feat: 十轮网关优化 - 安全加固/可观测性/性能/可靠性

- SSE Keepalive Ping (15s心跳防止代理断连)
- Timing HTTP 头 (X-Timing-Queue/Inference/Total-Ms)
- Adapter Request-ID 传播到后端
- Session 清理日志回调
- Server 安全加固 (ReadHeaderTimeout/MaxHeaderBytes 防 slowloris)
- Usage Tracker 数据保留清理 (retentionDays + 定期清理)
- Config Reload 后 Adapter Registry 更新 (RegisterIfAbsent + RWMutex)
- Rate Limiter 空闲 Bucket 清理 (30分钟过期)
- Shutdown Drain 超时可配置 (ShutdownDrainSeconds)
- Config 模型字段校验增强 (provider/endpoint/actual_model)
- Auth 过期 Key 自动清理 (5分钟扫描)
- Admin API Rate Limiting
- Adapter Health Check 独立超时 (每个 adapter 3s)
- TCP 连接阶段超时 (DialContext 5s + KeepAlive 30s)
- 幂等键缓存、审计日志、Gzip 中间件、CORS Expose Headers
- Backpressure 响应头、熔断器 Prometheus 指标
- 连接池优化、Trace-ID 全链路传播

internal/observability/metrics.go

@@ -27,6 +27,7 @@ type Metrics struct {
 	runningTasks      int64
 	activeSessions    int64
 	backpressureLevel int64
+	breakerState      int64 // 0=closed, 1=open, 2=half_open
 
 	// Histograms (simplified as buckets)
 	gatewayLatencyBuckets    map[string]int64
@@ -110,6 +111,11 @@ func (m *Metrics) SetBackpressureLevel(level int) {
 	atomic.StoreInt64(&m.backpressureLevel, int64(level))
 }
 
+// SetBreakerState 设置熔断器状态 gauge（0=closed, 1=open, 2=half_open）。
+func (m *Metrics) SetBreakerState(state int) {
+	atomic.StoreInt64(&m.breakerState, int64(state))
+}
+
 // ObserveGatewayLatency records gateway latency in a histogram bucket.
 func (m *Metrics) ObserveGatewayLatency(ms int64) {
 	bucket := latencyBucket(ms)
@@ -142,34 +148,77 @@ func (m *Metrics) Handler() http.HandlerFunc {
 		w.Header().Set("Content-Type", "text/plain; version=0.0.4; charset=utf-8")
 
 		// Counters
+		fmt.Fprintf(w, "# HELP edgeai_requests_total Total requests by status\n")
+		fmt.Fprintf(w, "# TYPE edgeai_requests_total counter\n")
 		m.mu.RLock()
 		for key, val := range m.requestsTotal {
 			fmt.Fprintf(w, "edgeai_requests_total{%s} %d\n", key, val)
 		}
+		fmt.Fprintf(w, "# HELP edgeai_tasks_total Total tasks by final state\n")
+		fmt.Fprintf(w, "# TYPE edgeai_tasks_total counter\n")
 		for key, val := range m.tasksTotal {
 			fmt.Fprintf(w, "edgeai_tasks_total{%s} %d\n", key, val)
 		}
 		m.mu.RUnlock()
 
