[fix][evaluation] keep shared sandbox execution records in experiment workspace

backend/modules/evaluation/application/eval_openapi_app.go

@@ -1680,14 +1680,12 @@ func (e *EvalOpenAPIApplication) GetEvalTargetOutputFieldContentOApi(ctx context
 		return nil, errorx.NewByCode(errno.ResourceNotFoundCode, errorx.WithExtraMsg("eval target record not found for the given experiment result"))
 	}
 
-	// ★ 跨空间共享: 评测对象执行记录随执行落在冻结的来源空间(TargetSpaceID>0), 按 space_id 严格过滤;
-	// 用调用方空间读会查不到 record → target 输出/session 丢失。按来源空间读。
 	recordSpaceID := req.GetWorkspaceID()
 	if exptDetail, gErr := e.manager.GetDetail(ctx, req.GetExperimentID(), req.GetWorkspaceID(), entity.NewSession(ctx)); gErr == nil && exptDetail != nil && exptDetail.TargetSpaceID > 0 {
 		recordSpaceID = exptDetail.TargetSpaceID
 	}
 
-	record, err := e.targetSvc.GetRecordByID(ctx, recordSpaceID, targetRecordID)
+	record, err := service.GetExperimentTargetRecord(ctx, e.targetSvc, req.GetWorkspaceID(), recordSpaceID, targetRecordID)
 	if err != nil {
 		return nil, err
 	}

backend/modules/evaluation/domain/service/expt_manage_execution_impl.go

@@ -5,6 +5,7 @@ package service
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"strconv"
 	"strings"
@@ -1379,7 +1380,7 @@ func (e *ExptMangerImpl) listItemTurnIDs(ctx context.Context, spaceID, exptID in
 // 传入的 itemIDs 在 Terminated 分支是被终止的 item、在其它终态分支是未完成的 item。
 //
 // targetSourceSpaceID: 跨空间共享场景下评测对象的来源空间 (Expt.TargetSpaceID, 0=同消费方空间);
-// EvalTargetRecord 落库时 SpaceID = 来源空间, 用消费方 spaceID 查会漏, 沙箱清不掉。
+// Both consumer-owned records and legacy source-owned attempts must be reclaimed.
 func (e *ExptMangerImpl) terminateSandboxExecutesForCancelledItems(ctx context.Context, spaceID, targetSourceSpaceID, exptID int64, exptRunID *int64, terminatedItemIDs []int64, waitForDestroy ...bool) error {
 	if e.evalTargetService == nil || exptRunID == nil || *exptRunID <= 0 || len(terminatedItemIDs) == 0 {
 		if len(terminatedItemIDs) > 0 && len(waitForDestroy) > 0 && waitForDestroy[0] {
@@ -1390,12 +1391,16 @@ func (e *ExptMangerImpl) terminateSandboxExecutesForCancelledItems(ctx context.C
 	if lifecycle, ok := e.evalTargetService.(interface {
 		TerminateSandboxAttemptsForRunItems(context.Context, int64, int64, []int64, ...bool) error
 	}); ok {
-		if err := lifecycle.TerminateSandboxAttemptsForRunItems(ctx, resolveLoadSpaceID(spaceID, targetSourceSpaceID), *exptRunID, terminatedItemIDs, waitForDestroy...); err != nil {
-			logs.CtxWarn(ctx, "[SandboxDestroy] discover pre-dispatch attempts failed, run_id=%d, err=%v", *exptRunID, err)
-			if len(waitForDestroy) > 0 && waitForDestroy[0] {
-				return err
+		var cleanupErr error
+		for _, recordSpaceID := range ExperimentTargetRecordSpaces(spaceID, targetSourceSpaceID) {
+			if err := lifecycle.TerminateSandboxAttemptsForRunItems(ctx, recordSpaceID, *exptRunID, terminatedItemIDs, waitForDestroy...); err != nil {
+				logs.CtxWarn(ctx, "[SandboxDestroy] discover pre-dispatch attempts failed, run_id=%d, err=%v", *exptRunID, err)
+				cleanupErr = errors.Join(cleanupErr, err)
 			}
 		}
+		if cleanupErr != nil && len(waitForDestroy) > 0 && waitForDestroy[0] {
+			return cleanupErr
+		}
 	}
 	turnRunLogs, err := e.turnResultRepo.MGetItemTurnRunLogs(ctx, exptID, *exptRunID, terminatedItemIDs, spaceID)
 	if err != nil {
@@ -1416,12 +1421,9 @@ func (e *ExptMangerImpl) terminateSandboxExecutesForCancelledItems(ctx context.C
 	for id := range recordIDSet {
 		recordIDs = append(recordIDs, id)
 	}
-	// 跨空间: EvalTargetRecord.SpaceID = 来源空间, TerminateAsyncRecordsAndDestroySandbox 的
-	// List/Version/Destroy 全链路都用它，否则 space 过滤把 record 过滤空、Destroy RPC 发不出。
-	targetSpaceID := resolveLoadSpaceID(spaceID, targetSourceSpaceID)
-	return e.evalTargetService.TerminateAsyncRecordsAndDestroySandbox(
-		ctx,
-		targetSpaceID,
+
+	return terminateExperimentTargetRecords(
+		ctx, e.evalTargetService, spaceID, targetSourceSpaceID,
 		recordIDs,
 		int32(errno.AsyncEvalTargetTerminatedCode),
 		"async eval target terminated: experiment cancelled",

