最近 ${formatNumber(hours)} 小时内观察过的表这次不再观察`,
+  scopeObserveNone: '还没有有数据目录的表：取值观察只看有数据目录的表，请先起草数据目录',
   /** 已选、筛选结果里有剖析不了的表：不交给服务端，写明有几张 */
   scopeLeftOut: (n: number) => `另有 ${formatNumber(n)} 张不剖析（视图、表结构中已没有，或没有可剖析的项）`,
   scopeEmpty: '这个范围里没有可剖析的表',
@@ -3230,6 +3235,7 @@ export const PROFILE_TEXT = {
     busy: '这个数据源已有剖析任务',
     tampered: '数据文件核对未通过',
     nothing: '没有可剖析的表',
+    nothingObserve: '没有要观察的表',
     other: '无法开始剖析',
   } as Record<string, string>,
   blockedNext: {
@@ -3239,6 +3245,7 @@ export const PROFILE_TEXT = {
     busy: '同一个数据源同一时间只有一个剖析任务：可以查看它的进度，等它结束后再开始新的剖析。',
     tampered: '请重新上传表格，或联系管理员核对数据文件。',
     nothing: '可以改选「全部表」，或在表清单中选定要剖析的表。',
+    nothingObserve: '观察过的表要过一段时间才再观察；要剖析，可以回到上一步改选「需要剖析的表」或「全部表」。',
     other: '请稍后重试。',
   } as Record<string, string>,
   busyAt: (who: string | null, done: number, total: number) =>
diff --git a/frontend/src/lib/termsEval.ts b/frontend/src/lib/termsEval.ts
index 04e268cd..348f9404 100644
--- a/frontend/src/lib/termsEval.ts
+++ b/frontend/src/lib/termsEval.ts
@@ -411,7 +411,8 @@ export const EVAL_TEXT = {
     traced: '可追溯准确率',
     tracedHint: '必需的期望全部经引用对上（答案里的数点得开、来自查询的那一格或口径卡的指标）',
     verdicts: '结论',
-    notRun: (n: number) => `${formatNumber(n)} 项没评（停下、出错、到了上限）`,
+    notRun: (n: number) => `${formatNumber(n)} 项没评（停止、中断，或者到了上限、止损之后没开评）`,
+    errors: (n: number) => `${formatNumber(n)} 项出错`,
     cost: '花费',
     costHint: '助手（建图、挑表、复核）+ 写作（答案运行里的模型调用）+ 结论裁判；算期望不调模型',
     perCase: (v: string) => `每题约 ${v}`,
@@ -473,6 +474,8 @@ export const EVAL_TEXT = {
     cleaned: '明细已清理',
     seq: '题',
     question: '问法',
+    /** 逐题表里实际问的问题下面一行：题目原文和实际问的不同（带变量的模板，或者题目之后改过）时写原文 */
+    questionTemplate: (text: string) => `题目原文：${text}`,
     verdict: '结论',
     expected: '期望',
     got: '答案里的数',
diff --git a/frontend/src/pages/catalog/ProfileDialog.tsx b/frontend/src/pages/catalog/ProfileDialog.tsx
index 24d253dc..5eba1092 100644
--- a/frontend/src/pages/catalog/ProfileDialog.tsx
+++ b/frontend/src/pages/catalog/ProfileDialog.tsx
@@ -20,8 +20,8 @@ import {
 } from '../../lib/terms'
 import { STATUS_TONE, StatusChip } from './parts'
 import {
-  PROFILE_MAX_TABLES, PROFILE_RETRYABLE, UNFINISHED_REASONS, coverageText, fillKey, isProfilable, profileBlockOf, profileScopeTables,
-  secondsText, sizeText, spanText,
+  PROFILE_MAX_TABLES, PROFILE_OBSERVE_EVERY_H, PROFILE_RETRYABLE, UNFINISHED_REASONS, coverageText, fillKey, isProfilable, profileBlockOf,
+  profileScopeTables, secondsText, sizeText, spanText,
 } from './profile'
 import type { ProfileBlock } from './profile'
 import { isJobActive } from './draftJob'
@@ -32,16 +32,18 @@ import type { ProfileStartBody } from './profileJob'
 // 数据目录页的「数据剖析」。剖析是服务端的后台任务（profileJob.ts）：这里选范围、看预算、建任务；之后显示页面跟着的
 // 那个任务（进行中的逐表进度、结束后的结果），任务本身的状态在页面上，关掉对话框照常进行。照起草对话框的结构：
 //
-// 开始前：范围（已选 / 当前筛选 / 需要剖析的表（默认）/ 全部表，个数按清单上每行的 profile_due、profile_checks 算）、按表写的
-// 预算和这次最多约几条查询、醒目一句「只读查询、一条一条发、可随时停止」；同一个数据源有起草任务时按钮写「排队剖析」。
+// 开始前：范围（已选 / 当前筛选 / 需要剖析的表（默认）/ 全部表 / 取值观察，个数按清单上每行的 profile_due、profile_checks、
+// 有没有数据目录算；取值观察只观察不剖析，最近观察过的服务端这次不观察，张数是上限）、按表写的预算和这次最多约几条查询、
+// 醒目一句「只读查询、一条一条发、可随时停止」；同一个数据源有起草任务时按钮写「排队剖析」。
 // 进行中：几张 / 共几张、已发几条查询、用时，按表数的真进度条，停止；逐表一行（状态、各类检查的进度、查询数、用时、正在做
 // 哪一项），点一行展开这张表的检查清单、发现和发过的查询。
 // 结束后：汇总（各状态几张、核实了什么）、止损和停止的原因照实写、按「有发现 / 有跳过 / 失败」筛、重试没做完的表；码值的
 // 「填写含义」照旧（填完回到这里，接着下一列）。
-// 开始不了（409）：每种情况一句说法和下一步（未开启就地打开剖析设置；正忙写清是谁的、到哪了，可以去看它的进度）。
+// 开始不了（409）：每种情况一句说法和下一步（未开启就地打开剖析设置；正忙写清是谁的、到哪了，可以去看它的进度；取值观察
+// 挑不出表时说「没有要观察的表」，不给「改选全部表」）。
 // ===========================================================================
 
-type Scope = 'selected' | 'filtered' | 'needed' | 'all'
+type Scope = 'selected' | 'filtered' | 'needed' | 'all' | 'observe'
 type Filter = 'all' | 'findings' | 'skipped' | 'failed'