+		fmt.Fprintf(w, "# HELP edgeai_tokens_input_total Total input tokens processed\n")
+		fmt.Fprintf(w, "# TYPE edgeai_tokens_input_total counter\n")
 		fmt.Fprintf(w, "edgeai_tokens_input_total %d\n", atomic.LoadInt64(&m.tokensInputTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_tokens_output_total Total output tokens generated\n")
+		fmt.Fprintf(w, "# TYPE edgeai_tokens_output_total counter\n")
 		fmt.Fprintf(w, "edgeai_tokens_output_total %d\n", atomic.LoadInt64(&m.tokensOutputTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_cancellations_total Total cancelled requests\n")
+		fmt.Fprintf(w, "# TYPE edgeai_cancellations_total counter\n")
 		fmt.Fprintf(w, "edgeai_cancellations_total %d\n", atomic.LoadInt64(&m.cancellationsTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_queue_timeouts_total Total queue timeouts\n")
+		fmt.Fprintf(w, "# TYPE edgeai_queue_timeouts_total counter\n")
 		fmt.Fprintf(w, "edgeai_queue_timeouts_total %d\n", atomic.LoadInt64(&m.queueTimeoutsTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_first_token_timeouts_total Total first token timeouts\n")
+		fmt.Fprintf(w, "# TYPE edgeai_first_token_timeouts_total counter\n")
 		fmt.Fprintf(w, "edgeai_first_token_timeouts_total %d\n", atomic.LoadInt64(&m.firstTokenTimeoutsTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_inference_timeouts_total Total inference timeouts\n")
+		fmt.Fprintf(w, "# TYPE edgeai_inference_timeouts_total counter\n")
 		fmt.Fprintf(w, "edgeai_inference_timeouts_total %d\n", atomic.LoadInt64(&m.inferenceTimeoutsTotal))
+
+		fmt.Fprintf(w, "# HELP edgeai_degraded_requests_total Total requests served with degradation\n")
+		fmt.Fprintf(w, "# TYPE edgeai_degraded_requests_total counter\n")
 		fmt.Fprintf(w, "edgeai_degraded_requests_total %d\n", atomic.LoadInt64(&m.degradedRequestsTotal))
 
 		// Gauges
+		fmt.Fprintf(w, "# HELP edgeai_queue_length Current queue length\n")
+		fmt.Fprintf(w, "# TYPE edgeai_queue_length gauge\n")
 		fmt.Fprintf(w, "edgeai_queue_length %d\n", atomic.LoadInt64(&m.queueLength))
+
+		fmt.Fprintf(w, "# HELP edgeai_running_tasks Current running tasks\n")
+		fmt.Fprintf(w, "# TYPE edgeai_running_tasks gauge\n")
 		fmt.Fprintf(w, "edgeai_running_tasks %d\n", atomic.LoadInt64(&m.runningTasks))
+
+		fmt.Fprintf(w, "# HELP edgeai_active_sessions Current active sessions\n")
+		fmt.Fprintf(w, "# TYPE edgeai_active_sessions gauge\n")
 		fmt.Fprintf(w, "edgeai_active_sessions %d\n", atomic.LoadInt64(&m.activeSessions))
+
+		fmt.Fprintf(w, "# HELP edgeai_backpressure_level Current backpressure level (0-3)\n")
+		fmt.Fprintf(w, "# TYPE edgeai_backpressure_level gauge\n")
 		fmt.Fprintf(w, "edgeai_backpressure_level %d\n", atomic.LoadInt64(&m.backpressureLevel))
 
+		fmt.Fprintf(w, "# HELP edgeai_circuit_breaker_state Circuit breaker state (0=closed, 1=open, 2=half_open)\n")
+		fmt.Fprintf(w, "# TYPE edgeai_circuit_breaker_state gauge\n")
+		fmt.Fprintf(w, "edgeai_circuit_breaker_state %d\n", atomic.LoadInt64(&m.breakerState))
+
 		// Histograms
+		fmt.Fprintf(w, "# HELP edgeai_gateway_latency_bucket Gateway latency distribution in milliseconds\n")
+		fmt.Fprintf(w, "# TYPE edgeai_gateway_latency_bucket histogram\n")
 		m.mu.RLock()
 		for bucket, count := range m.gatewayLatencyBuckets {
 			fmt.Fprintf(w, "edgeai_gateway_latency_bucket{%s} %d\n", bucket, count)
 		}
+		fmt.Fprintf(w, "# HELP edgeai_first_token_latency_bucket First token latency distribution in milliseconds\n")
+		fmt.Fprintf(w, "# TYPE edgeai_first_token_latency_bucket histogram\n")
 		for bucket, count := range m.firstTokenLatencyBuckets {
 			fmt.Fprintf(w, "edgeai_first_token_latency_bucket{%s} %d\n", bucket, count)
 		}