backend/modules/evaluation/domain/service/expt_manage_execution_impl_test.go

@@ -3892,7 +3892,7 @@ func TestExptMangerImpl_emitSandboxAgentExperimentFinished(t *testing.T) {
 
 // TestExptMangerImpl_terminateSandboxExecutesForCancelledItems_CrossSpace 覆盖实验取消 / 终态时
 // 沙箱销毁传给 TerminateAsyncRecordsAndDestroySandbox 的 spaceID 解析：
-//   - targetSourceSpaceID > 0 → 用来源空间 (跨空间共享评测对象)
+//   - targetSourceSpaceID > 0 → 消费方记录与旧来源空间记录均需回收
 //   - targetSourceSpaceID = 0 → 回退到消费方 spaceID (同空间/老数据)
 //   - 早退分支: evalTargetService = nil / exptRunID = nil or <=0 / len(itemIDs) = 0 均不发 RPC
 //   - record 全空 / MGet 出错 亦不发 RPC
@@ -3903,14 +3903,19 @@ func TestExptMangerImpl_terminateSandboxExecutesForCancelledItems_CrossSpace(t *
 	consumerSpaceID := int64(3)
 	itemIDs := []int64{10}
 
-	t.Run("targetSourceSpaceID>0 → 用来源空间 (99)", func(t *testing.T) {
+	t.Run("shared target → clean consumer and legacy source records", func(t *testing.T) {
 		ctrl := gomock.NewController(t)
 		defer ctrl.Finish()
 		mgr := newTestExptManager(ctrl)
 
 		mgr.turnResultRepo.(*repoMocks.MockIExptTurnResultRepo).
 			EXPECT().MGetItemTurnRunLogs(ctx, exptID, *exptRunID, itemIDs, consumerSpaceID).
 			Return([]*entity.ExptTurnResultRunLog{{ItemID: 10, TargetResultID: 500}}, nil)
+		mgr.evalTargetService.(*svcMocks.MockIEvalTargetService).
+			EXPECT().TerminateAsyncRecordsAndDestroySandbox(
+			gomock.Any(), consumerSpaceID, []int64{500}, int32(errno.AsyncEvalTargetTerminatedCode),
+			gomock.Any(), false,
+		).Times(1)
 		mgr.evalTargetService.(*svcMocks.MockIEvalTargetService).
 			EXPECT().TerminateAsyncRecordsAndDestroySandbox(
 			gomock.Any(), int64(99), []int64{500}, int32(errno.AsyncEvalTargetTerminatedCode),

backend/modules/evaluation/domain/service/expt_result_aggr_impl.go

@@ -161,13 +161,8 @@ func (e *ExptAggrResultServiceImpl) computeEvaluatorAggrGroup(ctx context.Contex
 func (e *ExptAggrResultServiceImpl) buildExptTargetMtrAggregatorGroup(ctx context.Context, spaceID, exptID int64) (*targetMtrAggrGroup, error) {
 	const queryInterval = time.Millisecond * 30
 
