diff --git a/src/data/bzzoiro.py b/src/data/bzzoiro.py
index e41cb98..8c51fe8 100644
--- a/src/data/bzzoiro.py
+++ b/src/data/bzzoiro.py
@@ -27,6 +27,26 @@ from src.db.models import League, Match, MatchStats, Team
 logger = logging.getLogger(__name__)
 
 
+def _to_date(value):
+    """把 datetime / date / str 统一成 `date`。"""
+    if value is None:
+        return None
+    if hasattr(value, "date") and callable(value.date):
+        return value.date()
+    return value
+
+
+def _match_key(home_team_id: int, away_team_id: int, match_date) -> tuple[int, int, str]:
+    """比赛去重键:(主队, 客队, 天级日期 ISO 字符串)。
+
+    统一在这里构造,避免"预加载时用 str(date)、写入时用 isoformat()"这类
+    隐式格式依赖 —— 两者当前恰好相等,但一旦有人改动其一就会静默失配,
+    导致所有比赛被判为不存在而重复插入。
+    """
+    d = _to_date(match_date)
+    return (home_team_id, away_team_id, d.isoformat() if d is not None else "")
+
+
 def _fetch_json_sync(path: str, params: dict | None = None, max_retries: int = 3) -> dict | list:
     """同步 HTTP(bzzoiro 客户端保持同步,在 async 函数里 run_in_executor)。"""
     base = settings.BZZOIRO_BASE.rstrip("/")
@@ -153,7 +173,9 @@ class BzzoiroSource:
             # === 批量优化: 预加载球队和已有比赛到内存 ===
             team_name_to_id: dict[str, int] = {}
             existing_matches: dict[tuple[int, int, str], Match] = {}  # 完整对象,避免重复查询
-            normalized_matches: list = []  # 缓存规范化结果,避免重复调用
+            # (NormalizedMatch, 原始 event) 成对保存:后续写 source_record_id 时
+            # 必须用配对的那条 event,不能依赖外层循环变量残留值。
+            normalized_matches: list[tuple] = []
 
             if raw_events:
                 # 一次遍历: 收集球队名 + 规范化
@@ -167,7 +189,7 @@ class BzzoiroSource:
                             logger.debug("normalize skip: %s", e)
                             league_r["errors"].append(f"normalize: {e}")
                             continue
-                        normalized_matches.append(nm)
+                        normalized_matches.append((nm, raw))
                         all_team_names.add(nm.home_team)
                         all_team_names.add(nm.away_team)
 
@@ -179,10 +201,10 @@ class BzzoiroSource:
                 # 预加载已有比赛(完整对象)
                 stmt = select(Match).where(Match.league_id == league.id)
                 for m in (await db.execute(stmt)).scalars():
-                    key = (m.home_team_id, m.away_team_id, str(m.match_date_date))
+                    key = _match_key(m.home_team_id, m.away_team_id, m.match_date_date)
                     existing_matches[key] = m
 
-            for nm in normalized_matches:
+            for nm, raw in normalized_matches:
                 # 球队: 内存查找 + 按需创建
                 home_team_id = team_name_to_id.get(nm.home_team)
                 if home_team_id is None:
@@ -201,8 +223,7 @@ class BzzoiroSource:
                     team_name_to_id[nm.away_team] = away_team_id
 
                 # 查找已有比赛: 内存查找
-                date_key = nm.date.date().isoformat() if hasattr(nm.date, "date") else str(nm.date)
-                match_key = (home_team_id, away_team_id, date_key)
+                match_key = _match_key(home_team_id, away_team_id, nm.date)
                 existing_match = existing_matches.get(match_key)
 
                 if existing_match is None:
@@ -212,7 +233,7 @@ class BzzoiroSource:
                         home_team_id=home_team_id,
                         away_team_id=away_team_id,
                         match_date=nm.date,
-                        match_date_date=nm.date.date() if hasattr(nm.date, "date") else nm.date,
+                        match_date_date=_to_date(nm.date),
                         match_status=nm.match_status,
                         home_goals=nm.home_goals,
                         away_goals=nm.away_goals,
@@ -263,7 +284,14 @@ class BzzoiroSource:
                         existing_match.match_stage = nm.match_stage
                         changed = True
                     if existing_match.stats is None and (nm.home_xg is not None or nm.away_xg is not None):
-                        existing_match.stats = MatchStats(match_id=existing_match.id)
+                        now = datetime.now(timezone.utc)
+                        existing_match.stats = MatchStats(
+                            match_id=existing_match.id,
+                            source="bzzoiro",
+                            source_event_id=str(raw.get("id", "")),
+                            retrieved_at=now,
+                            available_at=now,
+                        )
                         db.add(existing_match.stats)
                         await db.flush()
                     if existing_match.stats is not None:
diff --git a/src/db/models.py b/src/db/models.py
index dac2e43..fe28fc0 100644
--- a/src/db/models.py
+++ b/src/db/models.py
@@ -75,10 +75,20 @@ class Match(Base):
     created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=_utcnow)
     updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), default=_utcnow, onupdate=_utcnow)
 
