fix: 治理 Prometheus 指标基数爆炸和内存缓存无界增长

- 删除高基数标签 key_id/model，移除未使用的指标 concurrency_slots_in_use/streaming_request_duration_seconds
- HTTP 客户端池：命名客户端添加 LRU 淘汰上限，tunnel 客户端添加 LRU 淘汰和 last_used_time 追踪
- ResilienceManager：last_errors 改用 deque(maxlen)，error_stats/circuit_breakers 添加上限淘汰
- HealthMonitor：_circuit_history 改用 deque(maxlen)
- ProviderHealthTracker：清理已无记录的过期 key
- PrometheusPlugin：动态指标数量添加上限，超限后拒绝创建
- 中间件使用路由模板替代实际路径，防止动态路径段导致标签爆炸

src/core/metrics.py

@@ -4,11 +4,11 @@ Prometheus metrics for monitoring
 
 from prometheus_client import Counter, Gauge, Histogram
 
-# 并发槽位占用时长分布
+# 并发槽位占用时长分布（按异常类型聚合，不按 key_id 拆分）
 concurrency_slot_duration_seconds = Histogram(
     "concurrency_slot_duration_seconds",
     "Duration of concurrency slot occupation in seconds",
-    ["key_id", "exception"],
+    ["exception"],
     buckets=[0.1, 0.5, 1, 5, 10, 30, 60, 120, 300, 600],  # 0.1s 到 10 分钟
 )
 
@@ -16,20 +16,7 @@ concurrency_slot_duration_seconds = Histogram(
 concurrency_slot_release_total = Counter(
     "concurrency_slot_release_total",
     "Total number of concurrency slot releases",
-    ["key_id", "exception"],
-)
-
-# 当前并发槽位使用数
-concurrency_slots_in_use = Gauge(
-    "concurrency_slots_in_use", "Current number of concurrency slots in use", ["key_id"]
-)
-
-# 流式请求时长分布
-streaming_request_duration_seconds = Histogram(
-    "streaming_request_duration_seconds",
-    "Duration of streaming requests in seconds",
-    ["key_id", "status"],
-    buckets=[1, 5, 10, 30, 60, 120, 300, 600, 1800],  # 1s 到 30 分钟
+    ["exception"],
 )
 
 # 请求总数（按类型）
@@ -118,5 +105,5 @@ billing_invariant_violation_total = Counter(
 antigravity_degradation_total = Counter(
     "aether_antigravity_degradation_total",
     "Count of Antigravity signature degradation (rectification) events",
-    ["stage", "model"],
+    ["stage"],
 )

src/core/provider_health.py

@@ -110,6 +110,16 @@ class ProviderHealthTracker:
             if all(current_time - t > self.recovery_time for t in self.successes[provider_name]):
                 self.priority_adjustments[provider_name] = 0
 
+        # 清理已无记录且无优先级调整的 key，防止 dict 无限增长
+        if (
+            not self.failures[provider_name]
+            and not self.successes[provider_name]
+            and self.priority_adjustments.get(provider_name, 0) == 0
+        ):
+            self.failures.pop(provider_name, None)
+            self.successes.pop(provider_name, None)
+            self.priority_adjustments.pop(provider_name, None)
+
     def _get_status_label(self, failure_rate: float, recent_failures: int) -> str:
         """根据失败率返回状态标签"""
         if recent_failures >= self.failure_threshold:

src/core/resilience.py

@@ -11,6 +11,7 @@ import threading
 import time
 import traceback
 import uuid
+from collections import deque
 from collections.abc import Callable
 from contextlib import asynccontextmanager
 from datetime import datetime, timezone
@@ -115,11 +116,15 @@ class CircuitBreaker:
 class ResilienceManager:
     """系统韧性管理器"""
 
+    _MAX_ERROR_STATS = 500
+    _MAX_CIRCUIT_BREAKERS = 200
+    _MAX_LAST_ERRORS = 100
+
     def __init__(self) -> None:
         self.error_patterns: list[ErrorPattern] = []
         self.circuit_breakers: dict[str, CircuitBreaker] = {}
         self.error_stats: dict[str, int] = {}
-        self.last_errors: list[dict[str, Any]] = []
+        self.last_errors: deque[dict[str, Any]] = deque(maxlen=self._MAX_LAST_ERRORS)
         self._setup_default_patterns()
 
     def _setup_default_patterns(self) -> None:
@@ -203,6 +208,11 @@ class ResilienceManager:
     def get_circuit_breaker(self, key: str) -> CircuitBreaker:
         """获取或创建熔断器"""
         if key not in self.circuit_breakers:
+            # 淘汰已恢复的旧熔断器，防止无界增长
+            if len(self.circuit_breakers) >= self._MAX_CIRCUIT_BREAKERS:
+                closed_keys = [k for k, cb in self.circuit_breakers.items() if cb.state == "closed"]
+                for k in closed_keys:
+                    del self.circuit_breakers[k]
             self.circuit_breakers[key] = CircuitBreaker()
         return self.circuit_breakers[key]
 
@@ -226,13 +236,18 @@ class ResilienceManager:
         }
 
         self.last_errors.append(error_info)
-        # 只保留最近100个错误
-        if len(self.last_errors) > 100:
-            self.last_errors.pop(0)
 
-        # 更新错误统计
+        # 更新错误统计（超上限时淘汰计数最低的条目）
         error_key = f"{type(error).__name__}:{operation}"
         self.error_stats[error_key] = self.error_stats.get(error_key, 0) + 1
+        if len(self.error_stats) > self._MAX_ERROR_STATS:
+            min_key = min(
+                (k for k in self.error_stats if k != error_key),
+                key=lambda k: self.error_stats[k],
+                default=None,
+            )
+            if min_key is not None:
+                del self.error_stats[min_key]
 
         # 查找匹配的错误处理模式
         pattern = self._find_matching_pattern(error)