-	// ★ 跨空间共享: 评测对象执行记录随执行落来源空间(冻结 TargetSpaceID), 聚合读 record 也需按来源空间,
-	// 否则跨空间实验 target latency/token 聚合取不到 record。载入实验取冻结来源空间。
-	// 此处 fail-closed: 载入失败必须返回错误让上层重试, 不可退回调用方空间继续 ——
-	// 跨空间实验(TargetSpaceID>0)下带错空间继续会让 BatchGetRecordByIDs 全查不到,
-	// latency/token 被静默聚合成 0 并落库, 事后无法与真实 0 区分且不会自愈。
-	// 例外: 实验已被软删 / spaceID 不匹配时 GetByID 返回 ResourceNotFound —— 这是确定性结果,
-	// 重试不会好转, fail-closed 会让 MQ 永久重投且锁只能靠 TTL 过期; 此类按调用方空间降级继续。
+	// Resolve the frozen source workspace for legacy execution records.
+
 	exptDO, exptErr := e.experimentRepo.GetByID(ctx, exptID, spaceID)
 	if exptErr != nil {
 		if statusErr, ok := errorx.FromStatusError(exptErr); !ok || statusErr.Code() != errno.ResourceNotFoundCode {
@@ -213,7 +208,7 @@ func (e *ExptAggrResultServiceImpl) buildExptTargetMtrAggregatorGroup(ctx contex
 	}
 
 	for _, resIDs := range gslice.Chunk(targetResultIDs, 50) {
-		records, err := e.evalTargetSvc.BatchGetRecordByIDs(ctx, recordSpaceID, resIDs)
+		records, err := BatchGetExperimentTargetRecords(ctx, e.evalTargetSvc, spaceID, recordSpaceID, resIDs)
 		if err != nil {
 			return nil, err
 		}

backend/modules/evaluation/domain/service/expt_result_impl.go

@@ -2443,10 +2443,8 @@ func (e *ExptResultBuilder) buildTargetOutput(ctx context.Context) error {
 		targetResultIDs = append(targetResultIDs, turnResult.TargetResultID)
 		targetResultID2turnResultID[turnResult.TargetResultID] = turnResult.ID
 	}
-	// ★ 跨空间共享: 评测对象执行记录(eval_target_record)随执行落在来源空间(冻结 TargetSpaceID),
-	// 按 space_id 严格过滤; 用调用方空间读会查不到 → target 输出(含 Ext[sandbox_session_id])整体丢失,
-	// 前端 case 旁不显示 aiosession。按来源空间读 (0=同调用方空间)。
-	targetRecords, err := e.evalTargetService.BatchGetRecordByIDs(ctx, resolveLoadSpaceID(e.SpaceID, e.exptDO.TargetSpaceID), targetResultIDs)
+
+	targetRecords, err := BatchGetExperimentTargetRecords(ctx, e.evalTargetService, e.SpaceID, e.exptDO.TargetSpaceID, targetResultIDs)
 	if err != nil {
 		return err
 	}