-    league: Mapped[League] = relationship(back_populates="matches")
-    home_team: Mapped[Team] = relationship(foreign_keys=[home_team_id], back_populates="home_matches")
-    away_team: Mapped[Team] = relationship(foreign_keys=[away_team_id], back_populates="away_matches")
-    stats: Mapped["MatchStats | None"] = relationship(back_populates="match", cascade="all, delete-orphan")
+    league: Mapped[League] = relationship(back_populates="matches", lazy="selectin")
+    # lazy="selectin": 这些关系在业务里几乎总是一起读取(切片/回测/展示)。
+    # 默认的 lazy="select" 在 async SQLAlchemy 下,于 session 之外或未显式
+    # eager-load 时访问会抛 MissingGreenlet —— 已因此导致 form/stats/h2h
+    # 三个专家切片静默失败。统一改为预加载,从根上消除这类问题。
+    home_team: Mapped[Team] = relationship(
+        foreign_keys=[home_team_id], back_populates="home_matches", lazy="selectin"
+    )
+    away_team: Mapped[Team] = relationship(
+        foreign_keys=[away_team_id], back_populates="away_matches", lazy="selectin"
+    )
+    stats: Mapped["MatchStats | None"] = relationship(
+        back_populates="match", cascade="all, delete-orphan", lazy="selectin"
+    )
     predictions: Mapped[list["Prediction"]] = relationship(back_populates="match", cascade="all, delete-orphan")
 
     __table_args__ = (
diff --git a/src/db/repositories.py b/src/db/repositories.py
index 9d08e16..cc1e512 100644
--- a/src/db/repositories.py
+++ b/src/db/repositories.py
@@ -37,12 +37,17 @@ class MatchRepository:
     async def find_by_teams_and_date(
         self, league_id: int, home_team_id: int, away_team_id: int, date
     ) -> Match | None:
-        """按联赛+主队+客队+日期查找比赛(天级匹配)。"""
+        """按联赛+主队+客队+日期查找比赛(天级匹配)。
+
+        预加载 stats:调用方(understat 回填)会读取 existing.stats,
+        async session 下惰性加载会抛 MissingGreenlet。
+        """
         if hasattr(date, "date"):
             date = date.date()
 
         stmt = (
             select(Match)
+            .options(selectinload(Match.stats))
             .where(Match.league_id == league_id)
             .where(Match.home_team_id == home_team_id)
             .where(Match.away_team_id == away_team_id)
diff --git a/src/llm/backtest.py b/src/llm/backtest.py
index 25fb45e..1d947bc 100644
--- a/src/llm/backtest.py
+++ b/src/llm/backtest.py
@@ -8,10 +8,12 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass, field
+from datetime import datetime
 
-from sqlalchemy import and_, select
+from sqlalchemy import select
+from sqlalchemy.orm import selectinload
 
-from src.db.models import League, Match
+from src.db.models import Match
 from src.db.unit_of_work import get_uow
 from src.llm.eval import settle_prediction
 from src.llm.predict import predict_match
@@ -38,6 +40,23 @@ class BacktestMatchResult:
     prediction_id: int
 
 
+@dataclass
+class BacktestCandidate:
+    """回测候选比赛(字段快照,不持有 ORM 对象)。
+
+    session 关闭后仍可安全读取:所有需要的关系字段已在查询时物化为普通值,
+    避免在 session 之外访问惰性加载的关系属性(会抛 MissingGreenlet)。
+    """
+    match_id: int
+    league_code: str | None
+    home_team: str
+    away_team: str
+    match_date: datetime
+    home_goals: int
+    away_goals: int
+
+
+
 @dataclass
 class BacktestSummary:
     """回测汇总统计。"""
@@ -66,10 +85,20 @@ async def _get_historical_matches(
     date_from: str | None = None,
     date_to: str | None = None,
     limit: int = 50,
-) -> list[Match]:
-    """查询已完赛且有比分的比赛(回测候选)。"""
+) -> list[BacktestCandidate]:
+    """查询已完赛且有比分的比赛(回测候选)。
+
+    返回普通值快照而非 ORM 对象:调用方在 session 关闭后仍需使用这些字段,
+    而 league / home_team / away_team 是惰性加载关系,在 async 下于 session
+    之外访问会抛 MissingGreenlet。这里用 selectinload 预加载后立即物化。
+    """
     stmt = (
         select(Match)
+        .options(
+            selectinload(Match.league),
+            selectinload(Match.home_team),
+            selectinload(Match.away_team),
+        )
         .where(Match.match_status == "finished")
         .where(Match.home_goals.is_not(None))
         .where(Match.away_goals.is_not(None))
@@ -83,7 +112,19 @@ async def _get_historical_matches(
 
