fix(media): 统一图片结果 URL 化并限制同步 Base64

将上游 URL 直接持久化，内联媒体经对象存储后仅保留 URL 与内部定位元数据；异步轮询、任务详情和幂等重放统一使用零对象读取的 URL 投影，并增加 64KiB 响应门禁。

OpenAI 图片接口接受 url 与 b64_json，同步 Base64 限制为 20MiB 和每 Pod 2 并发；新增历史结果迁移清零门禁、结果指标和 API GOMEMLIMIT。

验证：API go test ./...、go vet、聚焦 race、pnpm openapi、pnpm lint/test/build、迁移安全检查与 docker compose config 均通过。

apps/api/internal/runner/service.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"net/http"
 	"os"
 	"regexp"
 	"strconv"
@@ -243,6 +244,13 @@ func (s *Service) observeObjectStorage(event string, provider string, bytes int,
 	}
 }
 
+func (s *Service) observeResultStorage(source string) {
+	observer, ok := s.billingMetrics.(interface{ ObserveResultStorage(string) })
+	if ok {
+		observer.ObserveResultStorage(source)
+	}
+}
+
 func (s *Service) Execute(ctx context.Context, task store.GatewayTask, user *auth.User) (Result, error) {
 	return s.execute(ctx, task, user, nil)
 }
@@ -982,7 +990,7 @@ candidatesLoop:
 						}
 						walletReservationFinalized = true
 						s.logger.Warn("task succeeded but billing requires manual review", "taskID", task.ID, "error_category", "billing_calculation_failed")
-						output, hydrateErr := s.HydrateTaskResult(ctx, task.ID, response.Result)
+						output, hydrateErr := s.taskResultForSynchronousResponse(ctx, task, response.Result)
 						if hydrateErr != nil {
 							return Result{Task: review}, hydrateErr
 						}
@@ -1075,7 +1083,7 @@ candidatesLoop:
 					// time after the task is already durably complete.
 					return Result{Task: finished, Output: response.Result}, nil
 				}
-				output, hydrateErr := s.HydrateTaskResult(ctx, task.ID, response.Result)
+				output, hydrateErr := s.taskResultForSynchronousResponse(ctx, task, response.Result)
 				if hydrateErr != nil {
 					return Result{Task: finished}, hydrateErr
 				}
@@ -1272,6 +1280,33 @@ candidatesLoop:
 	return Result{Task: failed, Output: failed.Result}, lastErr
 }
 
+func (s *Service) taskResultForSynchronousResponse(ctx context.Context, task store.GatewayTask, result map[string]any) (map[string]any, error) {
+	if taskRequestsInlineMediaResult(task.Request) {
+		if rawBytes := SynchronousInlineResultBytes(result); rawBytes > MaxSynchronousInlineResponseBytes {
+			return nil, &clients.ClientError{
+				Code:       "response_format_too_large",
+				Message:    "synchronous Base64 response exceeds the 20 MiB limit",
+				StatusCode: http.StatusRequestEntityTooLarge,
+				Retryable:  false,
+				Details: map[string]any{
+					"task_id":         task.ID,
+					"query_url":       "/api/v1/ai/result/" + task.ID,
+					"limit_bytes":     MaxSynchronousInlineResponseBytes,
+					"actual_bytes":    rawBytes,
+					"response_format": "url",
+				},
+			}
+		}
+		return s.HydrateTaskResult(ctx, task.ID, result)
+	}
+	return s.ProjectTaskResultURLs(ctx, task.ID, result)
+}
+
+func taskRequestsInlineMediaResult(request map[string]any) bool {
+	value, _ := request["response_format"].(string)
+	return strings.EqualFold(strings.TrimSpace(value), "b64_json")
+}
+
 func pricingCandidateKey(candidate store.RuntimeModelCandidate) string {
 	return firstNonEmptyString(candidate.PlatformModelID, candidate.PlatformID+":"+candidate.ModelName)
 }