backend/modules/evaluation/domain/service/expt_retry_failure_generation_test.go

@@ -83,6 +83,9 @@ func runRetryGenerationPreEval(t *testing.T, f *retryGenerationFixture) ([]*enti
 	turnRepo.EXPECT().GetItemTurnResults(gomock.Any(), int64(1), int64(10), int64(3)).Return(f.turns, nil).Times(repeats)
 	turnRepo.EXPECT().BatchGetTurnEvaluatorResultRef(gomock.Any(), int64(3), turnIDs).Return(refs, nil).Times(2 * repeats)
 	idgen.EXPECT().GenMultiIDs(gomock.Any(), len(f.turns)).Return(turnIDs, nil).Times(repeats)
+	if f.targetSpace != 3 {
+		targetSvc.EXPECT().GetRecordByID(gomock.Any(), int64(3), int64(201)).Return(nil, nil).AnyTimes()
+	}
 	targetSvc.EXPECT().GetRecordByID(gomock.Any(), f.targetSpace, int64(201)).DoAndReturn(func(ctx context.Context, _, _ int64) (*entity.EvalTargetRecord, error) {
 		assert.True(t, contexts.CtxWriteDB(ctx))
 		return &entity.EvalTargetRecord{ID: 201, ExperimentRunID: 101, Status: gptr.Of(entity.EvalTargetRunStatusSuccess)}, nil
@@ -496,7 +499,7 @@ func TestRetryFailure_SourceLogMembershipPreservesNewFormatIdentity(t *testing.T
 			}, nil)
 			turnRepo.EXPECT().MGetItemTurnRunLogs(gomock.Any(), int64(1), int64(101), []int64{10}, int64(3)).Return([]*entity.ExptTurnResultRunLog{nil, origin}, nil)
 			sources := &retryEvaluatorSourceLogs{repo: turnRepo, event: &entity.ExptItemEvalEvent{ExptID: 1, EvalSetItemID: 10, SpaceID: 3}, logsByRun: make(map[int64]map[int64]*entity.ExptTurnResultRunLog)}
-			targetID, got, selectErr := failRetrySelectTurnRunLogRefs(context.Background(), 3, true, tr, targetSvc, recordSvc, []*entity.ExptTurnEvaluatorResultRef{{ExptTurnResultID: tr.ID, EvaluatorVersionID: 401, Alias: "judge-a", EvaluatorResultID: 301}, {ExptTurnResultID: tr.ID, InlineKey: "inline-a", SourceType: int32(entity.EvaluatorRecordSourceTypeInline), EvaluatorResultID: 302}}, sources)
+			targetID, got, selectErr := failRetrySelectTurnRunLogRefs(context.Background(), 3, 0, true, tr, targetSvc, recordSvc, []*entity.ExptTurnEvaluatorResultRef{{ExptTurnResultID: tr.ID, EvaluatorVersionID: 401, Alias: "judge-a", EvaluatorResultID: 301}, {ExptTurnResultID: tr.ID, InlineKey: "inline-a", SourceType: int32(entity.EvaluatorRecordSourceTypeInline), EvaluatorResultID: 302}}, sources)
 			require.NoError(t, selectErr)
 			assert.Equal(t, int64(201), targetID)
 			if !tt.registered && !tt.inline {

backend/modules/evaluation/domain/service/expt_retry_failure_stage_resume_test.go

@@ -91,7 +91,7 @@ func TestFailRetrySelectTurnRunLogRefs_TargetRequirementControlsEvaluatorReuse(t
 			}
 
 			targetID, evaluatorResults, selectErr := failRetrySelectTurnRunLogRefs(
-				context.Background(), 1, tt.targetRequired,
+				context.Background(), 1, 0, tt.targetRequired,
 				&entity.ExptTurnResult{
 					EvaluatorResults: &entity.EvaluatorResults{
 						EvalVerIDToResID: map[int64]int64{10: evaluatorRecordID},
@@ -425,6 +425,8 @@ func TestExptFailRetryExec_BuildPagePlan_MultiSetGroupsTargetReadsBySourceSpace(
 		{ItemID: 12, ItemConfig: &entity.ExptItemConfig{TargetSourceSpaceID: rowTargetSpace}},
 	}, nil)
 	success := entity.EvalTargetRunStatusSuccess
+	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), int64(3), []int64{110}).Return(nil, nil)
+	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), int64(3), []int64{100, 120}).Return(nil, nil)
 	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), topTargetSpace, []int64{110}).Return([]*entity.EvalTargetRecord{{ID: 110, Status: &success}}, nil)
 	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), rowTargetSpace, []int64{100, 120}).Return([]*entity.EvalTargetRecord{
 		{ID: 100, Status: &success},
@@ -483,6 +485,7 @@ func TestExptFailRetryExec_BuildPagePlan_MultiSetMissingItemRefFallsBackToExperi
 		ItemID: itemID, ItemConfig: nil,
 	}}, nil)
 	success := entity.EvalTargetRunStatusSuccess
+	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), int64(3), []int64{30}).Return(nil, nil)
 	targetSvc.EXPECT().BatchGetRecordByIDs(gomock.Any(), int64(99), []int64{30}).Return([]*entity.EvalTargetRecord{{ID: 30, Status: &success}}, nil)
 	idgen.EXPECT().GenMultiIDs(gomock.Any(), 1).Return([]int64{900}, nil)
 

