fix(intelligent-eval): respect time-window slots before nudging analyst

冒烟发现 analyst_nudge 不考虑时间窗口：首会话完成后 10 分钟即催促，
但后续时段尚未到期，导致报告提前收敛、漏掉计划内会话。催促闸门新增
"窗口未结束且会话数未达计划则跳过"；worker 触发指令同步明确仅当所有
时段会话都达终态才转 analyst。

backend/agenteval/intelligent_eval/lifecycle.py

@@ -31,7 +31,7 @@ from agenteval.intelligent_eval.repository import (
     IntelligentEvalRepository,
     IntelligentEvalSessionRepository,
 )
-from agenteval.storage.db import IntelligentEvalSessionDB, utc_now
+from agenteval.storage.db import IntelligentEvalSessionDB, as_utc, utc_now
 from agenteval.storage.repository import TargetRepository
 
 # 合法转换表：当前状态 → 允许的目标状态集合
@@ -691,12 +691,37 @@ def enforce_executing_ceiling(session: Session) -> int:
     return failed
 
 
+def _window_has_pending_future_slots(eval_db, sessions, now) -> bool:
+    """窗口未结束且会话数未达计划 → 未来时段还要建会话，不该催 analyst。"""
+    if not eval_db.plan or not eval_db.started_at:
+        return False
+    from agenteval.intelligent_eval.domain import parse_time_slot
+
+    plan = eval_db.get_plan()
+    slots = plan.get("time_distribution") or []
+    end_hours: list[float] = []
+    planned = 0
+    for slot in slots:
+        parsed = parse_time_slot(slot.get("time_slot", ""))
+        if parsed is None:
+            continue
+        end_hours.append(parsed[1])
+        planned += int(slot.get("sessions", 0) or 0)
+    if not end_hours:
+        return False
+    window_end = (as_utc(eval_db.started_at) + timedelta(hours=max(end_hours))).replace(tzinfo=None)
+    if now >= window_end:
+        return False
+    expected = max(planned, int(plan.get("estimated_sessions", 0) or 0))
+    return len(sessions) < expected
+
+
 def evals_needing_analyst_nudge(session: Session) -> list[str]:
     """返回需要平台催促 analyst 的 executing 评估 id 列表。
 
-    条件（ADR-0011）：存在会话且全部终态、末会话终态已满
-    ANALYST_NUDGE_DELAY_MINUTES、催促次数 < ANALYST_NUDGE_MAX、距上次催促
-    已满 ANALYST_NUDGE_DELAY_MINUTES（冷却，避免每分钟连发）。
+    条件（ADR-0011）：存在会话且全部终态、时间窗口内无未到期时段欠账、
+    末会话终态已满 ANALYST_NUDGE_DELAY_MINUTES、催促次数 < ANALYST_NUDGE_MAX、
+    距上次催促已满 ANALYST_NUDGE_DELAY_MINUTES（冷却，避免每分钟连发）。
     """
     from sqlmodel import select
 
@@ -718,6 +743,10 @@ def evals_needing_analyst_nudge(session: Session) -> list[str]:
         ).all()
         if not sessions or any(s.status not in _TERMINAL_SESSION_STATUSES for s in sessions):
             continue
+        # 冒烟教训：窗口未结束且会话数未达计划时，未来时段到期后还要建会话，
+        # 此时催促 analyst 会让报告提前收敛（漏掉后续时段的证据）
+        if _window_has_pending_future_slots(row, sessions, now):
+            continue
         closed_moments = [s.closed_at for s in sessions if s.closed_at is not None]
         if not closed_moments:
             continue

backend/agenteval/web/app.py

@@ -190,7 +190,9 @@ async def _trigger_intelligent_worker(nudge_eval_ids: Optional[list[str]] = None
         "仅执行当前时间对应时段（time_distribution 中当前 offset 所在时段）内欠账的会话"
         "——按评估 started_at 与当前时间精确判断当前时段，只创建该时段计划内的会话，"
         "绝不创建未来时段的会话，未来时段到期后平台会再次触发你；"
-        "若所有会话已完成则调用 agenteval-intelligent-analyst skill 完成分析并生成报告。"
+        "仅当所有时段计划内的会话都已达到终态（completed/failed/expired）时，"
+        "才调用 agenteval-intelligent-analyst skill 完成分析并生成报告；"
+        "若还有未到期的未来时段，保持等待、本次不做任何操作。"
     )
     try:
         proc = await asyncio.to_thread(

tests/unit/test_intelligent_eval_watchdog_characterization.py

@@ -48,7 +48,7 @@ def _make_executing_eval(db_session: Session, **overrides) -> IntelligentEvalDB:
         name=overrides.pop("name", "wd-eval"),
         target_id="t1",
         status=IntelligentEvalStatus.EXECUTING.value,
-        started_at=utc_now() - timedelta(hours=3),
+        started_at=overrides.pop("started_at", utc_now() - timedelta(hours=3)),
         **overrides,
     )
     db_session.add(ev)
@@ -256,6 +256,50 @@ def test_analyst_nudge_skipped_with_running_session(db_session: Session):
     assert evals_needing_analyst_nudge(db_session) == []
 
 
+def test_analyst_nudge_skipped_while_future_slot_pending(db_session: Session):
+    """冒烟教训：窗口内还有未到期时段（会话数未达计划）时不催 analyst。"""
+    from agenteval.intelligent_eval.lifecycle import evals_needing_analyst_nudge
+
+    ev = _make_executing_eval(
+        db_session,
+        started_at=utc_now() - timedelta(minutes=20),
+        plan=json.dumps(
+            {
+                "estimated_sessions": 2,
+                "time_distribution": [
+                    {"time_slot": "0-30min", "sessions": 1},
+                    {"time_slot": "30-60min", "sessions": 1},
+                ],
+            }
+        ),
+    )
+    _make_terminal_session(db_session, ev.id, closed_minutes_ago=15)
+
+    assert evals_needing_analyst_nudge(db_session) == []
+
+
+def test_analyst_nudge_fires_after_window_elapsed_with_deficit(db_session: Session):
+    """窗口已结束但会话欠账：不再等未来时段，催促 analyst 按现有证据收敛。"""
+    from agenteval.intelligent_eval.lifecycle import evals_needing_analyst_nudge
+
+    ev = _make_executing_eval(
+        db_session,
+        started_at=utc_now() - timedelta(minutes=70),
+        plan=json.dumps(
+            {
+                "estimated_sessions": 2,
+                "time_distribution": [
+                    {"time_slot": "0-30min", "sessions": 1},
+                    {"time_slot": "30-60min", "sessions": 1},
+                ],
+            }
+        ),
+    )
+    _make_terminal_session(db_session, ev.id, closed_minutes_ago=30)
+
+    assert evals_needing_analyst_nudge(db_session) == [ev.id]
+
+
 def test_analyst_nudge_capped_and_cooled_down(db_session: Session):
     """催促上限 3 次；两次催促间隔未满 10min 不重复。"""
     from agenteval.intelligent_eval.lifecycle import (