     stmt = stmt.order_by(Match.match_date.desc()).limit(limit)
     result = await db.execute(stmt)
-    return list(result.scalars().all())
+    # 在 session 内物化为纯数据,切断与 ORM 会话的耦合
+    return [
+        BacktestCandidate(
+            match_id=m.id,
+            league_code=m.league.code if m.league else None,
+            home_team=m.home_team.name if m.home_team else "?",
+            away_team=m.away_team.name if m.away_team else "?",
+            match_date=m.match_date,
+            home_goals=m.home_goals,
+            away_goals=m.away_goals,
+        )
+        for m in result.scalars().all()
+    ]
 
 
 async def run_backtest(
@@ -109,32 +150,34 @@ async def run_backtest(
         BacktestSummary 含逐场结果 + 汇总统计
     """
     async with get_uow() as session:
-        matches = await _get_historical_matches(
+        candidates = await _get_historical_matches(
             session, league_id=league_id, date_from=date_from, date_to=date_to, limit=limit
         )
 
-    summary = BacktestSummary(total=len(matches), scored=0)
+    summary = BacktestSummary(total=len(candidates), scored=0)
 
-    for m in matches:
+    for c in candidates:
         try:
             # 预测 (build_context 内部已用 before=match_date 防泄漏,
             # injuries_slice 也使用 as_of=match_date 过滤 retrieved_at)
-            result = await predict_match(m.id, mode=mode, model=model)
+            # 回测必须禁用结果缓存: 否则命中缓存会复用同一 prediction_id,
+            # 导致 settle 反复覆盖同一条记录(见 P1-3)。
+            result = await predict_match(c.match_id, mode=mode, model=model, use_cache=False)
 
             # 用实际比分 settle
-            await settle_prediction(result.prediction_id, m.home_goals, m.away_goals)
+            await settle_prediction(result.prediction_id, c.home_goals, c.away_goals)
 
-            actual = _actual_1x2(m.home_goals, m.away_goals)
+            actual = _actual_1x2(c.home_goals, c.away_goals)
             correct = result.pred_1x2 == actual
 
             bt = BacktestMatchResult(
-                match_id=m.id,
-                league_code=m.league.code if m.league else None,
-                home_team=m.home_team.name if m.home_team else "?",
-                away_team=m.away_team.name if m.away_team else "?",
-                match_date=m.match_date.strftime("%Y-%m-%d") if m.match_date else "?",
-                actual_home=m.home_goals,
-                actual_away=m.away_goals,
+                match_id=c.match_id,
+                league_code=c.league_code,
+                home_team=c.home_team,
+                away_team=c.away_team,
+                match_date=c.match_date.strftime("%Y-%m-%d") if c.match_date else "?",
+                actual_home=c.home_goals,
+                actual_away=c.away_goals,
                 actual_1x2=actual,
                 pred_home=result.pred_home_goals,
                 pred_away=result.pred_away_goals,
@@ -146,8 +189,10 @@ async def run_backtest(
             summary.results.append(bt)
             summary.scored += 1
 
-        except Exception as e:
-            logger.warning("backtest match %s failed: %s", m.id, e)
+        except Exception:
+            # 用 exception 而非 warning:保留堆栈,否则集成层缺陷(如惰性加载
+            # 在 session 外触发)会只剩一行无堆栈的 warning,极难定位。
+            logger.exception("backtest match %s failed", c.match_id)
 
