fix(openai): 스트림 정지 종료를 복구한다

semantic filter 비활성 경로도 endpoint codec과 terminal sink를 유지해야 논리적 finish 뒤 정지가 열린 응답으로 남지 않는다.
2026-08-13 11:41:23 +09:00 · 2026-08-13 11:41:23 +09:00 · 2fcc1093c7
commit 2fcc1093c7
parent eaf8fbaf21
7 changed files with 521 additions and 54 deletions
--- a/apps/edge/internal/openai/responses_stream_gate.go
+++ b/apps/edge/internal/openai/responses_stream_gate.go
@ -1112,7 +1112,6 @@ func (s *Server) buildOpenAIResponsesStreamGateRuntime(dc *responsesDispatchCont
 // than retaining caller-derived Run/PrepareRun state from a generic tunnel
 // runtime.
 func (s *Server) buildOpenAIResponsesStreamGateRuntimeFromAttempt(dc *responsesDispatchContext, initial openAIAttemptTransport, dispatch edgeservice.RunDispatch, closeInitial func(), sink openAIStreamGateSink, registry streamgate.FilterRegistrySnapshot, stallStates ...*openAIStallRecoveryState) (*streamgate.RequestRuntime, *openAIStreamGateUsageHolder, error) {
-	semanticEnabled := s.streamGateSemanticEnabled()
 	var stallState *openAIStallRecoveryState
 	if len(stallStates) > 0 {
 		stallState = stallStates[0]
@ -1177,12 +1176,10 @@ func (s *Server) buildOpenAIResponsesStreamGateRuntimeFromAttempt(dc *responsesD
 			}
 			assembler := &providerChatAssembler{streaming: attemptDC.req.Stream}
 			rewriter := newProviderModelRewriter(attemptDC.req.Stream, "")
-			var tunnelSource *openAITunnelEventSource
-			if semanticEnabled {
-				tunnelSource = newOpenAITunnelEndpointEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, openAIRebuildEndpointResponses, codecState)
-			} else {
-				tunnelSource = newOpenAITunnelEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, codecState)
-			}
+			tunnelSource := newOpenAITunnelEndpointEventSource(
+				transport.tunnel.Stream(), transport.tunnel.WaitTimeout(),
+				rewriter, assembler, openAIRebuildEndpointResponses, codecState,
+			)
 			src = &openAIStreamGateUsageTrackingTunnelSource{openAITunnelEventSource: tunnelSource, usage: usage, attempt: transport.usage}
 		default:
 			return nil, fmt.Errorf("openai responses unsupported attempt path %q", transport.path)
@ -1278,12 +1275,7 @@ func (s *Server) runOpenAIResponsesStreamGateAttempt(w http.ResponseWriter, dc *
 	var sink openAIStreamGateSink = normalized
 	if dc.poolDispatch != nil {
 		if dc.responsesRequestContext.envelope.Stream {
-			if s.streamGateSemanticEnabled() {
-				sink = newOpenAIResponsesPoolReleaseSink(w, holder, selector)
-			} else {
-				flusher, _ := w.(http.Flusher)
-				sink = newOpenAICompositeReleaseSink(selector, normalized, newOpenAITunnelReleaseSink(w, flusher))
-			}
+			sink = newOpenAIResponsesPoolReleaseSink(w, holder, selector)
 		} else {
 			tunnel := newOpenAIBufferedTunnelReleaseSink(w, nil, "")
 			sink = newOpenAICompositeReleaseSink(selector, normalized, tunnel)
--- a/apps/edge/internal/openai/stream_gate_pipeline_test.go
+++ b/apps/edge/internal/openai/stream_gate_pipeline_test.go
@ -387,7 +387,7 @@ func TestOpenAITunnelCodecTerminalWire(t *testing.T) {
 			endpoint: openAIRebuildEndpointChat,
 			frames: [][]byte{
 				[]byte("data: {\"choices\":[{\"delta\":{\"content\":\"answer\"}}]}\n\n"),
-				[]byte("data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n"),
+				[]byte("data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":2,\"completion_tokens\":1}}\n\n"),
 				[]byte("data: [DONE]\n\n"),
 			},
 		},
@ -413,12 +413,19 @@ func TestOpenAITunnelCodecTerminalWire(t *testing.T) {
 			state := &openAITunnelCodecState{}
 			codec := newOpenAITunnelEndpointCodec(tc.endpoint, state)
 			var events []streamgate.NormalizedEvent
-			for _, frame := range tc.frames {
+			for index, frame := range tc.frames {
 				decoded, err := codec.decode(frame, false)
 				if err != nil {
 					t.Fatalf("decode: %v", err)
 				}
 				events = append(events, decoded...)