 const FILTERS: readonly Filter[] = ['all', 'findings', 'skipped', 'failed']
 
@@ -114,6 +116,10 @@ export function ProfileDialog({
       tables: needed, leftOut: 0, body: { scope: 'needed' } })
     const all = profileScopeTables(rows, 'all')
     out.push({ key: 'all', label: PT.scopeAll(all.length), hint: PT.scopeAllHint, tables: all, leftOut: 0, body: { scope: 'all' } })
+    const watched = profileScopeTables(rows, 'observe')
+    out.push({ key: 'observe', label: PT.scopeObserve(watched.length),
+      hint: watched.length ? PT.scopeObserveHint(PROFILE_OBSERVE_EVERY_H) : PT.scopeObserveNone, tables: watched, leftOut: 0,
+      body: { scope: 'observe' } })
     return out
   }, [rows, visible, selected, filtered])
   const [scope, setScope] = useState<Scope>(() => {
@@ -260,8 +266,9 @@ export function ProfileDialog({
           <div className="space-y-1.5" role="radiogroup" aria-label={PT.scope}>
             {scopes.map((s) => {
               const over = !!s.body.tables && s.tables.length > PROFILE_MAX_TABLES
-              const note = over ? PT.scopeTooMany(PROFILE_MAX_TABLES) : !s.tables.length && s.key !== 'needed' ? PT.scopeEmpty
-                : s.leftOut ? PT.scopeLeftOut(s.leftOut) : s.hint
+              const note = over ? PT.scopeTooMany(PROFILE_MAX_TABLES)
+                : !s.tables.length && s.key !== 'needed' && s.key !== 'observe' ? PT.scopeEmpty
+                  : s.leftOut ? PT.scopeLeftOut(s.leftOut) : s.hint
               return (
                 <label key={s.key} className={clsx('flex cursor-pointer items-start gap-2 rounded-md border px-3 py-2 text-xs',
                   picked.key === s.key ? 'border-[var(--accent)] bg-accent-soft' : 'hover:bg-hover')} data-profile-scope={s.key}
@@ -307,7 +314,8 @@ export function ProfileDialog({
       </>
     )
   } else if (phase === 'blocked' && blocked) {
-    body = <Blocked blocked={blocked} onSettings={onSettings} onGoSource={onGoSource} onViewActive={() => void viewActive()}
+    body = <Blocked blocked={blocked} observing={blocked.body.scope === 'observe'} onSettings={onSettings} onGoSource={onGoSource}
+                    onViewActive={() => void viewActive()}
                     onChooseAll={() => { setScope('all'); void begin({ scope: 'all' }) }} onRetry={() => void begin(blocked.body)} />
     footer = (
       <>
@@ -384,32 +392,35 @@ function Budget({ settings: s, sourceName, tables, onSettings }: {
 }
 
 /** 开始不了（409）：是哪种情况、服务端的原话、下一步和对应的按钮 */
-function Blocked({ blocked: b, onSettings, onGoSource, onViewActive, onChooseAll, onRetry }: {
+function Blocked({ blocked: b, observing, onSettings, onGoSource, onViewActive, onChooseAll, onRetry }: {
   blocked: { kind: ProfileBlock; message: string; active: CatalogProfileJob | null }
+  /** 选的是取值观察：挑不出表时说「没有要观察的表」，不给「改选全部表」（那是剖析，不是观察） */
+  observing: boolean
   onSettings: () => void
   onGoSource: () => void
   onViewActive: () => void
   onChooseAll: () => void
   onRetry: () => void
 }) {
+  const nothingObserve = observing && b.kind === 'nothing'
   const action = b.kind === 'disabled'
     ? <button type="button" className="btn btn-sm btn-primary" onClick={onSettings} data-profile-enable="">{PT.enable}</button>
     : b.kind === 'busy'
       ? <button type="button" className="btn btn-sm btn-primary" onClick={onViewActive} data-profile-view-active="">{PT.viewActive}</button>
       : b.kind === 'nothing'
-        ? <button type="button" className="btn btn-sm" onClick={onChooseAll} data-profile-choose-all="">{PT.chooseAll}</button>
+        ? (observing ? null : <button type="button" className="btn btn-sm" onClick={onChooseAll} data-profile-choose-all="">{PT.chooseAll}</button>)
         : b.kind === 'other'
