feat(routing): 完善平台满载避让与故障轮转

为未配置并发上限的平台增加基于运行和等待任务数的软负载，并按非满载、有效优先级、缓存亲和与负载稳定排序。\n\n平台模型 RPM、TPM 和并发额度竞争失败时只轮转候选，不触发冷却、禁用或降级；用户组额度保持不可绕过。补齐异步冷却排队恢复、满载原因、选择原因与低基数指标。\n\n验证：go test ./...；go vet ./...；PostgreSQL 原子额度、准入队列、Worker 容量回收及故障策略 HTTP acceptance。

apps/api/internal/securityevents/metrics.go

@@ -79,6 +79,13 @@ type Metrics struct {
 	taskAdmissionCancelled      atomic.Uint64
 	taskAdmissionExpired        atomic.Uint64
 	taskAdmissionMigrated       atomic.Uint64
+	candidateNormalRotation     atomic.Uint64
+	candidateFullAvoided        atomic.Uint64
+	candidateQuotaRaceRotated   atomic.Uint64
+	candidateCooldownSkipped    atomic.Uint64
+	candidateDisabledSkipped    atomic.Uint64
+	candidateAllFullQueued      atomic.Uint64
+	candidateRoutingOther       atomic.Uint64
 	taskAdmissionWaitBuckets    [11]atomic.Uint64
 	taskAdmissionWaitMicros     atomic.Uint64
 }
@@ -239,6 +246,25 @@ func (m *Metrics) ObserveTaskAdmission(event string) {
 	}
 }
 
+func (m *Metrics) ObserveCandidateRouting(reason string) {
+	switch reason {
+	case "normal_rotation":
+		m.candidateNormalRotation.Add(1)
+	case "full_avoided":
+		m.candidateFullAvoided.Add(1)
+	case "quota_race_rotated":
+		m.candidateQuotaRaceRotated.Add(1)
+	case "cooldown_skipped":
+		m.candidateCooldownSkipped.Add(1)
+	case "disabled_skipped":
+		m.candidateDisabledSkipped.Add(1)
+	case "all_full_queued":
+		m.candidateAllFullQueued.Add(1)
+	default:
+		m.candidateRoutingOther.Add(1)
+	}
+}
+
 func (m *Metrics) ObserveTaskAdmissionWait(wait time.Duration) {
 	if wait < 0 {
 		wait = 0
@@ -436,6 +462,15 @@ func (m *Metrics) Handler(provider MetricsSnapshotProvider, issuer, audience str
 			{"concurrent", m.providerQuotaWaitConcurrent.Load()},
 			{"other", m.providerQuotaWaitOther.Load()},
 		})
+		outcomeCounters(w, "easyai_gateway_candidate_routing_total", "Candidate routing decisions by bounded reason.", []outcomeValue{
+			{"normal_rotation", m.candidateNormalRotation.Load()},
+			{"full_avoided", m.candidateFullAvoided.Load()},
+			{"quota_race_rotated", m.candidateQuotaRaceRotated.Load()},
+			{"cooldown_skipped", m.candidateCooldownSkipped.Load()},
+			{"disabled_skipped", m.candidateDisabledSkipped.Load()},
+			{"all_full_queued", m.candidateAllFullQueued.Load()},
+			{"other", m.candidateRoutingOther.Load()},
+		})
 		platformModelRateLimitUtilizationGauges(w, modelRateLimits)
 		plainGauge(w, "easyai_gateway_postgres_pool_max_connections", "Maximum PostgreSQL connections in this process pool.", int64(postgresPool.MaxConnections))
 		plainGauge(w, "easyai_gateway_postgres_pool_total_connections", "Current PostgreSQL connections in this process pool.", int64(postgresPool.TotalConnections))

apps/api/internal/securityevents/metrics_test.go

@@ -46,6 +46,12 @@ func TestMetricsExposeBoundedOutcomesAndState(t *testing.T) {
 	metrics.SetWorkerLoad(12, 3, 8, 2, 5, 1, "busy")
 	metrics.ObserveProviderQuotaWait("rpm")
 	metrics.ObserveProviderQuotaWait("tpm_total")
+	metrics.ObserveCandidateRouting("normal_rotation")
+	metrics.ObserveCandidateRouting("full_avoided")
+	metrics.ObserveCandidateRouting("quota_race_rotated")
+	metrics.ObserveCandidateRouting("cooldown_skipped")
+	metrics.ObserveCandidateRouting("disabled_skipped")
+	metrics.ObserveCandidateRouting("all_full_queued")
 	metrics.ObserveAsyncWorkerResize("success")
 	metrics.ObserveConcurrencyLeaseRenewal("success")
 	metrics.ObserveConcurrencyLeaseRenewal("lost")
@@ -86,6 +92,12 @@ func TestMetricsExposeBoundedOutcomesAndState(t *testing.T) {
 		`easyai_gateway_worker_pressure_state 1`,
 		`easyai_gateway_provider_quota_waits_total{outcome="rpm"} 1`,
 		`easyai_gateway_provider_quota_waits_total{outcome="tpm"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="normal_rotation"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="full_avoided"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="quota_race_rotated"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="cooldown_skipped"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="disabled_skipped"} 1`,
+		`easyai_gateway_candidate_routing_total{outcome="all_full_queued"} 1`,
 		`easyai_gateway_platform_model_rate_limit_utilization{platform_model_id="platform-model-1",metric="concurrent"} 0.750000`,
 		`easyai_gateway_async_worker_resizes_total{outcome="success"} 1`,
 		`easyai_gateway_concurrency_lease_renewals_total{outcome="success"} 1`,