+				if index < len(tc.frames)-1 {
+					for _, event := range events {
+						if event.Kind() == streamgate.EventKindTerminal {
+							t.Fatalf("protocol finish created a Core terminal before [DONE]/END: frame=%d events=%v", index, eventKinds(events))
+						}
+					}
+				}
 			}
 			decoded, err := codec.finishTransport(nil, false)
 			if err != nil {
@ -810,6 +817,85 @@ func TestOpenAITunnelHTTPErrorLifecycle(t *testing.T) {
 	}
 }

+// TestOpenAIDirectResponsesSemanticDisabledStallTerminal drives the generic
+// direct Responses tunnel runtime (not the provider-pool Responses wrapper).
+// Once raw Responses SSE is visible, a typed stall must discard a staged
+// response.completed and append exactly one native error plus [DONE].
+func TestOpenAIDirectResponsesSemanticDisabledStallTerminal(t *testing.T) {
+	srv := NewServer(config.EdgeOpenAIConf{TimeoutSec: 5}, &providerFakeRunService{}, nil)
+	rawBody := []byte(`{"model":"client-model","stream":true,"input":"hi"}`)
+	requestCtx := newTestRequestContext(t, routeDispatch{Adapter: "openai-compat", Target: "served-model", TimeoutSec: 5}, rawBody)
+	req := openAITunnelStreamGateRequest{
+		route:   routeDispatch{Adapter: "openai-compat", Target: "served-model", TimeoutSec: 5},
+		ingress: requestCtx.ingress, endpoint: openAIRebuildEndpointResponses,
+		method: http.MethodPost, path: "/v1/responses", operation: string(config.OperationResponses),
+		stream: true, modelGroupKey: "client-model", semanticSet: true, semanticEnabled: false,
+		authorize:   func(context.Context) (map[string]string, error) { return nil, nil },
+		rewriteBody: func(body []byte, _ string) ([]byte, error) { return body, nil },
+	}
+	frames := bufferedTunnelFrames(
+		&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_RESPONSE_START, StatusCode: http.StatusOK, Headers: map[string]string{"Content-Type": "text/event-stream"}},
+		&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_BODY, Body: []byte(
+			"data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp-direct\",\"object\":\"response\",\"status\":\"in_progress\"},\"sequence_number\":1}\n\n" +
+				"data: {\"type\":\"response.output_text.delta\",\"item_id\":\"msg-direct\",\"output_index\":0,\"content_index\":0,\"delta\":\"direct-prefix\",\"sequence_number\":2}\n\n" +
+				"data: {\"type\":\"response.completed\",\"response\":{\"id\":\"raw-provider-terminal-must-stay-hidden\",\"object\":\"response\",\"status\":\"completed\"},\"sequence_number\":3}\n\n",
+		)},
+		&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_ERROR, Failure: confirmedStallFailure("available")},
+	)
+	closeCount := 0
+	handle := &countingDirectTunnelHandle{
+		fakeTunnelHandle: fakeTunnelHandle{dispatch: edgeservice.RunDispatch{
+			RunID: "direct-responses-stall", ModelGroupKey: "client-model", Adapter: "openai-compat",
+			Target: "served-model", ProviderID: "provider-a", ExecutionPath: string(edgeservice.ProviderPoolPathTunnel),
+		}, frames: frames},
+		closed: &closeCount,
+	}
+	fctx, err := srv.openAITunnelOutputFilterContext(req)
+	if err != nil {
+		t.Fatalf("openAITunnelOutputFilterContext: %v", err)
+	}
+	stallState, registration, err := openAIStallRecoveryRegistration(fctx)
+	if err != nil {
+		t.Fatalf("openAIStallRecoveryRegistration: %v", err)
+	}
+	registry, err := openAIStreamGateRegistrySnapshotFor(srv.streamGateConfig(), fctx, registration)
+	if err != nil {
+		t.Fatalf("openAIStreamGateRegistrySnapshotFor: %v", err)
+	}
+	w := newRecordingResponseWriter()
+	sink := newOpenAIResponsesPoolReleaseSink(w, &openAIResponsesResultHolder{}, newOpenAIStreamGateCodecSelector(openAIStreamGateCodecTunnel))
+	runtime, _, err := srv.buildOpenAITunnelStreamGateRuntime(req, handle, sink, registry, stallState)
+	if err != nil {
+		t.Fatalf("buildOpenAITunnelStreamGateRuntime: %v", err)
+	}
+	runErr := runtime.Run(t.Context())
+	committed, success := sink.terminalStatus()
+	if closeErr := runtime.CloseRequestResources(t.Context(), false); closeErr != nil {
+		t.Fatalf("CloseRequestResources: %v", closeErr)