     # 汇总统计
     if summary.scored > 0:
diff --git a/src/llm/context_builder.py b/src/llm/context_builder.py
index c40934f..eee7699 100644
--- a/src/llm/context_builder.py
+++ b/src/llm/context_builder.py
@@ -328,9 +328,19 @@ async def _load_match(db, match_id: int) -> Match:
 
 
 async def _get_form(db, team_id: int, before, *, limit: int = 5) -> list[Match]:
-    """某队近 N 场(已完赛)。before=None 表示不限制(预测赛前的场景由调用方保证)。"""
+    """某队近 N 场(已完赛)。before=None 表示不限制(预测赛前的场景由调用方保证)。
+
+    必须预加载 stats / home_team / away_team:切片函数会读取这些关系,
+    而 async session 下惰性加载会抛 MissingGreenlet。
+    (models.py 已声明 lazy="selectin",此处显式声明以固化查询意图。)
+    """
     stmt = (
         select(Match)
+        .options(
+            selectinload(Match.stats),
+            selectinload(Match.home_team),
+            selectinload(Match.away_team),
+        )
         .where(Match.match_status == "finished")
         .where(Match.home_goals.is_not(None))
         .where((Match.home_team_id == team_id) | (Match.away_team_id == team_id))
@@ -344,9 +354,13 @@ async def _get_form(db, team_id: int, before, *, limit: int = 5) -> list[Match]:
 
 
 async def _get_h2h(db, home_id: int, away_id: int, before, *, limit: int = 5) -> list[Match]:
-    """两队交锋史。"""
+    """两队交锋史。需预加载 home_team / away_team(切片输出队名)。"""
     stmt = (
         select(Match)
+        .options(
+            selectinload(Match.home_team),
+            selectinload(Match.away_team),
+        )
         .where(Match.match_status == "finished")
         .where(Match.home_goals.is_not(None))
         .where(
@@ -363,9 +377,17 @@ async def _get_h2h(db, home_id: int, away_id: int, before, *, limit: int = 5) ->
 
 
 async def _get_home_away(db, team_id: int, side: str, before, *, limit: int = 10) -> list[Match]:
-    """某队主场/客场近 N 场。side='home' 取主场,'away' 取客场。"""
+    """某队主场/客场近 N 场。side='home' 取主场,'away' 取客场。
+
+    当前只用标量字段,但统一预加载以免后续扩展时踩坑。
+    """
     stmt = (
         select(Match)
+        .options(
+            selectinload(Match.stats),
+            selectinload(Match.home_team),
+            selectinload(Match.away_team),
+        )
         .where(Match.match_status == "finished")
         .where(Match.home_goals.is_not(None))
         .order_by(Match.match_date.desc())
diff --git a/src/llm/predict.py b/src/llm/predict.py
index eb07b50..0872ede 100644
--- a/src/llm/predict.py
+++ b/src/llm/predict.py
@@ -81,11 +81,22 @@ async def predict_match(
     model: str | None = None,
     prompt_version: str | None = None,
     mode: str = "multi",
+    use_cache: bool = True,
 ) -> "PredictResult | MultiPredictResult":
-    """预测入口。mode=multi(默认)走多 agent;mode=single 走单次调用。"""
+    """预测入口。mode=multi(默认)走多 agent;mode=single 走单次调用。
+
+    Args:
+        use_cache: 是否允许返回进程内缓存结果。回测必须传 False——
+            缓存命中不会新建 prediction 行,调用方会对同一个 prediction_id
+            反复 settle,把不同比赛的真实比分覆盖到同一条记录上。
+    """
     if mode == "single":
         return await _predict_single(
-            match_id, provider=provider, model=model, prompt_version=prompt_version
+            match_id,
+            provider=provider,
+            model=model,
+            prompt_version=prompt_version,
+            use_cache=use_cache,
         )
     from src.llm.agents.orchestrator import predict_match_multi
 
@@ -98,6 +109,7 @@ async def _predict_single(
     provider: LLMProvider | None = None,
     model: str | None = None,
     prompt_version: str | None = None,
+    use_cache: bool = True,
 ) -> PredictResult:
     """单次调用路径(原有实现)。"""
     if provider is None:
@@ -107,10 +119,11 @@ async def _predict_single(
     version = prompt_version or "v1"
 
     # 0. 查缓存(同 match+provider+model+version 5 分钟内直接返)
-    cached = _get_cached(match_id, settings.LLM_PROVIDER, provider.model, version)
-    if cached is not None:
-        logger.debug("predict cache hit match=%s", match_id)
-        return cached
+    if use_cache:
+        cached = _get_cached(match_id, settings.LLM_PROVIDER, provider.model, version)
+        if cached is not None:
+            logger.debug("predict cache hit match=%s", match_id)
+            return cached
 
     # 1. 拼上下文
     ctx = await build_context(match_id)
@@ -191,6 +204,7 @@ async def _predict_single(
             raw=resp.raw,
         )
 
-        # 5. 写入缓存
-        _set_cached(match_id, settings.LLM_PROVIDER, provider.model, version, result)
+        # 5. 写入缓存(仅当允许缓存时)
+        if use_cache:
+            _set_cached(match_id, settings.LLM_PROVIDER, provider.model, version, result)
         return result