           ? <button type="button" className="btn btn-sm" onClick={onRetry} data-profile-retry-start="">{PT.retryStart}</button>
           : <button type="button" className="btn btn-sm" onClick={onGoSource} data-profile-go-source="">{PT.goSource}</button>
   return (
     <Notice tone={b.kind === 'busy' || b.kind === 'nothing' ? 'info' : 'warn'} attr={{ 'data-profile-blocked': b.kind }}>
-      <p className="font-medium">{PT.blocked[b.kind]}</p>
+      <p className="font-medium">{PT.blocked[nothingObserve ? 'nothingObserve' : b.kind]}</p>
       <p className="mt-0.5 break-words text-dim">{b.message}</p>
       {b.kind === 'busy' && b.active && (
         <p className="tnum mt-0.5 text-dim" data-profile-busy-at="">{PT.busyAt(b.active.created_by, b.active.done, b.active.total)}</p>
       )}
-      <p className="mt-0.5 text-dim">{PT.blockedNext[b.kind]}</p>
-      <div className="mt-2">{action}</div>
+      <p className="mt-0.5 text-dim">{PT.blockedNext[nothingObserve ? 'nothingObserve' : b.kind]}</p>
+      {action && <div className="mt-2">{action}</div>}
     </Notice>
   )
 }
diff --git a/frontend/src/pages/catalog/profile.ts b/frontend/src/pages/catalog/profile.ts
index a6d41fb2..3c502116 100644
--- a/frontend/src/pages/catalog/profile.ts
+++ b/frontend/src/pages/catalog/profile.ts
@@ -167,9 +167,19 @@ export const isProfilable = (r: Pick<CatalogTableRow, 'in_schema' | 'is_view' |
 /** 服务端「需要剖析」的口径：能剖析、并且 profile_due（从没剖析过、上次没做完、之后出现了新的待核实关系） */
 export const isProfileDue = (r: CatalogTableRow): boolean => isProfilable(r) && r.profile_due === true
 
-/** 范围的几张表：needed 需要剖析的、all 全部能剖析的（服务端挑表的顺序：有待核实关系的在前，再按使用次数） */
-export function profileScopeTables(rows: CatalogTableRow[], scope: 'needed' | 'all'): string[] {
-  return rows.filter(scope === 'needed' ? isProfileDue : isProfilable).map((r) => r.table_name)
+/** 能不能做取值观察：在表结构里、不是视图、有数据目录（version > 0）。同服务端 _pick_observing 的前两条；观察到期没有
+ * （上一次观察在 PROFILE_OBSERVE_EVERY_H 小时以前）清单里没有，由服务端判 */
+export const isObservable = (r: Pick<CatalogTableRow, 'in_schema' | 'is_view' | 'version'>): boolean =>
+  r.in_schema && !r.is_view && r.version > 0
+
+/** 上一次观察多久以前才再观察（同服务端 catalog_profile_observe.OBSERVE_EVERY_H） */
+export const PROFILE_OBSERVE_EVERY_H = 12
+
+/** 范围的几张表：needed 需要剖析的、all 全部能剖析的（服务端挑表的顺序：有待核实关系的在前，再按使用次数）、observe 有数据
+ * 目录、可以观察的（其中最近观察过的服务端这次不观察，张数是上限） */
+export function profileScopeTables(rows: CatalogTableRow[], scope: 'needed' | 'all' | 'observe'): string[] {
+  const test = scope === 'needed' ? isProfileDue : scope === 'observe' ? isObservable : isProfilable
+  return rows.filter(test).map((r) => r.table_name)
 }
 
 /** 因为运行原因没查成的项（计入 unfinished、使表成为部分完成）；其余跳过的原因是按规则不查 */
diff --git a/frontend/src/pages/catalog/profileJob.ts b/frontend/src/pages/catalog/profileJob.ts
index 839b8516..55e3888d 100644
--- a/frontend/src/pages/catalog/profileJob.ts
+++ b/frontend/src/pages/catalog/profileJob.ts
@@ -21,7 +21,7 @@ export const PROFILE_POLL_MS = 1500
 
 export const retryableTasks = (j: CatalogProfileJob): CatalogProfileTaskRow[] => j.tasks.filter((t) => PROFILE_RETRYABLE.has(t.status))
 
-export interface ProfileStartBody { tables?: string[]; scope?: 'needed' | 'all' }
+export interface ProfileStartBody { tables?: string[]; scope?: 'needed' | 'all' | 'observe' }
 
 export function useProfileJob(sourceId: string, active: CatalogActiveProfileJob | null | undefined, on: {
   /** 这一次取回来新结束的表（写进目录了，或者失败、停止了）：重取清单，正在看的表在里面就重新载入 */
diff --git a/frontend/src/pages/evals/EvaluationPage.tsx b/frontend/src/pages/evals/EvaluationPage.tsx
index 11d473df..bef39bc5 100644
--- a/frontend/src/pages/evals/EvaluationPage.tsx
+++ b/frontend/src/pages/evals/EvaluationPage.tsx
@@ -221,7 +221,8 @@ function Overview({ run }: { run: EvalRun }) {
       </div>
       <Section title={O.verdicts} attr="verdicts">
