fix(acceptance): 按隔离额度执行平台降级路由

让 acceptance 候选负载读取同一 Run ID 下的并发、RPM 和 TPM 计数器，并在单个平台模型额度不足时继续尝试下一候选，生产与 canary 额度作用域保持不变。\n\n补齐本地容量控制器对宁波和香港测试 Worker 的最小 RBAC，避免初次应用生产反亲和配置时发生滚动死锁，并将突发负载对齐到完整限流窗口。\n\n验证：\n- go test ./internal/runner ./internal/store -count=1\n- go vet ./internal/runner ./internal/store\n- bash -n scripts/acceptance/local-cluster.sh scripts/acceptance/provider-burst.sh scripts/acceptance/run-local-acceptance.sh\n- shellcheck -x scripts/acceptance/local-cluster.sh scripts/acceptance/provider-burst.sh scripts/acceptance/run-local-acceptance.sh\n- kubectl apply --dry-run=client -f deploy/kubernetes/local-acceptance/capacity-controller-rbac.yaml

apps/api/internal/runner/admission.go

@@ -103,6 +103,7 @@ func (s *Service) buildTaskAdmissionPlanForCurrentBinding(
 		CacheAffinityKey:    cacheAffinityKeys.Primary,
 		CacheAffinityKeys:   cacheAffinityKeys.Lookup,
 		CacheAffinityPolicy: runnerPolicy.CacheAffinityPolicy,
+		QuotaScopePrefix:    acceptanceQuotaScopePrefix(task),
 	})
 	if err == nil {
 		candidates, err = filterCandidatesByRequestedPlatform(candidates, body)
@@ -120,6 +121,7 @@ func (s *Service) buildTaskAdmissionPlanForCurrentBinding(
 		return taskAdmissionPlan{}, err
 	}
 	candidates, _ = pinCandidatesToTaskAdmission(candidates, admission)
+	var candidateRateLimitErr error
 	for _, candidate := range candidates {
 		available, availabilityErr := s.store.RuntimeCandidateAvailable(ctx, candidate.PlatformID, candidate.PlatformModelID)
 		if availabilityErr != nil {
@@ -151,6 +153,13 @@ func (s *Service) buildTaskAdmissionPlanForCurrentBinding(
 			s.rateLimitReservations(ctx, user, candidate, body),
 		)
 		if err := s.store.CheckRateLimits(ctx, reservations); err != nil {
+			var limitErr *store.RateLimitExceededError
+			if errors.As(err, &limitErr) && limitErr.ScopeType == "platform_model" {
+				if candidateRateLimitErr == nil {
+					candidateRateLimitErr = err
+				}
+				continue
+			}
 			return taskAdmissionPlan{}, err
 		}
 		return taskAdmissionPlan{
@@ -162,6 +171,9 @@ func (s *Service) buildTaskAdmissionPlanForCurrentBinding(
 			Eligible:  true,
 		}, nil
 	}
+	if candidateRateLimitErr != nil {
+		return taskAdmissionPlan{}, candidateRateLimitErr
+	}
 	return taskAdmissionPlan{}, store.ErrNoModelCandidate
 }
 
@@ -239,6 +251,13 @@ func acceptanceScopeKey(task store.GatewayTask, scopeKey string) string {
 	return "acceptance:" + runID + ":" + scopeKey
 }
 
+func acceptanceQuotaScopePrefix(task store.GatewayTask) string {
+	if task.RunMode != "acceptance" {
+		return ""
+	}
+	return acceptanceScopeKey(task, "")
+}
+
 func (s *Service) loadAsyncTaskAdmission(ctx context.Context, task store.GatewayTask) (*store.TaskAdmission, error) {
 	if !task.AsyncMode {
 		return nil, nil

apps/api/internal/runner/admission_test.go

@@ -117,6 +117,19 @@ func TestAcceptanceInfrastructureReservationsIsolateEveryMetric(t *testing.T) {
 	}
 }
 
