安全加固 + 数据管线修复 + 质量改进(第2批)

安全加固:
- /api/v1/predict 限流改用 get_client_ip 防 X-Forwarded-For 伪造
- 新增 TRUST_PROXY_HEADERS/REQUIRE_ADMIN_AUTH 配置,默认 fail-closed
- 生产环境管理接口未配置鉴权时拒绝(503),不再放行

数据管线修复:
- injuries IntegrityError 改用 begin_nested(SAVEPOINT)隔离批次
- injuries 空名单 vs 未配置语义区分(has_data 精确标记)
- bzzoiro 统计字段映射(shots/possession/cards) + 入库条件放宽
- available_at 语义收紧(回测防泄漏)
- team_names NFKD 去变音双重查找 + 补齐变体键

预测系统改进:
- multi-agent 全专家失败时 status=degraded 跳过终裁
- agent_weights 独立持久化到 predictions 表

新增迁移:
- 0014_predictions_agent_weights.py

新增测试(8个文件):
- test_ip_spoofing.py: 限流防伪造
- test_require_admin_fail_closed.py: 生产 fail-closed
- test_injuries_integrity_rollback.py: SAVEPOINT 隔离
- test_injuries_slice_semantic.py: 空名单 vs 未配置
- test_available_at.py: 回测防泄漏
- test_bzzoirot_stats.py: 统计字段映射
- test_team_names_normalize.py: NFKD 变体
- test_multi_agent_degraded.py: 全失败 degraded
- test_agent_weights_persist.py: 权重持久化

src/llm/context_builder.py

@@ -40,11 +40,23 @@ def _outcome(home_goals: int, away_goals: int, side: str) -> str:
 
 
 def _is_stats_available(stats, before) -> bool:
-    """检查统计数据在 cutoff 时间是否已可用。"""
+    """检查统计数据在 cutoff 时间是否已可用。
+
+    available_at 语义:该条统计「对外可被使用」的最早时间,
+    至少不得早于比赛结束。用于回测防泄漏。
+
+    规则:
+      - before is None(实盘):available_at 为 None 时允许(兼容旧数据)
+      - before is not None(回测):available_at 为 None 视为不可用(保守)
+      - available_at > cutoff:不可用(数据在 cutoff 之后才生成)
+    """
     if before is None:
+        # 实盘模式:无时间信息时允许(兼容旧数据)
         return True
+    # 回测模式(cutoff 不为 None):
+    # available_at 为 None → 无法确认是否在 cutoff 前可用,保守视为不可用
     if stats.available_at is None:
-        return True  # 无时间信息时保守处理:允许使用
+        return False
     return stats.available_at <= before
 
 
@@ -330,34 +342,62 @@ async def injuries_slice(header: MatchHeader, *, before=None, db: AsyncSession |
 
     before=cutoff: 只使用 cutoff 之前已采集的伤停数据,防回测泄漏。
     db: 可选共享 session(见模块 docstring)。
+
+    语义区分:
+      - 查询成功 + 空结果 → has_data=True(明确知道「无人伤停」)
+      - 源未配置 / 查询失败 → has_data=False(无法判断,跳过 LLM)
     """
-    from src.data.injuries import get_injuries_for_match
+    from src.data.injuries import get_injuries_for_match, InjuryQueryResult
 
     cutoff = before or header.match_dt
     if db is not None:
-        home_injuries = await get_injuries_for_match(db, header.home_team_id, cutoff, as_of=cutoff)
-        away_injuries = await get_injuries_for_match(db, header.away_team_id, cutoff, as_of=cutoff)
+        home_result = await get_injuries_for_match(db, header.home_team_id, cutoff, as_of=cutoff)
+        away_result = await get_injuries_for_match(db, header.away_team_id, cutoff, as_of=cutoff)
     else:
         async with AsyncSessionLocal() as new_db:
-            home_injuries = await get_injuries_for_match(new_db, header.home_team_id, cutoff, as_of=cutoff)
-            away_injuries = await get_injuries_for_match(new_db, header.away_team_id, cutoff, as_of=cutoff)
+            home_result = await get_injuries_for_match(new_db, header.home_team_id, cutoff, as_of=cutoff)
+            away_result = await get_injuries_for_match(new_db, header.away_team_id, cutoff, as_of=cutoff)
+
+    # 判断是否有有效查询结果
+    # 两队都成功查询(即使为空) → has_data=True
+    # 任一查询失败或源未配置 → has_data=False
+    both_succeeded = (
+        home_result.query_status == "success"
+        and away_result.query_status == "success"
+    )
+    any_configured = (
+        home_result.query_status != "source_not_configured"
+        or away_result.query_status != "source_not_configured"
+    )
 
     lines = ["── 阵容完整性 ──"]
     n_records = 0
-    for label, injuries in (("主队", home_injuries), ("客队", away_injuries)):
-        if injuries:
-            n_records += len(injuries)
-            lines.append(f"  {label}伤停({len(injuries)}人):")
-            for inj in injuries[:8]:  # 最多显示 8 条
+
+    for label, result in (("主队", home_result), ("客队", away_result)):
+        if result.query_status == "source_not_configured":
+            lines.append(f"  {label}: 伤停源未配置")
+        elif result.query_status == "query_error":
+            lines.append(f"  {label}: 查询异常")
+        elif result.records:
+            n_records += len(result.records)
+            lines.append(f"  {label}伤停({len(result.records)}人):")
+            for inj in result.records[:8]:
                 reason = inj.reason or inj.injury_type or "未知"
                 lines.append(f"    - {inj.player_name}: {reason}")
-            if len(injuries) > 8:
-                lines.append(f"    ...及其他 {len(injuries) - 8} 人")
+            if len(result.records) > 8:
+                lines.append(f"    ...及其他 {len(result.records) - 8} 人")
         else:
-            lines.append(f"  {label}: 无伤停数据")
+            # 查询成功但无人伤停
+            lines.append(f"  {label}: 当前无伤停记录")
 
-    if n_records == 0:
-        return SliceResult(text="── 阵容完整性 ──\n  无数据", has_data=False, n_records=0)
+    # 决定 has_data:
+    # - 两队都成功查询(即使为空) → True(明确知道名单)
+    # - 源未配置且无数据 → False
+    has_data = both_succeeded or (any_configured and n_records > 0)
+
+    if not has_data:
+        # 保留详细状态文案(伤停源未配置/查询异常),而非通用「无数据」
+        return SliceResult(text="\n".join(lines), has_data=False, n_records=0)
 
     return SliceResult(text="\n".join(lines), has_data=True, n_records=n_records)