refactor: defer ProviderAPIKey 大 JSON 字段，减少调度器和模型获取的内存占用

- candidate_builder: defer adjustment_history/utilization_samples/upstream_metadata，
  号池路径在释放 DB 连接前预计算 _pool_account_state 避免后续 N+1 查询
- pool/manager: 优先读取预计算的 _pool_account_state，fallback 到实时解析
- fetch_scheduler: 三处 ProviderAPIKey 查询改用 defer/load_only 排除无关字段

src/services/model/fetch_scheduler.py

@@ -19,7 +19,7 @@ from dataclasses import dataclass
 from datetime import datetime, timezone
 from typing import Any
 
-from sqlalchemy.orm import joinedload
+from sqlalchemy.orm import defer, joinedload, load_only
 
 from src.core.cache_service import CacheService
 from src.core.crypto import crypto_service
@@ -284,7 +284,19 @@ class ModelFetchScheduler:
         """更新 Key 的错误信息（独立事务）"""
         try:
             with create_session() as db:
-                key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
+                key = (
+                    db.query(ProviderAPIKey)
+                    .options(
+                        defer(ProviderAPIKey.adjustment_history),
+                        defer(ProviderAPIKey.utilization_samples),
+                        defer(ProviderAPIKey.health_by_format),
+                        defer(ProviderAPIKey.circuit_breaker_by_format),
+                        defer(ProviderAPIKey.upstream_metadata),
+                        defer(ProviderAPIKey.allowed_models),
+                    )
+                    .filter(ProviderAPIKey.id == key_id)
+                    .first()
+                )
                 if key:
                     key.last_models_fetch_at = datetime.now(timezone.utc)
                     key.last_models_fetch_error = error_msg
@@ -384,7 +396,18 @@ class ModelFetchScheduler:
         with create_session() as db:
             key = (
                 db.query(ProviderAPIKey)
-                .options(joinedload(ProviderAPIKey.provider))
+                .options(
+                    load_only(
+                        ProviderAPIKey.id,
+                        ProviderAPIKey.is_active,
+                        ProviderAPIKey.auto_fetch_models,
+                        ProviderAPIKey.api_key,
+                        ProviderAPIKey.auth_type,
+                        ProviderAPIKey.auth_config,
+                        ProviderAPIKey.provider_id,
+                        ProviderAPIKey.proxy,
+                    ),
+                )
                 .filter(ProviderAPIKey.id == key_id)
                 .first()
             )
@@ -482,7 +505,18 @@ class ModelFetchScheduler:
 
         with create_session() as db:
             # 重新获取 Key（因为之前的连接已关闭）
-            key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
+            # defer 不需要的大 JSON 字段，减少内存占用
+            key = (
+                db.query(ProviderAPIKey)
+                .options(
+                    defer(ProviderAPIKey.adjustment_history),
+                    defer(ProviderAPIKey.utilization_samples),
+                    defer(ProviderAPIKey.health_by_format),
+                    defer(ProviderAPIKey.circuit_breaker_by_format),
+                )
+                .filter(ProviderAPIKey.id == key_id)
+                .first()
+            )
             if not key:
                 logger.warning(f"Key {key_id} 在更新时不存在")
                 return "error"

src/services/provider/pool/manager.py

@@ -221,16 +221,22 @@ class PoolManager:
                         pass
 
         # --- 3. Classify candidates -----------------------------------
-        # Pre-compute account states to avoid repeated dict parsing inside loop.
+        # Use precomputed account states from CandidateBuilder when available
+        # (upstream_metadata is deferred on pool keys to save memory).
+        # Fall back to on-the-fly resolution for non-pool candidates.
         account_states: dict[str, Any] = {}
         for c in candidates:
             kid = str(c.key.id)
             if kid not in account_states:
-                account_states[kid] = resolve_pool_account_state(
-                    provider_type=provider_type,
-                    upstream_metadata=getattr(c.key, "upstream_metadata", None),
-                    oauth_invalid_reason=getattr(c.key, "oauth_invalid_reason", None),
-                )
+                precomputed = getattr(c.key, "_pool_account_state", None)
+                if precomputed is not None:
+                    account_states[kid] = precomputed
+                else:
+                    account_states[kid] = resolve_pool_account_state(
+                        provider_type=provider_type,
+                        upstream_metadata=getattr(c.key, "upstream_metadata", None),
+                        oauth_invalid_reason=getattr(c.key, "oauth_invalid_reason", None),
+                    )
 
         sticky_candidate: ProviderCandidate | None = None
         available: list[ProviderCandidate] = []
@@ -582,16 +588,20 @@ class PoolManager:
                         pass
 