+func TestAcceptanceQuotaScopePrefixMatchesIsolatedCounters(t *testing.T) {
+	acceptanceTask := store.GatewayTask{RunMode: "acceptance", AcceptanceRunID: "run-1"}
+	if got := acceptanceQuotaScopePrefix(acceptanceTask); got != "acceptance:run-1:" {
+		t.Fatalf("acceptance quota scope prefix=%q", got)
+	}
+	if got := acceptanceQuotaScopePrefix(store.GatewayTask{RunMode: "acceptance_canary", AcceptanceRunID: "run-1"}); got != "" {
+		t.Fatalf("canary must use production quota scope, got %q", got)
+	}
+	if got := acceptanceQuotaScopePrefix(store.GatewayTask{RunMode: "production"}); got != "" {
+		t.Fatalf("production quota scope changed, got %q", got)
+	}
+}
+
 func TestAcceptanceAdmissionScopesLeaveProductionPolicyUnchanged(t *testing.T) {
 	input := []store.AdmissionScope{{
 		ScopeType:       "platform_model",

apps/api/internal/runner/service.go

@@ -372,6 +372,7 @@ func (s *Service) executeWithToken(ctx context.Context, task store.GatewayTask,
 		CacheAffinityKey:    cacheAffinityKeys.Primary,
 		CacheAffinityKeys:   cacheAffinityKeys.Lookup,
 		CacheAffinityPolicy: runnerPolicy.CacheAffinityPolicy,
+		QuotaScopePrefix:    acceptanceQuotaScopePrefix(task),
 	})
 	if err == nil {
 		candidates, err = filterCandidatesByRequestedPlatform(candidates, body)

apps/api/internal/store/candidates.go

@@ -17,6 +17,7 @@ type ListModelCandidatesOptions struct {
 	CacheAffinityKey    string
 	CacheAffinityKeys   []string
 	CacheAffinityPolicy map[string]any
+	QuotaScopePrefix    string
 }
 
 func (s *Store) ListModelCandidates(ctx context.Context, model string, modelType string, user *auth.User, options ...ListModelCandidatesOptions) ([]RuntimeModelCandidate, error) {
@@ -102,7 +103,7 @@ LEFT JOIN (
 	    AND released_at IS NULL
 	    AND expires_at > statement_timestamp()
 	  GROUP BY scope_key
-	) con ON con.scope_key = m.id::text
+	) con ON con.scope_key = $5::text || m.id::text
 LEFT JOIN (
 	  SELECT queued_sources.platform_model_id, COUNT(DISTINCT queued_sources.task_id) AS waiting
 	  FROM (
@@ -137,7 +138,7 @@ LEFT JOIN (
 	    AND metric = 'rpm'
 	    AND reset_at > now()
 	  ORDER BY scope_key, window_start DESC
-	) rpm ON rpm.scope_key = m.id::text
+	) rpm ON rpm.scope_key = $5::text || m.id::text
 LEFT JOIN (
 	  SELECT scope_key, SUM(used_value) AS used_value, SUM(reserved_value) AS reserved_value
 	  FROM gateway_rate_limit_counters
@@ -145,7 +146,7 @@ LEFT JOIN (
 	    AND metric LIKE 'tpm%'
 	    AND reset_at > now()
 	  GROUP BY scope_key
-	) tpm ON tpm.scope_key = m.id::text
+	) tpm ON tpm.scope_key = $5::text || m.id::text
 WHERE p.status = 'enabled'
   AND p.deleted_at IS NULL
   AND m.enabled = true
@@ -170,7 +171,7 @@ WHERE p.status = 'enabled'
 	         COALESCE(s.running_count, 0) ASC,
 	         COALESCE(s.waiting_count, 0) ASC,
 	         COALESCE(s.last_assigned_at, to_timestamp(0)) ASC,
-	         m.created_at ASC`, exactModel, modelType, listOptions.CacheAffinityKeys, cacheAffinityStaleAfterSeconds(listOptions.CacheAffinityPolicy))
+	         m.created_at ASC`, exactModel, modelType, listOptions.CacheAffinityKeys, cacheAffinityStaleAfterSeconds(listOptions.CacheAffinityPolicy), listOptions.QuotaScopePrefix)
 	if err != nil {
 		return nil, err
 	}
@@ -472,6 +473,7 @@ func normalizeListModelCandidatesOptions(modelType string, options ...ListModelC
 	}
 	out := options[0]
 	out.CacheAffinityKey = strings.TrimSpace(out.CacheAffinityKey)
+	out.QuotaScopePrefix = strings.TrimSpace(out.QuotaScopePrefix)
 	out.CacheAffinityKeys = normalizedCacheAffinityKeys(out.CacheAffinityKey, out.CacheAffinityKeys)
 	if len(out.CacheAffinityKeys) > 0 {
 		out.CacheAffinityKey = out.CacheAffinityKeys[0]

