test(acceptance): 增加视频容量阶梯场景

新增 30 至 45 秒模拟上游等待的视频容量场景，使单 Worker 自适应控制在每级突发负载中获得连续占满窗口，同时保持提交、轮询、结果处理和回调全链路。\n\n未修改生产 Worker 和厂商限流逻辑。已执行完整 Go 测试、go vet、gofmt、bash -n、ShellCheck 和生产验收脚本测试。

apps/api/cmd/acceptance-load/main.go

@@ -279,7 +279,7 @@ func parseOptions() (options, error) {
 		return options{}, errors.New("requests must be between 0 and 10000")
 	}
 	switch opts.profile {
-	case "simulated-smoke", "simulated-all", "gemini-baseline", "gemini-multi-image", "gemini-large", "gemini-peak", "video-throughput", "video-recovery",
+	case "simulated-smoke", "simulated-all", "gemini-baseline", "gemini-multi-image", "gemini-large", "gemini-peak", "video-throughput", "video-capacity", "video-recovery",
 		"mixed-soak", "mixed-overload":
 		if opts.emulatorURL == "" {
 			return options{}, errors.New("AI_GATEWAY_ACCEPTANCE_EMULATOR_URL is required for simulated profiles")
@@ -365,6 +365,12 @@ func run(ctx context.Context, opts options) ([]phaseResult, error) {
 					requests = opts.requestCount
 				}
 				result = runVideo(ctx, client, opts, name, opts.shardRequestCount(requests), false, opts.emulatorFixtureURLs(), false, opts.shardIndex, opts.shardCount)
+			case "video-capacity":
+				requests := 1200
+				if opts.requestCount > 0 {
+					requests = opts.requestCount
+				}
+				result = runVideo(ctx, client, opts, name, opts.shardRequestCount(requests), false, opts.emulatorFixtureURLs(), false, opts.shardIndex, opts.shardCount)
 			case "video-recovery":
 				result = runVideo(ctx, client, opts, name, opts.shardRequestCount(96), true, opts.emulatorFixtureURLs(), false, opts.shardIndex, opts.shardCount)
 			case "mixed-soak", "mixed-overload":
@@ -659,6 +665,9 @@ func runVideo(
 			if longRun {
 				prompt += " acceptance-long-recovery"
 			}
+			if name == "video-capacity" {
+				prompt += " acceptance-capacity-ladder"
+			}
 			if !realUpstream && logicalIndex%4 == 0 {
 				prompt += " acceptance-force-conversion"
 			}

apps/api/internal/acceptanceemulator/server.go

@@ -674,9 +674,24 @@ func videoDelay(body map[string]any, longRun bool) time.Duration {
 	if longRun {
 		return 2*time.Minute + time.Duration(seed%61)*time.Second
 	}
+	if videoPromptContains(body, "acceptance-capacity-ladder") {
+		return 30*time.Second + time.Duration(seed%16)*time.Second
+	}
 	return 5*time.Second + time.Duration(seed%11)*time.Second
 }
 
+func videoPromptContains(body map[string]any, marker string) bool {
+	content, _ := body["content"].([]any)
+	for _, raw := range content {
+		item, _ := raw.(map[string]any)
+		if strings.TrimSpace(stringValue(item["type"])) == "text" &&
+			strings.Contains(strings.ToLower(stringValue(item["text"])), marker) {
+			return true
+		}
+	}
+	return false
+}
+
 func requestOrigin(r *http.Request) string {
 	scheme := strings.TrimSpace(r.Header.Get("X-Forwarded-Proto"))
 	if scheme == "" {

apps/api/internal/acceptanceemulator/server_test.go

@@ -228,6 +228,20 @@ func TestScaledDelayStaysOnWholeSecondsWithinProfile(t *testing.T) {
 	}
 }
 
+func TestVideoCapacityDelay(t *testing.T) {
+	body := map[string]any{
+		"seed": json.Number("7"),
+		"content": []any{map[string]any{
+			"type": "text",
+			"text": "capacity acceptance-capacity-ladder",
+		}},
+	}
+	delay := videoDelay(body, false)
+	if delay < 30*time.Second || delay > 45*time.Second {
+		t.Fatalf("video capacity delay=%s, want 30s..45s", delay)
+	}
+}
+
 func TestVolcesProtocolAcceptsThreeSixAndNineReferenceImages(t *testing.T) {
 	server := New(Config{Wait: func(context.Context, time.Duration) error { return nil }})
 	httpServer := httptest.NewServer(server.Handler())

scripts/cluster/run-production-acceptance.sh

@@ -3819,7 +3819,7 @@ run_single_node_acceptance() {
     apply_single_node_capacity "$slots" || break
     requests=$((slots * 3))
     (( requests < 32 )) && requests=32
-    if ! run_single_node_profile image-video video-throughput "$slots" "$requests"; then
+    if ! run_single_node_profile image-video video-capacity "$slots" "$requests"; then
       break
     fi
     single_node_stable_slots=$slots