+	}
+	body := w.body.String()
+	if runErr != nil || !committed || success || w.code != http.StatusOK || closeCount != 1 {
+		t.Fatalf("runtime=(err=%v committed=%v success=%v status=%d closes=%d body=%q)", runErr, committed, success, w.code, closeCount, body)
+	}
+	if !strings.Contains(body, "direct-prefix") || strings.Contains(body, "raw-provider-terminal-must-stay-hidden") || strings.Contains(body, "provider body") || strings.Contains(body, "raw provider metadata") {
+		t.Fatalf("direct Responses terminal leaked or lost wire: %q", body)
+	}
+	if strings.Count(body, `"type":"error"`) != 1 || strings.Count(body, "data: [DONE]") != 1 || strings.Contains(body, `"type":"response.completed"`) {
+		t.Fatalf("direct Responses terminal count mismatch: %q", body)
+	}
+}
+
+type countingDirectTunnelHandle struct {
+	fakeTunnelHandle
+	closed *int
+}
+
+func (h *countingDirectTunnelHandle) Close() {
+	if h.closed != nil {
+		*h.closed = *h.closed + 1
+	}
+}
+
 // TestOpenAITunnelHTTPErrorRawPassthroughRuntime closes the production
 // source -> Core -> release-sink gap for unmatched provider errors. A non-2xx
 // body stays opaque even when it looks like endpoint success/tool evidence,
--- a/apps/edge/internal/openai/stream_gate_runtime.go
+++ b/apps/edge/internal/openai/stream_gate_runtime.go
@ -647,6 +647,48 @@ func newOpenAITunnelEndpointEventSource(stream edgeservice.ProviderTunnelStream,
 	return source
 }

+// decodeEndpointTunnelBody keeps endpoint parsing active for known Chat and
+// Responses frames while preserving the legacy behavior of opaque SSE data.
+// The latter remains a release event per complete provider frame; recognized
+// endpoint finish wire stays staged until [DONE], END, or a typed error owns
+// the terminal.
+func decodeEndpointTunnelBody(codec *openAITunnelEndpointCodec, body []byte) ([]streamgate.NormalizedEvent, error) {
+	if codec == nil || codec.terminal {
+		return nil, nil
+	}
+	codec.pending = append(codec.pending, body...)
+	var out []streamgate.NormalizedEvent
+	for {
+		frame, rest, ok := takeOpenAISSEFrame(codec.pending)
+		if !ok {
+			break
+		}
+		codec.pending = rest
+		data := openAISSEData(frame)
+		if strings.TrimSpace(data) != "" && strings.TrimSpace(data) != "[DONE]" && !json.Valid([]byte(data)) {
+			payload := append(append([]byte(nil), codec.stagedWire...), frame...)
+			codec.stagedWire = nil
+			codec.state.pushRelease(payload)
+			event, err := streamgate.NewTextDeltaEvent(streamGateChannelDefault, string(frame), time.Now())
+			if err != nil {
+				return nil, err
+			}
+			out = append(out, event)
+			continue
+		}
+		events, err := codec.decodeFrame(frame)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, events...)
+		if codec.terminal {
+			codec.pending = nil
+			break
+		}
+	}
+	return out, nil
+}
+
 func (s *openAITunnelEventSource) NextEvent(ctx context.Context) (streamgate.NormalizedEvent, error) {
 	s.mu.Lock()
 	if len(s.pending) > 0 {
@ -756,7 +798,7 @@ func (s *openAITunnelEventSource) translateFrame(frame *iop.ProviderTunnelFrame)
 			return events, nil
 		}
 		if s.codec != nil {
-			decoded, err := s.codec.decode(rewritten, false)
+			decoded, err := decodeEndpointTunnelBody(s.codec, rewritten)
 			if err != nil {
 				return nil, err
 			}
@ -775,12 +817,18 @@ func (s *openAITunnelEventSource) translateFrame(frame *iop.ProviderTunnelFrame)
 		return nil, nil

 	case iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_ERROR:
-		if s.codec == nil && (frame.GetFailure() == nil || frame.GetFailure().GetCode() != openAIStallFailureCode) {
+		if frame.GetFailure() == nil || frame.GetFailure().GetCode() != openAIStallFailureCode {
 			message := frame.GetError()
 			if message == "" {
 				message = "provider tunnel failed"
 			}
 			s.compatState.setCompatibilityProviderTerminal(message)
+			// The generic provider-error compatibility path owns its legacy
+			// pre/post-start behavior and must not acquire an endpoint-authored
+			// SSE terminal merely because endpoint parsing is always enabled.