         <p>{distText(s.verdicts as Record<string, number>, T.verdict)}</p>
-        {!!s.not_run && <p className="text-dim">{O.notRun(s.not_run)}</p>}
+        {!!s.not_run && <p className="text-dim" data-eval-not-run={s.not_run}>{O.notRun(s.not_run)}</p>}
+        {!!s.errors && <p className="text-dim" data-eval-errors={s.errors}>{O.errors(s.errors)}</p>}
         <p className="text-dim">{O.outcomes}：{distText(s.outcomes as Record<string, number>, T.outcome)}</p>
         {s.text && Object.keys(s.text).length > 0 && <p className="text-dim">{O.text}：{distText(s.text as Record<string, number>, T.textOutcome)}</p>}
       </Section>
@@ -370,7 +371,10 @@ function Results({ id, live, focusCase }: { id: string; live: boolean; focusCase
                     {r.arm && <div className="text-2xs text-faint">{T.run.arm(r.arm)}</div>}
                   </td>
                   <td className="max-w-xs px-2 py-1.5">
-                    <div className="line-clamp-2 break-words">{r.question ?? '—'}</div>
+                    <div className="line-clamp-2 break-words" data-eval-result-question="">{r.question ?? '—'}</div>
+                    {r.question_template && r.question_template !== r.question && (
+                      <div className="line-clamp-2 break-words text-2xs text-faint" data-eval-result-template="">{R.questionTemplate(r.question_template)}</div>
+                    )}
                     {r.path && <div className="text-2xs text-faint">{R.path}：{T.path[r.path]}</div>}
                   </td>
                   <td className="px-2 py-1.5">
diff --git a/frontend/src/typesEval.ts b/frontend/src/typesEval.ts
index 82f851be..39d3e185 100644
--- a/frontend/src/typesEval.ts
+++ b/frontend/src/typesEval.ts
@@ -393,11 +393,14 @@ export interface EvalSpend {
 export type EvalStatus = 'queued' | 'running' | 'succeeded' | 'partial' | 'failed' | 'cancelled' | 'interrupted'
 export type EvalTrigger = 'manual' | 'schedule' | 'pair'
 
+/** 一臂的计数（服务端 runner._counts）：done 是真的跑过的（评完、出错），not_run 是停止、中断、到了上限、止损没跑的，
+ * errors 是出错的（不含没跑的）；done + not_run + running + pending = total */
 export interface EvalCounts {
   total?: number
   done?: number
   running?: number
   pending?: number
+  not_run?: number
   passed?: number
   failed?: number
   ungradable?: number
@@ -428,7 +431,9 @@ export interface EvalSummary {
   done?: number
   cases?: number
   verdicts?: Partial<Record<EvalVerdict, number>>
+  /** 没跑的（停止、中断、到了上限、止损没开评）；出错的另在 errors */
   not_run?: number
+  errors?: number
   accuracy?: EvalRate
   traced_accuracy?: EvalRate
   outcomes?: Partial<Record<EvalOutcome, number>>
@@ -558,7 +563,10 @@ export interface EvalResult {
   error: string | null
   started_at: string | null
   finished_at: string | null
+  /** 这一题实际问的问题（变量按这次评测的截至时间渲染好了）；题目之后改过、又没有作答运行可取时为 null */
   question?: string | null
+  /** 题目的原文（可能带 {{ schedule.… }} 变量的模板）；题目删了为 null */
+  question_template?: string | null
   /** 这一题的花费：助手（含复核）+ 写作 + 结论裁判，服务端加好的 */
   cost?: { cost_usd: number | null; unpriced_calls: number }
 }
diff --git a/scripts/check-catalog-profile.mjs b/scripts/check-catalog-profile.mjs
index 6fac3c71..7d45bde2 100644
--- a/scripts/check-catalog-profile.mjs
+++ b/scripts/check-catalog-profile.mjs
@@ -421,7 +421,9 @@ function dueOf(state, src, t) {
 }
 /** 范围 needed / all：服务端的口径（有待核实关系的在前，这里按使用次数，同清单顺序） */
 const scopeTables = (state, src, scope) => state.tables[src].filter((t) => !t.missing).sort((a, b) => b.usage - a.usage)
-  .filter((t) => (scope === 'all' ? checksOf(t) > 0 && !t.is_view : dueOf(state, src, t) != null)).map((t) => t.name)
+  .filter((t) => (scope === 'all' ? checksOf(t) > 0 && !t.is_view
+    // 取值观察：有数据目录、不是视图的表；state.observedRecently 时都算最近观察过（服务端 12 小时内不再观察）