backend/modules/evaluation/domain/service/expt_run_item_event_impl.go

@@ -1298,7 +1298,7 @@ func (e *ExptRecordEvalModeSubmit) PostEval(ctx context.Context, eiec *entity.Ex
 // failRetrySelectTurnRunLogRefs keeps successful stages that belong to the reused target.
 func failRetrySelectTurnRunLogRefs(
 	ctx context.Context,
-	spaceID int64,
+	spaceID, sourceSpaceID int64,
 	targetRequired bool,
 	tr *entity.ExptTurnResult,
 	evalTarget IEvalTargetService,
@@ -1322,7 +1322,7 @@ func failRetrySelectTurnRunLogRefs(
 	if evalTarget == nil {
 		return 0, nil, errors.New("eval target service is nil")
 	}
-	targetRec, err := evalTarget.GetRecordByID(ctx, spaceID, tr.TargetResultID)
+	targetRec, err := GetExperimentTargetRecord(ctx, evalTarget, spaceID, sourceSpaceID, tr.TargetResultID)
 	if err != nil {
 		return 0, nil, err
 	}
@@ -1571,9 +1571,7 @@ func (e *ExptRecordEvalModeFailRetry) PreEval(ctx context.Context, eiec *entity.
 		runLog.Status = entity.TurnRunState_Processing
 		runLog.ExptRunID = eiec.Event.ExptRunID
 		runLog.ErrMsg = ""
-		// 跨空间共享: Target 记录随执行落来源空间(冻结 TargetSpaceID), 失败重试选引用时须按来源空间读;
-		// 用调用方空间读会得 nil → 误判 Target 非 Success → 清零 target_result_id 触发无谓重跑 Target。
-		targetID, evalIDs, err := failRetrySelectTurnRunLogRefs(snapshotCtx, resolveLoadSpaceID(eiec.Event.SpaceID, eiec.TargetSourceSpaceID()), !shouldSkipTargetNode(eiec.Expt), tr, e.evalTargetService, e.evaluatorRecordSvc, refsByTurnResultID[tr.ID], sourceLogs)
+		targetID, evalIDs, err := failRetrySelectTurnRunLogRefs(snapshotCtx, eiec.Event.SpaceID, eiec.TargetSourceSpaceID(), !shouldSkipTargetNode(eiec.Expt), tr, e.evalTargetService, e.evaluatorRecordSvc, refsByTurnResultID[tr.ID], sourceLogs)
 		if err != nil {
 			return fmt.Errorf("%w: %w", errRetryStartDependencyFailure, err)
 		}

backend/modules/evaluation/domain/service/expt_run_item_event_impl_test.go

@@ -1678,7 +1678,7 @@ func Test_failRetrySelectTurnRunLogRefs(t *testing.T) {
 					}}, nil)
 			}
 			gotTargetID, gotEvalResults, selectErr := failRetrySelectTurnRunLogRefs(
-				context.Background(), tt.spaceID, targetRequired, tt.tr, evalTarget, evalRecord, refs, sourceLogs,
+				context.Background(), tt.spaceID, 0, targetRequired, tt.tr, evalTarget, evalRecord, refs, sourceLogs,
 			)
 			if tt.name == "TargetResultID > 0, evalTarget is nil -> dependency error" || tt.name == "TargetResultID > 0, GetRecordByID returns error -> returns 0, nil" {
 				require.Error(t, selectErr)

backend/modules/evaluation/domain/service/expt_run_item_impl.go

@@ -443,9 +443,7 @@ func (e *ExptItemEvalCtxExecutor) buildExptTurnEvalCtx(ctx context.Context, turn
 	}
 