+			s.compatState.mu.Lock()
+			s.compatState.endpoint = ""
+			s.compatState.mu.Unlock()
 		}
 		ev, err := newOpenAIProviderErrorEventFromFailure(frame.GetFailure(), streamGateErrorTunnelFailed)
 		if err != nil {
@ -1238,12 +1286,10 @@ func (s *Server) newOpenAIChatAttemptEventSourceFactory(
 			rewriter := newProviderModelRewriter(dc.req.Stream, dc.req.Model)
 			state := openAITunnelCodecStateForSink(cfg.sink)
 			state.reset()
-			var tunnelSrc *openAITunnelEventSource
-			if cfg.semanticEnabled {
-				tunnelSrc = newOpenAITunnelEndpointEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, openAIRebuildEndpointChat, state)
-			} else {
-				tunnelSrc = newOpenAITunnelEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, state)
-			}
+			tunnelSrc := newOpenAITunnelEndpointEventSource(
+				transport.tunnel.Stream(), transport.tunnel.WaitTimeout(),
+				rewriter, assembler, openAIRebuildEndpointChat, state,
+			)
 			dispatch := transport.tunnel.Dispatch()
 			tunnelSrc.onTerminal = func(obs *providerAssembledObservation, bodyBytes int) {
 				s.logger.Info("openai chat completion passthrough closed",
@ -1806,13 +1852,14 @@ func (s *Server) buildOpenAITunnelStreamGateRuntime(
 	if len(stallStates) > 0 {
 		stallState = stallStates[0]
 	}
-	semanticEnabled := req.semanticEnabled
-	if !req.semanticSet {
-		// Direct runtime fixtures predate the product-level semantic switch.
-		// Product callers always set semanticSet explicitly.
-		semanticEnabled = true
-	}
 	usage := &openAIStreamGateUsageHolder{}
+	var responseSink *openAIResponsesPoolReleaseSink
+	if req.endpoint == openAIRebuildEndpointResponses {
+		responseSink, _ = sink.(*openAIResponsesPoolReleaseSink)
+		if responseSink != nil {
+			responseSink.setUsageHolder(usage)
+		}
+	}

 	recoverySource := newOpenAIRecoverySourceStore(req.ingress)
 	rebuilder, err := newOpenAIRequestRebuilder(req.ingress, req.endpoint, recoverySource, s.openAIResumeContextWindowTokens(req.modelGroupKey))
@ -1832,13 +1879,15 @@ func (s *Server) buildOpenAITunnelStreamGateRuntime(
 		assembler := &providerChatAssembler{streaming: req.stream}
 		rewriter := newProviderModelRewriter(req.stream, req.requestModel)
 		state := openAITunnelCodecStateForSink(sink)
-		state.reset()
-		var src *openAITunnelEventSource
-		if semanticEnabled {
-			src = newOpenAITunnelEndpointEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, req.endpoint, state)
-		} else {
-			src = newOpenAITunnelEventSource(transport.tunnel.Stream(), transport.tunnel.WaitTimeout(), rewriter, assembler, state)
+		if responseSink != nil {
+			state = openAIResponsesTunnelCodecStateForSink(responseSink)
+			responseSink.bindAttempt(req.stream, transport.tunnel.Dispatch())
 		}
+		state.reset()
+		src := newOpenAITunnelEndpointEventSource(
+			transport.tunnel.Stream(), transport.tunnel.WaitTimeout(),
+			rewriter, assembler, req.endpoint, state,
+		)
 		tracking := &openAIStreamGateUsageTrackingTunnelSource{openAITunnelEventSource: src, usage: usage, attempt: transport.usage}
 		return newOpenAIRecoverySourceEventSource(tracking, recoverySource), nil
 	}
@ -1863,13 +1912,15 @@ func (s *Server) buildOpenAITunnelStreamGateRuntime(
 	initialAssembler := &providerChatAssembler{streaming: req.stream}
 	initialRewriter := newProviderModelRewriter(req.stream, req.requestModel)
 	initialState := openAITunnelCodecStateForSink(sink)
-	initialState.reset()
-	var initialEventSource *openAITunnelEventSource
-	if semanticEnabled {
-		initialEventSource = newOpenAITunnelEndpointEventSource(handle.Stream(), handle.WaitTimeout(), initialRewriter, initialAssembler, req.endpoint, initialState)
-	} else {
-		initialEventSource = newOpenAITunnelEventSource(handle.Stream(), handle.WaitTimeout(), initialRewriter, initialAssembler, initialState)
