feat(worker): 实现集群限流与自适应负载

保留平台模型 RPM、TPM 和并发策略语义，增加 PostgreSQL 集群级租约、饱和候选重选和多平台自动负载，避免突发任务固定等待首个平台。\n\n新增 Worker 实时负载采样、自适应 active/heavy 容量、心跳与管理端指标，并扩展本地 acceptance runner，覆盖三 Worker、同模型三平台 2/4/6 并发和 48 个带图视频突发任务。\n\n验证：go test ./...、go vet ./...、PostgreSQL 跨 Store 集成测试、gofmt、bash -n、ShellCheck 及本地集群 provider-burst 验收通过；48/48 成功，无越限、重复提交、重复计费、重复回调或终态资源泄漏。

apps/api/internal/runner/admission.go

@@ -205,23 +205,12 @@ func (s *Service) taskAdmissionScopes(
 }
 
 func acceptanceAdmissionScopes(task store.GatewayTask, scopes []store.AdmissionScope) []store.AdmissionScope {
-	if task.RunMode != "acceptance" && task.RunMode != "acceptance_canary" {
+	if task.RunMode != "acceptance" {
 		return scopes
 	}
 	out := append([]store.AdmissionScope(nil), scopes...)
 	for index := range out {
-		// Protocol-emulated acceptance measures Gateway and Worker capacity, so
-		// the isolated Run is bounded by the worker_capacity scope instead of a
-		// production supplier quota. The real acceptance_canary path deliberately
-		// retains the production platform-model concurrency limit.
-		if task.RunMode == "acceptance" && out[index].ScopeType == "platform_model" {
-			out[index].ConcurrentLimit = 0
-		}
-		if out[index].ConcurrentLimit <= 0 {
-			if task.RunMode != "acceptance" || out[index].ScopeType != "platform_model" {
-				continue
-			}
-		}
+		out[index].ScopeKey = acceptanceScopeKey(task, out[index].ScopeKey)
 		out[index].QueueLimit = acceptanceQueueLimit
 		out[index].MaxWaitSeconds = acceptanceQueueMaxWait
 	}
@@ -235,16 +224,21 @@ func acceptanceInfrastructureReservations(
 	if task.RunMode != "acceptance" {
 		return reservations
 	}
-	out := make([]store.RateLimitReservation, 0, len(reservations))
-	for _, reservation := range reservations {
-		if reservation.ScopeType == "platform_model" && reservation.Metric == "concurrent" {
-			continue
-		}
-		out = append(out, reservation)
+	out := append([]store.RateLimitReservation(nil), reservations...)
+	for index := range out {
+		out[index].ScopeKey = acceptanceScopeKey(task, out[index].ScopeKey)
 	}
 	return out
 }
 
+func acceptanceScopeKey(task store.GatewayTask, scopeKey string) string {
+	runID := strings.TrimSpace(task.AcceptanceRunID)
+	if runID == "" {
+		runID = "unbound"
+	}
+	return "acceptance:" + runID + ":" + scopeKey
+}
+
 func (s *Service) loadAsyncTaskAdmission(ctx context.Context, task store.GatewayTask) (*store.TaskAdmission, error) {
 	if !task.AsyncMode {
 		return nil, nil
@@ -265,6 +259,7 @@ func pinCandidatesToTaskAdmission(
 ) ([]store.RuntimeModelCandidate, bool) {
 	if admission == nil ||
 		(admission.Status != "waiting" && admission.Status != "admitted") ||
+		!admission.ReselectRequestedAt.IsZero() ||
 		len(candidates) < 2 {
 		return candidates, false
 	}
@@ -785,6 +780,24 @@ func (s *Service) dispatchWaitingAsyncTasks(ctx context.Context, admissions []st
 				return false, outcome.Err
 			}
 			if !outcome.Result.Admitted {
+				platformModelID := outcome.Result.Admission.PlatformModelID
+				if platformModelID == "" {
+					for _, input := range inputs {
+						if input.TaskID == outcome.TaskID {
+							platformModelID = input.PlatformModelID
+							break
+						}
+					}
+				}
+				if platformModelID != "" {
+					marked, markErr := s.store.RequestWaitingTaskAdmissionReselect(ctx, platformModelID)
+					if markErr != nil {
+						return false, markErr
+					}
+					if marked > 0 {
+						s.observeTaskAdmission("candidate_reselect_requested")
+					}
+				}
 				return true, nil
 			}
 		}