         # --- 3. Classify keys -------------------------------------------------
-        # Pre-compute account states to avoid repeated dict parsing inside loop.
+        # Use precomputed account states when available.
         account_states_sk: dict[str, Any] = {}
         for k in keys:
             kid = str(k.id)
             if kid not in account_states_sk:
-                account_states_sk[kid] = resolve_pool_account_state(
-                    provider_type=self.provider_type,
-                    upstream_metadata=getattr(k, "upstream_metadata", None),
-                    oauth_invalid_reason=getattr(k, "oauth_invalid_reason", None),
-                )
+                precomputed = getattr(k, "_pool_account_state", None)
+                if precomputed is not None:
+                    account_states_sk[kid] = precomputed
+                else:
+                    account_states_sk[kid] = resolve_pool_account_state(
+                        provider_type=self.provider_type,
+                        upstream_metadata=getattr(k, "upstream_metadata", None),
+                        oauth_invalid_reason=getattr(k, "oauth_invalid_reason", None),
+                    )
 
         sticky_key: ProviderAPIKey | None = None
         available: list[ProviderAPIKey] = []

src/services/scheduling/candidate_builder.py

@@ -35,6 +35,9 @@ from src.models.database import (
 )
 from src.services.health.monitor import health_monitor
 from src.services.provider.format import normalize_endpoint_signature
+from src.services.provider.pool.account_state import (
+    resolve_pool_account_state as _resolve_pool_account_state,
+)
 from src.services.scheduling.quota_skipper import is_key_quota_exhausted
 from src.services.scheduling.utils import release_db_connection_before_await
 
@@ -97,10 +100,13 @@ class CandidateBuilder:
             db.query(Provider)
             .options(
                 # 预加载 Provider 级别的 api_keys
-                # defer 排除调度热路径不需要的字段，减少内存占用
-                # 凭证类字段（api_key/auth_config）在执行阶段才由 get_provider_auth() 按需加载
-                # 注意：adjustment_history/utilization_samples 在并发检查阶段
-                # (AdaptiveReservationManager) 被访问，不可 defer
+                # defer 排除调度热路径不需要的大 JSON 字段，减少号池场景内存占用
+                # - 凭证类: api_key/auth_config 在执行阶段由 get_provider_auth() 按需加载
+                # - adjustment_history/utilization_samples: 仅 AdaptiveReservationManager
+                #   在并发检查时读取单个 key，号池/非号池均可 lazy load
+                # - upstream_metadata: 号池模式在 _build_candidates 中预计算账号封禁
+                #   状态并挂到 key._pool_account_state，排序阶段不再需要原始 JSON；
+                #   非号池模式 key 少，lazy load 可忽略
                 selectinload(Provider.api_keys)
                 .defer(ProviderAPIKey.api_key)
                 .defer(ProviderAPIKey.auth_config)
@@ -113,7 +119,10 @@ class CandidateBuilder:
                 .defer(ProviderAPIKey.last_models_fetch_at)
                 .defer(ProviderAPIKey.last_models_fetch_error)
                 .defer(ProviderAPIKey.max_probe_interval_minutes)
-                .defer(ProviderAPIKey.expires_at),
+                .defer(ProviderAPIKey.expires_at)
+                .defer(ProviderAPIKey.adjustment_history)
+                .defer(ProviderAPIKey.utilization_samples)
+                .defer(ProviderAPIKey.upstream_metadata),
                 # 预加载 endpoints（用于按 api_format 选择请求配置）
                 selectinload(Provider.endpoints),
                 # 同时加载 models 和 global_model 关系
@@ -570,15 +579,33 @@ class CandidateBuilder:
                 if pool_cfg is not None:
                     # 号池优化：跳过逐 key 的 _check_key_availability 检查，
                     # 直接收集全部 active key，将检查推迟到 PoolManager 排序后分页执行。
-                    # 在此释放 DB 连接，因为后续的 PoolManager 排序涉及大量 Redis 操作，
-                    # 避免在 Redis I/O 期间长时间占用 DB 连接池。
-                    release_db_connection_before_await(db)
 
                     pool_keys = list(keys_to_check)
 
                     if not pool_keys:
                         continue
 
+                    # 在释放 DB 连接前预计算账号封禁状态并挂到 Key 对象上。
+                    # upstream_metadata 是 deferred 字段，逐条 lazy load 会产生 N+1 查询；
+                    # 这里集中触发后，PoolManager 排序时直接读取 _pool_account_state 即可。
+                    provider_type_str = (
+                        str(getattr(provider, "provider_type", "") or "").strip().lower() or None
+                    )
+                    for pk in pool_keys:
+                        setattr(
+                            pk,
+                            "_pool_account_state",
+                            _resolve_pool_account_state(
+                                provider_type=provider_type_str,
+                                upstream_metadata=getattr(pk, "upstream_metadata", None),
+                                oauth_invalid_reason=getattr(pk, "oauth_invalid_reason", None),
+                            ),
+                        )
+
+                    # 释放 DB 连接，因为后续的 PoolManager 排序涉及大量 Redis 操作，
+                    # 避免在 Redis I/O 期间长时间占用 DB 连接池。
+                    release_db_connection_before_await(db)
+
                     provider_priority_raw = getattr(provider, "provider_priority", None)
                     try:
                         provider_priority = (