+	if responseSink != nil {
+		initialState = openAIResponsesTunnelCodecStateForSink(responseSink)
+		responseSink.bindAttempt(req.stream, dispatch)
 	}
+	initialState.reset()
+	initialEventSource := newOpenAITunnelEndpointEventSource(
+		handle.Stream(), handle.WaitTimeout(), initialRewriter, initialAssembler,
+		req.endpoint, initialState,
+	)
 	initialSource := &openAIStreamGateUsageTrackingTunnelSource{
 		openAITunnelEventSource: initialEventSource,
 		usage:                   usage,
@ -1948,8 +1999,14 @@ func (s *Server) runOpenAITunnelStreamGate(w http.ResponseWriter, r *http.Reques
 	req.semanticEnabled = s.streamGateSemanticEnabled()
 	req.semanticSet = true
 	flusher, _ := w.(http.Flusher)
-	var sink *openAITunnelReleaseSink
-	if req.stream {
+	var sink openAIStreamGateSink
+	if req.stream && req.endpoint == openAIRebuildEndpointResponses {
+		sink = newOpenAIResponsesPoolReleaseSink(
+			w,
+			&openAIResponsesResultHolder{},
+			newOpenAIStreamGateCodecSelector(openAIStreamGateCodecTunnel),
+		)
+	} else if req.stream {
 		sink = newOpenAITunnelReleaseSink(w, flusher)
 	} else {
 		sink = newOpenAIBufferedTunnelReleaseSink(w, flusher, req.requestModel)
--- a/apps/edge/internal/openai/stream_gate_stall_recovery_test.go
+++ b/apps/edge/internal/openai/stream_gate_stall_recovery_test.go
@ -244,6 +244,172 @@ func assertStallAttemptClosedOnce(t *testing.T, service *scriptedPoolRunService,
 	}
 }

+func logicalFinishStallAttempt(runID, content string, committed bool) scriptedPoolAttempt {
+	wire := []byte(fmt.Sprintf("data: {\"id\":\"logical-finish\",\"object\":\"chat.completion.chunk\",\"choices\":[{\"index\":0,\"delta\":{\"content\":%q},\"finish_reason\":\"stop\"}]}\n\n", content))
+	if committed {
+		// The first content frame opens the stream. The following finish-only
+		// frame has no semantic event, so its identifiable wire remains pending
+		// until [DONE]/END and must be discarded by the later stall terminal.
+		wire = append(
+			[]byte("data: {\"id\":\"logical-finish\",\"object\":\"chat.completion.chunk\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"committed-prefix\"},\"finish_reason\":null}]}\n\n"),
+			[]byte(fmt.Sprintf("data: {\"id\":%q,\"object\":\"chat.completion.chunk\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n", content))...,
+		)
+	}
+	return scriptedPoolAttempt{
+		path: string(edgeservice.ProviderPoolPathTunnel), runID: runID, provider: "provider-a", target: "served-a",
+		frames: bufferedTunnelFrames(
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_RESPONSE_START, StatusCode: http.StatusOK, Headers: map[string]string{"Content-Type": "text/event-stream"}},
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_BODY, Body: wire},
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_ERROR, Failure: confirmedStallFailure("available")},
+		),
+	}
+}
+
+func logicalFinishMatrixServer(service runService, holdRunes int) *Server {
+	budget := 1
+	srv := NewServer(config.EdgeOpenAIConf{
+		TimeoutSec: 5,
+		StreamEvidenceGate: config.StreamEvidenceGateConf{
+			Enabled: true, MaxRequestFaultRecovery: &budget,
+			Filters: []config.StreamGateFilterPolicyConf{{
+				Filter: config.StreamGateFilterRepeatGuard, Enforcement: config.StreamGateFilterEnforcementBlocking,
+				Priority: 10, HoldEvidenceRunes: holdRunes,
+			}},
+		},
+	}, service, nil)
+	srv.SetModelCatalog([]config.ModelCatalogEntry{{
+		ID: "matrix-model", ContextWindowTokens: 8192,
+		Providers: map[string]string{"provider-a": "served-a", "provider-b": "served-b"},
+	}})
+	return srv
+}
+
+// TestOpenAIStallAfterLogicalFinishMatrix drives the observed shape through
+// the production tunnel codec/Core/sink: finish_reason is pending protocol
+// wire, while only the typed stall terminal decides replay or sanitized close.