deploy/kubernetes/local-acceptance/capacity-controller-rbac.yaml

@@ -0,0 +1,17 @@
+apiVersion: rbac.authorization.k8s.io/v1
+kind: Role
+metadata:
+  name: easyai-capacity-controller
+  namespace: easyai
+rules:
+  - apiGroups: ["apps"]
+    resources: ["deployments"]
+    resourceNames: ["easyai-worker-ningbo", "easyai-worker-hongkong"]
+    verbs: ["get"]
+  - apiGroups: ["apps"]
+    resources: ["deployments/scale"]
+    resourceNames: ["easyai-worker-ningbo", "easyai-worker-hongkong"]
+    verbs: ["get", "update", "patch"]
+  - apiGroups: [""]
+    resources: ["pods"]
+    verbs: ["get", "list", "patch"]

scripts/acceptance/local-cluster.sh

@@ -599,12 +599,20 @@ render_and_apply_application() {
   chmod 0600 "$rendered"
   kubectl --context "$context" -n "$namespace" apply \
     -f "$repository_root/deploy/kubernetes/production/service-account-rbac.yaml" >/dev/null
+  kubectl --context "$context" -n "$namespace" apply \
+    -f "$manifest_root/capacity-controller-rbac.yaml" >/dev/null
   kubectl --context "$context" -n "$namespace" apply -f "$manifest_root/local-config.yaml" >/dev/null
   kubectl --context "$context" -n "$namespace" apply -f "$rendered" >/dev/null
+  kubectl --context "$context" -n "$namespace" scale \
+    deployment/easyai-worker-hongkong --replicas=0 >/dev/null
+  kubectl --context "$context" -n "$namespace" rollout status \
+    deployment/easyai-worker-hongkong --timeout=5m >/dev/null
   for workload in easyai-worker-ningbo easyai-worker-hongkong; do
     kubectl --context "$context" -n "$namespace" patch deployment "$workload" \
       --type=merge -p='{"spec":{"template":{"spec":{"affinity":null}}}}' >/dev/null
   done
+  kubectl --context "$context" -n "$namespace" scale \
+    deployment/easyai-worker-hongkong --replicas=2 >/dev/null
 
   for workload in easyai-api-ningbo easyai-worker-ningbo; do
     kubectl --context "$context" -n "$namespace" set env deployment/"$workload" \

scripts/acceptance/provider-burst.sh

@@ -543,6 +543,9 @@ provider_burst() {
   sample_resources "$resources" "$worker_sample_stop" &
   active_resource_sampler_pid=$!
   current_phase=provider_burst_load
+  while ((10#$(date -u '+%S') > 5)); do
+    sleep 1
+  done
   run_load video-provider-quota "$load_report" -requests "$requests"
   sleep 2
   touch "$worker_sample_stop"