+    : scope === 'observe' ? !!t.notes && !t.is_view && !state.observedRecently : dueOf(state, src, t) != null)).map((t) => t.name)
 
 // ---------------------------------------------------------------------------
 // 假服务端的起草任务：只做排队要用的部分（建、取、停），起草本身不推
@@ -560,6 +562,10 @@ function handlers(state, replies = {}) {
       if ((body?.tables?.length ?? 0) > 500) return json({ detail: '一次最多剖析 500 张表' }, 422)(route)
       const tables = body?.tables?.length ? [...new Set(body.tables)] : scopeTables(state, src, body?.scope ?? 'needed')
       if (!tables.length) {
+        if (body?.scope === 'observe') {
+          return json({ detail: '没有要观察的表：取值观察只看有数据目录的表，这些表最近 12 小时内都观察过了（或者还没有数据目录）',
+            code: 'profile_nothing' }, 409)(route)
+        }
         return json({ detail: '按「需要剖析的表」挑出 0 张表：表都剖析过、上次都做完了，也没有新的待核实关联关系。可以改选「全部表」，或指定要剖析的表',
           code: 'profile_nothing' }, 409)(route)
       }
@@ -793,10 +799,14 @@ await section('剖析 · 开始前：范围、个数与预算', async () => {
   check('页头写「数据剖析」', (await head(page).getAttribute('aria-label')) === '数据剖析' && (await head(page).getAttribute('data-catalog-profile')) === '')
   await openDialog(page)
   const scopes = await box(page).locator('[data-profile-scope]').evaluateAll((els) => els.map((e) => `${e.getAttribute('data-profile-scope')}:${e.getAttribute('data-count')}`))
+  const observeText = await inBox(page, '[data-profile-scope="observe"]')
   // 需要剖析：从没剖析过的订单、空日志表；上次没做完的访问记录、退款；剖析后有新关系的商品。全部：再加上客户，不含视图、
   // 没有可剖析项的配置表、只剩目录的旧表
-  check('没选表、没筛选：范围是「需要剖析的 5 张表」（默认）和「全部 6 张表」', scopes.join(',') === 'needed:5,all:6'
+  check('没选表、没筛选：范围是「需要剖析的 5 张表」（默认）、「全部 6 张表」和「取值观察：有数据目录的 1 张表」',
+    scopes.join(',') === 'needed:5,all:6,observe:1'
         && await box(page).locator('[data-profile-scope="needed"] input').isChecked(), scopes.join(','))
+  check('……取值观察写明只观察不剖析、看哪几样、最近 12 小时内观察过的不再观察', ['取值观察：有数据目录的 1 张表', '只观察、不剖析',
+    '码值分布、最新业务日期、按日行数和行数', '最近 12 小时内观察过的表这次不再观察'].every((x) => observeText.includes(x)), observeText)
   check('……写明需要剖析的口径（含上次是空表的不算）、全部不含视图', (await inBox(page, '[data-profile-scope="needed"]')).includes('从没剖析过、上次没做完，或剖析之后又出现了待核实关联关系、要取码值或日期范围的列的表。上次是空表的不算')
         && (await inBox(page, '[data-profile-scope="all"]')).includes('不含视图'))
   const budget = await inBox(page, '[data-profile-budget]')
@@ -840,6 +850,43 @@ await section('剖析 · 开始前：范围、个数与预算', async () => {
   await close()
 })
 
+await section('剖析 · 取值观察：范围、请求与挑不出表', async () => {
+  // 取值观察以前只能经定时维护触发（下一轮验收）：对话框的范围里多一项「取值观察」，建任务带 scope: observe；
+  // 有数据目录的表都在 12 小时内观察过时服务端 409 profile_nothing，说「没有要观察的表」，不给「改选全部表」（那是剖析）
+  const state = freshState()
+  const { page, sent, errors, natives, close } = await open(`/data/catalog/${S1}`, { handlers: handlers(state) })
+  await page.locator('[data-catalog-row]').first().waitFor()
+  await openDialog(page)
+  await box(page).locator('[data-profile-scope="observe"]').click()
+  check('选取值观察：预算按 1 张表估', (await inBox(page, '[data-profile-budget-total]')).startsWith('这次 1 张表'),
+    await inBox(page, '[data-profile-budget-total]'))
+  await page.locator('[data-profile-start]').click()
+  check('开始：建任务的请求带「取值观察」范围', await until(async () => (await phase(page)) === 'running')
+        && JSON.stringify(lastBody(sent, /profile-jobs$/)) === JSON.stringify({ scope: 'observe' }), JSON.stringify(lastBody(sent, /profile-jobs$/)))
+  const job = state.pjobs.at(-1)
+  check('……任务的范围是 observe、只含有数据目录的订单表', job.scope === 'observe' && job.tasks.map((t) => t.table_name).join(',') === 'orders',
+    `${job.scope} ${job.tasks.map((t) => t.table_name).join(',')}`)
+  for (const t of job.tasks) settle(state, job, t.table_name, { queries: 4, elapsed_ms: 2000 })
+  finish(job)
+  check('……观察完：对话框显示结束后的结果', await until(async () => (await phase(page)) === 'done'), await phase(page))
+  await page.keyboard.press('Escape')