+func TestOpenAIStallAfterLogicalFinishMatrix(t *testing.T) {
+	t.Run("pre-commit pending finish replays once", func(t *testing.T) {
+		const pending = "pending-finish-must-stay-hidden"
+		const recovered = "recovered-output-only"
+		service := newScriptedPoolRunService(
+			logicalFinishStallAttempt("logical-pre-a", pending, false),
+			stallMatrixSuccessAttempt(openAIRebuildEndpointChat, string(edgeservice.ProviderPoolPathTunnel), true, "logical-pre-b", "provider-b", recovered),
+		)
+		w := runStallMatrixHandler(t, logicalFinishMatrixServer(service, 500), openAIRebuildEndpointChat, true, nil)
+		body := w.Body.String()
+		if w.Code != http.StatusOK || service.poolSubmits() != 2 || !strings.Contains(body, recovered) || strings.Contains(body, pending) {
+			t.Fatalf("pre-commit recovery=(status=%d dispatches=%d body=%q)", w.Code, service.poolSubmits(), body)
+		}
+		if strings.Contains(body, "provider body") || strings.Contains(body, "raw provider metadata") {
+			t.Fatalf("raw failure data leaked: %q", body)
+		}
+		requests := stallPoolRequests(service)
+		if len(requests) != 2 || requests[1].AvoidProviderID != "provider-a" || !requests[1].AllowAvoidedProviderFallback {
+			t.Fatalf("pre-commit replay requests=%+v", requests)
+		}
+		if strings.Count(body, recovered) != 1 {
+			t.Fatalf("recovered output count is not one: %q", body)
+		}
+		assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "logical-pre-a")
+		assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "logical-pre-b")
+	})
+
+	t.Run("post-commit finish emits one sanitized terminal", func(t *testing.T) {
+		const pending = "pending-finish-after-commit"
+		service := newScriptedPoolRunService(
+			logicalFinishStallAttempt("logical-post-a", pending, true),
+			stallMatrixSuccessAttempt(openAIRebuildEndpointChat, string(edgeservice.ProviderPoolPathTunnel), true, "forbidden", "provider-b", "must-not-replay"),
+		)
+		w := runStallMatrixHandler(t, logicalFinishMatrixServer(service, 1), openAIRebuildEndpointChat, true, nil)
+		body := w.Body.String()
+		if w.Code != http.StatusOK || service.poolSubmits() != 1 || !strings.Contains(body, "committed-prefix") || strings.Contains(body, "must-not-replay") {
+			t.Fatalf("post-commit recovery=(status=%d dispatches=%d body=%q)", w.Code, service.poolSubmits(), body)
+		}
+		if strings.Count(body, `"type":"run_error"`) != 1 || strings.Count(body, "data: [DONE]") != 1 {
+			t.Fatalf("post-commit terminal count mismatch: %q", body)
+		}
+		if strings.Contains(body, "provider body") || strings.Contains(body, "raw provider metadata") || strings.Contains(body, pending) {
+			t.Fatalf("pending/raw bytes leaked after commit: %q", body)
+		}
+		assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "logical-post-a")
+	})
+}
+
+func semanticDisabledCommittedStallAttempt(endpoint, runID, pending string) scriptedPoolAttempt {
+	body := []byte(
+		"data: {\"id\":\"disabled-chat\",\"object\":\"chat.completion.chunk\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"disabled-committed-prefix\"},\"finish_reason\":null}]}\n\n" +
+			fmt.Sprintf("data: {\"id\":%q,\"object\":\"chat.completion.chunk\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n", pending),
+	)
+	if endpoint == openAIRebuildEndpointResponses {
+		body = []byte(
+			"data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp-disabled\",\"object\":\"response\",\"status\":\"in_progress\"},\"sequence_number\":1}\n\n" +
+				"data: {\"type\":\"response.output_text.delta\",\"item_id\":\"msg-disabled\",\"output_index\":0,\"content_index\":0,\"delta\":\"disabled-committed-prefix\",\"sequence_number\":2}\n\n" +
+				fmt.Sprintf("data: {\"type\":\"response.completed\",\"response\":{\"id\":%q,\"object\":\"response\",\"status\":\"completed\"},\"sequence_number\":3}\n\n", pending),
+		)
+	}
+	return scriptedPoolAttempt{
+		path: string(edgeservice.ProviderPoolPathTunnel), runID: runID, provider: "provider-a", target: "served-a",
+		frames: bufferedTunnelFrames(
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_RESPONSE_START, StatusCode: http.StatusOK, Headers: map[string]string{"Content-Type": "text/event-stream"}},
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_BODY, Body: body},
+			&iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_ERROR, Failure: confirmedStallFailure("available")},
+		),
+	}
+}
+
+// TestOpenAISemanticGateDisabledStallTerminalMatrix proves that endpoint
+// framing and the private typed-stall owner remain active when semantic filters
+// are disabled. Pre-commit stalls replay once; an opened Chat/Responses stream
+// never replays and receives one sanitized endpoint-native terminal.
