perf: 优化请求鉴权链路并批量化统计/调度查询

- 为 Pipeline/Context 增加按需读取请求体能力，支持 async 懒加载 JSON body
- 为 chat/cli/video/claude/openai-cli 适配器关闭默认预读，减少无效 body 读取与超时风险
- 将本地登录、JWT 用户加载、API Key 鉴权迁移到线程池隔离会话执行，避免阻塞事件循环
- 为 API Key 鉴权返回结构化余额结果，并在主请求会话中重新绑定 user/api_key 后再校验状态、过期和锁定信息
- 为 management/user token 前缀认证引入独立会话与结果回绑，避免跨会话对象写入失效

- 为 Usage 余额检查补充结构化返回，统一透出 remaining 与欠费/不可用文案映射
- 为用户与管理端活跃请求查询增加 maintain_status 开关，避免轮询指定 id 时误触发状态修复
- 重写 user_me usage 汇总逻辑，支持 group_by=None 的粗粒度聚合
- 修正 provider 维度成功率与平均响应时间统计，基于 success_count 和成功响应耗时汇总计算
- 前端 Usage 轮询由 setInterval 改为串行 setTimeout，避免并发轮询叠加

- 为 StatsAggregator 增加按本地日期批量计算百分位能力，替代逐天 fan-out 查询
- 为混合统计查询合并连续实时日期区间，并批量读取 StatsDaily，减少逐日查询次数
- 为用户日统计增加批量聚合入口，替代逐用户循环聚合
- 为系统配置导出改用 selectinload 预加载 provider 关联数据，减少 N+1 查询
- 为管理员用户列表增加钱包批量查询，避免逐用户回表

- 为调度器增加 provider 轻量引用预过滤，先按 allowed_providers 缩小范围再加载完整 provider 图
- 为 CandidateBuilder 增加 provider refs/provider_ids 查询能力，保留分页顺序
- 为模型缓存增加 provider_model_mappings 索引缓存与 model_mappings 规则缓存，减少重复全量扫描
- 为请求候选中间态改为 flush/batch commit，降低 pending/streaming 状态切换的事务往返
- 为钱包访问结果补充 balance_snapshot，并抽取余额快照复用逻辑

- 补充 pipeline、auth、admin users、user_me usage、stats aggregator、model cache、
  scheduler、wallet、request candidate 等回归与契约测试

tests/contracts/test_scheduler_list_all_candidates_contract.py

@@ -69,26 +69,33 @@ async def test_list_all_candidates_returns_provider_batch_count_even_when_candid
     global_model = _make_global_model(gid="gm1", name="gpt-4o")
 
     with patch.object(scheduler, "_ensure_initialized", new=AsyncMock(return_value=None)):
-        with patch.object(scheduler._candidate_builder, "_query_providers", return_value=providers):
-            with patch(
-                "src.services.scheduling.aware_scheduler.ModelCacheService.get_global_model_by_name",
-                new=AsyncMock(return_value=global_model),
-            ):
+        with patch.object(
+            scheduler._candidate_builder,
+            "_query_provider_refs",
+            return_value=[("p1", "p1"), ("p2", "p2")],
+        ):
+            with patch.object(scheduler._candidate_builder, "_query_providers") as query_providers:
                 with patch(
-                    "src.services.scheduling.aware_scheduler.SystemConfigService.is_format_conversion_enabled",
-                    return_value=True,
+                    "src.services.scheduling.aware_scheduler.ModelCacheService.get_global_model_by_name",
+                    new=AsyncMock(return_value=global_model),
                 ):
-                    candidates, global_model_id, provider_batch_count = (
-                        await scheduler.list_all_candidates(
-                            db=db,
-                            api_format="openai:chat",
-                            model_name="gpt-4o",
-                            affinity_key=None,
-                            user_api_key=user_api_key,  # type: ignore[arg-type]
-                            provider_offset=0,
-                            provider_limit=20,
+                    with patch(
+                        "src.services.scheduling.aware_scheduler.SystemConfigService.is_format_conversion_enabled",
+                        return_value=True,
+                    ):
+                        candidates, global_model_id, provider_batch_count = (
+                            await scheduler.list_all_candidates(
+                                db=db,
+                                api_format="openai:chat",
+                                model_name="gpt-4o",
+                                affinity_key=None,
+                                user_api_key=user_api_key,  # type: ignore[arg-type]
+                                provider_offset=0,
+                                provider_limit=20,
+                            )
                         )
-                    )
+
+    query_providers.assert_not_called()
 
     assert candidates == []
     assert global_model_id == "gm1"

tests/contracts/test_scheduler_provider_prefilter_contract.py

@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from src.services.scheduling.aware_scheduler import CacheAwareScheduler
+
+
+def _make_db() -> MagicMock:
+    db = MagicMock()
+    db.new = []
+    db.dirty = []
+    db.deleted = []
+    db.in_transaction.return_value = False
+    return db
+
+
+def _make_global_model() -> SimpleNamespace:
+    return SimpleNamespace(
+        id="gm1",
+        name="gpt-4o",
+        is_active=True,
+        config={},
+        supported_capabilities=[],
+    )
+
+
+@pytest.mark.asyncio
+async def test_list_all_candidates_prefilters_provider_graph_by_allowed_providers() -> None:
+    scheduler = CacheAwareScheduler()
+    scheduler.scheduling_mode = CacheAwareScheduler.SCHEDULING_MODE_FIXED_ORDER
+
+    db = _make_db()
+    global_model = _make_global_model()
+    user_api_key = SimpleNamespace(
+        id="ak1",
+        allowed_providers=["provider-b"],
+        allowed_models=None,
+        allowed_api_formats=None,
+        user=None,
+    )
+
+    filtered_provider = SimpleNamespace(
+        id="provider-b",
+        name="provider-b",
+        endpoints=[],
+        models=[],
+        provider_priority=2,
+    )
+
+    with patch.object(scheduler, "_ensure_initialized", new=AsyncMock(return_value=None)):
+        with patch.object(
+            scheduler._candidate_builder,
+            "_query_provider_refs",
+            return_value=[("provider-a", "provider-a"), ("provider-b", "provider-b")],
+        ) as refs_mock:
+            with patch.object(
+                scheduler._candidate_builder,
+                "_query_providers",
+                return_value=[filtered_provider],
+            ) as providers_mock:
+                with patch(
+                    "src.services.scheduling.aware_scheduler.ModelCacheService.get_global_model_by_name",
+                    new=AsyncMock(return_value=global_model),
+                ):
+                    with patch(
+                        "src.services.scheduling.aware_scheduler.SystemConfigService.is_format_conversion_enabled",
+                        return_value=True,
+                    ):
+                        with patch.object(
+                            scheduler._candidate_builder,
+                            "_build_candidates",
+                            new=AsyncMock(return_value=[]),
+                        ):
+                            candidates, global_model_id, provider_batch_count = (
+                                await scheduler.list_all_candidates(
+                                    db=db,
+                                    api_format="openai:chat",
+                                    model_name="gpt-4o",
+                                    affinity_key=None,
+                                    user_api_key=user_api_key,  # type: ignore[arg-type]
+                                    provider_offset=0,
+                                    provider_limit=20,
+                                )
+                            )
+
+    assert candidates == []
+    assert global_model_id == "gm1"
+    assert provider_batch_count == 2
+    refs_mock.assert_called_once()
+    providers_mock.assert_called_once_with(db=db, provider_ids=["provider-b"])