+  await until(async () => (await box(page).count()) === 0)
+
+  // 都在 12 小时内观察过：开始不了，说「没有要观察的表」，没有「改选全部表」
+  state.observedRecently = true
+  await openDialog(page)
+  await box(page).locator('[data-profile-scope="observe"]').click()
+  await page.locator('[data-profile-start]').click()
+  check('挑不出表：说「没有要观察的表」和服务端的原话（12 小时内都观察过）', await until(async () => (await inBox(page, '[data-profile-blocked]')).startsWith('没有要观察的表'))
+        && (await inBox(page, '[data-profile-blocked]')).includes('最近 12 小时内都观察过了'), await inBox(page, '[data-profile-blocked]'))
+  check('……不给「改选全部表」（那是剖析，不是观察），可以回到上一步', await box(page).locator('[data-profile-choose-all]').count() === 0
+        && await page.locator('[data-profile-back]').count() === 1)
+  check('写请求只有建任务', writes(sent, /./).every((s) => /profile-jobs$/.test(s.key)), JSON.stringify(writes(sent, /./).map((s) => s.key)))
+  check('没有原生对话框', natives.length === 0, natives.join(' | '))
+  check('没有运行时报错', errors.length === 0, errors[0] ?? '')
+  await close()
+})
+
 await section('剖析 · 进行中：逐表进度、后台继续与停止', async () => {
   const state = freshState()
   const { page, sent, errors, close } = await open(`/data/catalog/${S1}`, { handlers: handlers(state) })
diff --git a/scripts/check-evals.mjs b/scripts/check-evals.mjs
index 50e238e7..db3f53b1 100644
--- a/scripts/check-evals.mjs
+++ b/scripts/check-evals.mjs
@@ -161,13 +161,15 @@ const RESULTS = [
     scores: [{ id: 'e1', kind: 'scalar', required: true, outcome: 'traced', expected: 330, got: 330 }], traced: true, review: null, tier: null,
     judge: null, looks: { model: 1, system: 1 }, calls: 2, timings: { total_ms: 41000 }, usage: {}, sql_match: { status: 'proven', aspects: [] },
     details_artifact: 'd-1', current: null, error: null, started_at: iso(NOW - 2 * H), finished_at: iso(NOW - H), question: CASE_SQL.question,
-    cost: { cost_usd: 0.014, unpriced_calls: 0 } },
+    question_template: CASE_SQL.question, cost: { cost_usd: 0.014, unpriced_calls: 0 } },
   { id: 'r-2', eval_run_id: 'ev-1', seq: 2, case_id: 'case-2', case_seq: 2, case_hash: 'h2', repeat: 0, arm: null, status: 'ok',
     run_id: 'run-answer-2', gold_run_ids: ['run-gold-2'], gold: { status: 'ok' }, path: 'ran', verdict: 'fail',
     scores: [{ id: 'e1', kind: 'scalar', required: true, outcome: 'wrong', expected: 330, got: 310, reason: '答案里的数和期望对不上' }], traced: false,
     review: null, tier: null, judge: null, looks: { model: 1, system: 0 }, calls: 2, timings: { total_ms: 60000 }, usage: {},
     sql_match: { status: 'related', aspects: ['time'] }, details_artifact: 'd-2', current: null, error: '答案里的数和期望对不上',
-    started_at: iso(NOW - 2 * H), finished_at: iso(NOW - H), question: CASE_CAL.question, cost: { cost_usd: 0.02, unpriced_calls: 0 } },
+    started_at: iso(NOW - 2 * H), finished_at: iso(NOW - H), cost: { cost_usd: 0.02, unpriced_calls: 0 },
+    // 带变量的题：question 是实际问的（服务端按这次评测的截至时间渲染好），模板原文另放
+    question: '10 月 5 日到 10 月 11 日的周销售额？', question_template: '{{ schedule.last_week_start }} 到 {{ schedule.last_week_end }} 的周销售额？' },
   { id: 'r-3', eval_run_id: 'ev-1', seq: 3, case_id: 'case-3', case_seq: 3, case_hash: 'h3', repeat: 0, arm: null, status: 'ok',
     run_id: null, gold_run_ids: [], gold: null, path: 'asked', verdict: 'pass', scores: [{ id: 'e1', kind: 'ask', required: true, outcome: 'traced' }],
     traced: true, review: null, tier: null, judge: null, looks: { model: 0, system: 0 }, calls: 1, timings: { total_ms: 9000 }, usage: {},
@@ -254,6 +256,9 @@ function reply(state, method, path, url, body) {
     if (out) return out
   }
   if (path === '/auth/me' && state.me) return { json: state.me }
+  // 「提取模板」去画布之后，画布页给这张图起一段画布对话（POST /conversations）：早晚不定，以前偶尔赶在检查结束前发出去、
+  // 漏到服务端。照 check-ledger 的写法在这里接住，不进写请求的清单（不是评测页发的）