+func TestOpenAISemanticGateDisabledStallTerminalMatrix(t *testing.T) {
+	for _, endpoint := range []string{openAIRebuildEndpointChat, openAIRebuildEndpointResponses} {
+		t.Run(endpoint+"/pre-commit", func(t *testing.T) {
+			const recovered = "semantic-disabled-recovered"
+			service := newScriptedPoolRunService(
+				stallMatrixFailureAttempt(string(edgeservice.ProviderPoolPathTunnel), "disabled-pre-a", "provider-a", "available"),
+				stallMatrixSuccessAttempt(endpoint, string(edgeservice.ProviderPoolPathTunnel), true, "disabled-pre-b", "provider-b", recovered),
+			)
+			w := runStallMatrixHandler(t, stallMatrixServer(service, false, 1), endpoint, true, nil)
+			body := w.Body.String()
+			if w.Code != http.StatusOK || service.poolSubmits() != 2 || strings.Count(body, recovered) != 1 {
+				t.Fatalf("pre-commit response=(status=%d dispatches=%d body=%q)", w.Code, service.poolSubmits(), body)
+			}
+			if strings.Contains(body, "provider body") || strings.Contains(body, "raw provider metadata") {
+				t.Fatalf("raw failure data leaked: %q", body)
+			}
+			assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "disabled-pre-a")
+			assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "disabled-pre-b")
+		})
+
+		t.Run(endpoint+"/post-commit", func(t *testing.T) {
+			const pending = "semantic-disabled-pending-terminal"
+			service := newScriptedPoolRunService(
+				semanticDisabledCommittedStallAttempt(endpoint, "disabled-post-a", pending),
+				stallMatrixSuccessAttempt(endpoint, string(edgeservice.ProviderPoolPathTunnel), true, "forbidden", "provider-b", "must-not-replay"),
+			)
+			w := runStallMatrixHandler(t, stallMatrixServer(service, false, 1), endpoint, true, nil)
+			body := w.Body.String()
+			if w.Code != http.StatusOK || service.poolSubmits() != 1 || !strings.Contains(body, "disabled-committed-prefix") || strings.Contains(body, "must-not-replay") {
+				t.Fatalf("post-commit response=(status=%d dispatches=%d body=%q)", w.Code, service.poolSubmits(), body)
+			}
+			if strings.Contains(body, pending) || strings.Contains(body, "provider body") || strings.Contains(body, "raw provider metadata") {
+				t.Fatalf("pending/raw failure data leaked: %q", body)
+			}
+			if strings.Count(body, "data: [DONE]") != 1 {
+				t.Fatalf("DONE terminal count mismatch: %q", body)
+			}
+			if endpoint == openAIRebuildEndpointChat {
+				if strings.Count(body, `"type":"run_error"`) != 1 {
+					t.Fatalf("Chat error terminal count mismatch: %q", body)
+				}
+			} else if strings.Count(body, `"type":"error"`) != 1 || strings.Contains(body, `"type":"response.completed"`) {
+				t.Fatalf("Responses error terminal mismatch: %q", body)
+			}
+			assertStallAttemptClosedOnce(t, service, string(edgeservice.ProviderPoolPathTunnel), "disabled-post-a")
+		})
+	}
+}
+
 // TestOpenAIStallRecoveryMatrix proves S05 through the supported production
 // handlers and the production runtime adapter. It covers every endpoint/path/
 // semantic-policy recovery product, then exercises the shared budget and every
--- a/apps/node/internal/node/liveness_watchdog_test.go
+++ b/apps/node/internal/node/liveness_watchdog_test.go
@ -453,6 +453,79 @@ func TestTunnelSinkStallClaimSerializesAcceptedFrame(t *testing.T) {
 	}
 }

+// TestTunnelWatchdogFinishFrameWithoutEndStallsOnce reproduces a provider
+// stream that has emitted a logical OpenAI finish frame but has not closed its
+// HTTP response. The finish frame is body progress, not a transport terminal:
+// the watchdog must expire once and fence every later usage/END frame.