 	if tid := existTurnRunResult.TargetResultID; tid > 0 {
-		// ★ 跨空间共享: 评测对象执行记录随执行落来源空间(冻结 TargetSpaceID), 按来源空间读;
-		// 用调用方空间读会得 nil → 异步回调 validateEvalTargetCtx 报 "target result must not be nil"。
-		targetRecord, err := e.evalTargetService.GetRecordByID(recordCtx, resolveLoadSpaceID(spaceID, eiec.TargetSourceSpaceID()), tid)
+		targetRecord, err := GetExperimentTargetRecord(recordCtx, e.evalTargetService, spaceID, eiec.TargetSourceSpaceID(), tid)
 		if err != nil {
 			return nil, err
 		}

backend/modules/evaluation/domain/service/expt_run_scheduler_event_impl.go

@@ -935,13 +935,11 @@ func (e *ExptSchedulerImpl) sweepTerminatedSandboxItems(ctx context.Context, eve
 		return items, nil, nil
 	}
 
-	// 跨空间共享: EvalTargetRecord.SpaceID 是来源空间, 用消费方 event.SpaceID 查会漏。
 	var targetSourceSpaceID int64
 	if expt != nil {
 		targetSourceSpaceID = expt.TargetSpaceID
 	}
-	targetSpaceID := resolveLoadSpaceID(event.SpaceID, targetSourceSpaceID)
-	terminatedRecordIDs, statusMap := e.evalTargetService.CheckSandboxTerminated(ctx, targetSpaceID, recordIDs)
+	terminatedRecordIDs, statusMap := checkExperimentSandboxTerminated(ctx, e.evalTargetService, event.SpaceID, targetSourceSpaceID, recordIDs)
 	if len(terminatedRecordIDs) == 0 {
 		return items, nil, nil
 	}
@@ -983,10 +981,8 @@ func (e *ExptSchedulerImpl) sweepTerminatedSandboxItems(ctx context.Context, eve
 		event.ExptID, event.ExptRunID, terminatedItemIDs, firstStatus)
 
 	// 沙箱已经是终态，Destroy 不需要再发 EndCmd，zombieTimeout=false。
-	// targetSpaceID 复用上面 CheckSandboxTerminated 用的来源空间, 保持一致。
-	_ = e.evalTargetService.TerminateAsyncRecordsAndDestroySandbox(
-		ctx,
-		targetSpaceID,
+	_ = terminateExperimentTargetRecords(
+		ctx, e.evalTargetService, event.SpaceID, targetSourceSpaceID,
 		terminatedRecordIDs,
 		int32(errno.SandboxTerminatedBeforeReportCode),
 		fmt.Sprintf("sandbox execute reached terminal state (%s) before result was reported", firstStatus),
@@ -1110,17 +1106,12 @@ func (e *ExptSchedulerImpl) terminateZombieEvalTargetRecords(ctx context.Context
 		return nil
 	}
 
-	// 跨空间共享: EvalTargetRecord 落库时 SpaceID = 评测对象来源空间 (resolveLoadSpaceID 结果),
-	// 若这里仍用 event.SpaceID (消费方) 去查, DAO 的 SpaceID.Eq 过滤会返回空 → Destroy RPC 根本发不出,
-	// 导致沙箱在来源空间空跑到 session TTL 兜底才回收。与 expt_run_item_turn_impl.go:130 保持同一套解析。
 	var targetSourceSpaceID int64
 	if expt != nil {
 		targetSourceSpaceID = expt.TargetSpaceID
 	}
-	targetSpaceID := resolveLoadSpaceID(event.SpaceID, targetSourceSpaceID)
-	_ = e.evalTargetService.TerminateAsyncRecordsAndDestroySandbox(
-		ctx,
-		targetSpaceID,
+	_ = terminateExperimentTargetRecords(
+		ctx, e.evalTargetService, event.SpaceID, targetSourceSpaceID,
 		recordIDs,
 		int32(errno.AsyncEvalTargetZombieTimeoutCode),
 		"async eval target terminated: experiment item exceeded zombie timeout",