+  if (method === 'POST' && path === '/conversations') return { json: { id: 'cv-eval', kind: body?.kind ?? 'canvas', title: '', turns: [] } }
   if (method === 'GET' && path === '/datasources') return { json: [DS] }
   if (method === 'GET' && path === '/evals/suites') return { json: { suites: state.suites } }
   if (method === 'POST' && path === '/evals/suites') {
@@ -357,7 +362,11 @@ function reply(state, method, path, url, body) {
     }
     if (method === 'POST' && rest === '/stop') {
       write()
-      Object.assign(ev, { status: 'cancelled', running: false, current: [], stop_reason: null,
+      // 服务端（runner._counts、_close）：停下的原因记 stop_reason、不当成出错；没跑的题单独计 not_run，不进 errors，done 不等于 total
+      Object.assign(ev, { status: 'cancelled', running: false, current: [], stop_reason: '评测已停止', error: null,
+        counts: { total: 3, done: 1, not_run: 2, running: 0, pending: 0, passed: 1, failed: 0, ungradable: 0, skipped: 0, errors: 0 },
+        summary: { ...clone(SUMMARY), items: 3, done: 3, verdicts: { pass: 1, fail: 0, ungradable: 0, skipped: 0 }, not_run: 2, errors: 0,
+          accuracy: RATE(1, 1), traced_accuracy: RATE(1, 1) },
         actions: { stop: false, retry: true, rescore: true, pin: false, unpin: false, open: 2, cleaned: false } })
       return { json: ev }
     }
@@ -738,6 +747,11 @@ await section('评测详情：进度、正在做什么、停止和重试、重
   await page.locator('[data-eval-retry]').waitFor()
   check('停止之后：服务端说能重试，给「重试没做完的」，悬停写几项', (await page.locator('[data-eval-retry]').getAttribute('title')).includes('2 项')
     && writes(state, /\/stop$/).length === 1)
+  check('……停下的原因照服务端写「评测已停止」，不当成出错；没评的 2 项单列，不算出错', await until(async () =>
+    (await page.locator('[data-eval-stop-reason]').count()) === 1)
+    && (await text(page, '[data-eval-stop-reason]')).includes('评测已停止') && await page.locator('[data-eval-error]').count() === 0
+    && (await text(page, '[data-eval-not-run]')).startsWith('2 项没评') && await page.locator('[data-eval-errors]').count() === 0,
+  await text(page, '[data-eval-run-head]'))
   await page.locator('[data-eval-retry]').click()
   await until(() => writes(state, /\/retry$/).length === 1)
   check('重试：发了一次', writes(state, /\/retry$/).length === 1)
@@ -772,6 +786,11 @@ await section('逐题结果、和上一次比、A/B 配对表', async () => {
   check('?case= 打开逐题、定位并展开那道题', await page.locator('[data-eval-result-detail="r-2"]').count() === 1)
   check('逐题：结论、期望和答案里的数、SQL 差在哪', t2.includes('不通过') && t2.includes('330') && t2.includes('310') && t2.includes('时间口径'), t2)
   check('一题的花费照服务端加好的写', (await page.locator('[data-eval-result="r-1"]').innerText()).includes('$0.014'))
+  check('逐题的问法写实际问的（变量渲染好了），模板原文另起一行', await text(page, '[data-eval-result="r-2"] [data-eval-result-question]')
+      === '10 月 5 日到 10 月 11 日的周销售额？'
+    && (await text(page, '[data-eval-result="r-2"] [data-eval-result-template]')).startsWith('题目原文：{{ schedule.last_week_start }}')
+    && await page.locator('[data-eval-result="r-1"] [data-eval-result-template]').count() === 0,
+  await text(page, '[data-eval-result="r-2"] [data-eval-result-question]'))
   check('点开去运行详情', await page.locator('[data-eval-result-detail="r-2"] [data-eval-result-run]').getAttribute('href') === '/runs/run-answer-2')
   await page.locator('[data-eval-result-details]').waitFor()
   check('展开的明细写作答的经过和答案原文（明细工件）', (await text(page, '[data-eval-result-details]')).includes('自动重来了 1 次')
diff --git a/scripts/check-verify.mjs b/scripts/check-verify.mjs
index 0ae030a4..59e96516 100644
--- a/scripts/check-verify.mjs
+++ b/scripts/check-verify.mjs
@@ -450,9 +450,9 @@ await section('审批卡：没有预告时，「记入数据目录」只说必
   // 而避开本来对的一边。变体的 catalog 是服务端对这一对 SQL 的真实预告，按它核对
   const { CATALOG_TABLE_FIELD_LABEL: LABEL } = await import(await load('/src/lib/terms.ts'))
   // [变体, 说明, 服务端不写的那一种卡上照实说的那句（只在卡上提数据目录时有）]。变体挑的是收尾版差异说明下各有代表性
