fix(queue): 防止过期执行占位阻塞与重复提交

队列恢复后，仍在运行标记中的 River job 可能保留 waiting admission，形成全局 FIFO 队头阻塞；同时旧执行在租约被接管后仍可能创建 attempt 或切换到 submitting。

新增 stale admission 自动让出机制，并用当前 execution token 对 attempt 创建和上游提交状态切换做 fencing。任务、River job 和结算状态均不在让出流程中改写。

验证：go vet ./...；env -u AI_GATEWAY_TEST_DATABASE_URL go test ./... -count=1；真实 PostgreSQL 集成测试覆盖 stale admission 让出与旧 token 提交拦截。

apps/api/internal/runner/queue_worker.go

@@ -589,6 +589,20 @@ func (s *Service) recoverOrphanedAsyncRiverJobs(ctx context.Context) {
 				"cleanedTaskAdmissions", runtimeRecovery.CleanedTaskAdmissions,
 			)
 		}
+		yieldedAdmissions, err := s.store.YieldStaleAsyncTaskAdmissions(
+			ctx,
+			orphanedRiverJobWorkerStaleAfter,
+			orphanedRiverJobRecoveryBatchSize,
+		)
+		if err != nil {
+			if ctx.Err() == nil {
+				s.logger.Warn("yield stale async task admissions failed", "error", err)
+			}
+			return
+		}
+		if yieldedAdmissions > 0 {
+			s.logger.Warn("stale async task admissions yielded", "count", yieldedAdmissions)
+		}
 		recovered, err := s.store.RecoverOrphanedAsyncRiverJobs(
 			ctx,
 			orphanedRiverJobWorkerStaleAfter,