-  // 的几种（backend/tests/test_dual_check_fixture.py 的模块说明）：同一列上的时间范围不同只给兜底（fallback）
+  // 的几种（backend/tests/test_dual_check_fixture.py 的模块说明）：套着 DATE() 的时间范围不同只给兜底（fallback）
   const cases = [
-    ['fallback', '说不清的（同一列上的时间范围不同，只给兜底）'],
+    ['fallback', '说不清的（套着 DATE() 的时间范围不同，只给兜底）'],
     ['two_filters', '差两个筛选条件'],
     ['join', '差在关联方式'],
     ['parse', '一边解析失败'],
@@ -492,7 +492,7 @@ await section('审批卡：没有预告时，「记入数据目录」只说必
         && note.includes('按不同的列筛时间') && note.includes('采用的一边所用的时间列'), note)
     }
     if (name === 'fallback') {
-      // 收尾版差异说明归纳不了同一列上的时间范围：卡上写那句固定的兜底说明，默认展开对照，审核人逐行对照两条 SQL
+      // 收尾版差异说明归纳不了套着函数的时间范围：卡上写那句固定的兜底说明，默认展开对照，审核人逐行对照两条 SQL
       check('……写那句固定的兜底说明，不当成一条具体差异', (await text(gmv.locator('[data-dual-fallback]'))) === '两条 SQL 的写法不同，未能归纳成具体差异'
         && await gmv.locator('[data-dual-diff-item]').count() === 0, await text(gmv.locator('[data-dual-fallback]')))
       check('……默认展开对照', (await gmv.locator('[data-dual-compare-toggle]').getAttribute('aria-expanded')) === 'true')
@@ -532,7 +532,7 @@ await section('审批卡：没有预告时，「记入数据目录」只说必
         const marked = [...m.normDel, ...m.normAdd]
         check('……两边相同的条件（开始日期）不标，只标不同的那一行', m.normDel.length === 1 && m.normAdd.length === 1
           && !marked.some((l) => l.includes(">= '2026-09-01'"))
-          && marked.every((l) => l.includes("orders.paid_at < '2026-09-")), marked.join(' ‖ '))
+          && marked.every((l) => l.includes("DATE(orders.paid_at) < '2026-09-")), marked.join(' ‖ '))
       }
       if (name === 'two_filters') {
         check('……一边多出的两个条件各占一行', m.normDel.includes("orders.channel = 'web'")
diff --git a/scripts/dialect_it.py b/scripts/dialect_it.py
index 86538338..f5621d42 100644
--- a/scripts/dialect_it.py
+++ b/scripts/dialect_it.py
@@ -50,6 +50,9 @@ SETTINGS = ProfileSettings(enabled=True, table_max_queries=40, table_max_s=60, q
 N_ORDERS, N_EVENTS_OLD, N_EVENTS_NEW = 40, 10, 1500
 BASE = _dt.datetime(2026, 9, 1, 10, 0, 0)
 WINDOW_FROM = "2026-09-05"
+#: 按日行数的上界（观察当天，含）：WINDOW_UNTIL 在全部夹具日期之后；WINDOW_CUT 落在中间，之后的日子不进窗口
+WINDOW_UNTIL = "2026-09-30"
+WINDOW_CUT = "2026-09-07"
 
 failures: list[str] = []
 
@@ -80,6 +83,8 @@ EXPECT = {
     "pay_way": Counter(o["pay_way"] for o in ORDERS),
     "paid_at": (min(o["paid_at"] for o in VALID), max(o["paid_at"] for o in VALID)),
     "day_volume": Counter(o["paid_at"].date().isoformat() for o in IN_WINDOW),
+    "day_volume_cut": Counter(o["paid_at"].date().isoformat() for o in IN_WINDOW
+                              if o["paid_at"].date().isoformat() <= WINDOW_CUT),
     "window_status": Counter(o["status"] for o in IN_WINDOW),
 }
 
@@ -354,9 +359,13 @@ async def run_target(t: Target) -> None:
         # ---- 取值观察的新方法
         tbl = d.table(orders, "orders")
         paid_col = d.column(orders, _key(orders, "paid_at"))
-        res = await data_engine.run_query(source, d.day_volume_sql(tbl, paid_col, WINDOW_FROM))
+        res = await data_engine.run_query(source, d.day_volume_sql(tbl, paid_col, WINDOW_FROM, WINDOW_UNTIL))
         got_days = Counter({_day(row[0]): int(row[1]) for row in res.rows})
         check(got_days == EXPECT["day_volume"], f"按日行数和手算一致（无效日期不算进窗口）：{dict(got_days)}")
+        res = await data_engine.run_query(source, d.day_volume_sql(tbl, paid_col, WINDOW_FROM, WINDOW_CUT))
+        got_cut = Counter({_day(row[0]): int(row[1]) for row in res.rows})
+        check(got_cut == EXPECT["day_volume_cut"],
+              f"按日行数只数到观察当天（含当天的时分秒，之后的不算）：{dict(got_cut)}")
         status_col = d.column(orders, _key(orders, "status"))
         res = await data_engine.run_query(source, d.window_value_counts_sql(tbl, status_col, paid_col, WINDOW_FROM, 21))
         got_window = Counter({str(row[0]).rstrip(" "): int(row[1]) for row in res.rows})