+func TestTunnelWatchdogFinishFrameWithoutEndStallsOnce(t *testing.T) {
+	clock := newManualAttemptClock()
+	adapter := newControlledWatchdogAdapter("tunnel-finish-without-end")
+	n := newWatchdogNode(t, adapter, clock)
+	pipe := newWatchdogPipe(t)
+	done := make(chan error, 1)
+	go func() {
+		done <- n.OnProviderTunnelRequest(context.Background(), pipe.sess, &iop.ProviderTunnelRequest{
+			RunId: "tunnel-finish-without-end", TunnelId: "tunnel", Adapter: adapter.Name(), Target: "target", ResponseStallTimeoutMs: 1000,
+		})
+	}()
+	call := <-adapter.tunnelCalls
+
+	if err := call.sink.EmitTunnelFrame(context.Background(), runtime.ProviderTunnelFrame{
+		RunID: "tunnel-finish-without-end", TunnelID: "tunnel", Kind: runtime.ProviderTunnelFrameKindResponseStart,
+		StatusCode: 200, Headers: map[string]string{"Content-Type": "text/event-stream"},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	logicalFinish := []byte("data: {\"choices\":[{\"delta\":{\"content\":\"finished body\"},\"finish_reason\":\"stop\"}]}\n\n")
+	if err := call.sink.EmitTunnelFrame(context.Background(), runtime.ProviderTunnelFrame{
+		RunID: "tunnel-finish-without-end", TunnelID: "tunnel", Kind: runtime.ProviderTunnelFrameKindBody, Body: logicalFinish,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if frame := waitTunnelFrame(t, pipe.frames); frame.GetKind() != iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_RESPONSE_START {
+		t.Fatalf("response start = %+v", frame)
+	}
+	if frame := waitTunnelFrame(t, pipe.frames); frame.GetKind() != iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_BODY || string(frame.GetBody()) != string(logicalFinish) {
+		t.Fatalf("logical finish body = %+v", frame)
+	}
+
+	stallTimer := clock.waitTimer(t, 0)
+	requireTimerDurations(t, stallTimer, time.Second, time.Second, time.Second)
+	stallTimer.fire()
+	waitContextCanceled(t, call.ctx)
+	grace := clock.waitTimer(t, 1)
+	adapter.tunnelReturn <- nil
+	if err := <-done; err != errProviderResponseStalled {
+		t.Fatalf("tunnel result = %v", err)
+	}
+	terminal := waitTunnelFrame(t, pipe.frames)
+	if terminal.GetKind() != iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_ERROR ||
+		terminal.GetFailure().GetCode() != string(runtime.FailureCodeResponseStalled) ||
+		terminal.GetMetadata()["failure_code"] != string(runtime.FailureCodeResponseStalled) {
+		t.Fatalf("stall terminal = %+v", terminal)
+	}
+	requireTimerDurations(t, grace, defaultAttemptCloseGrace)
+
+	// These are the exact late frames an adapter can race after cancellation.
+	// Both must be accepted as no-ops by the fenced sink.
+	if err := call.sink.EmitTunnelFrame(context.Background(), runtime.ProviderTunnelFrame{
+		RunID: "tunnel-finish-without-end", TunnelID: "tunnel", Kind: runtime.ProviderTunnelFrameKindUsage,
+		Usage: &runtime.UsageStats{InputTokens: 1, OutputTokens: 1},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := call.sink.EmitTunnelFrame(context.Background(), runtime.ProviderTunnelFrame{
+		RunID: "tunnel-finish-without-end", TunnelID: "tunnel", Kind: runtime.ProviderTunnelFrameKindEnd, End: true,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	select {
+	case frame := <-pipe.frames:
+		t.Fatalf("late frame escaped the terminal fence: %+v", frame)
+	default:
+	}
+}
+
 func TestRunWatchdogStaleExpiryYieldsToProgress(t *testing.T) {
 	clock := newManualAttemptClock()
 	adapter := newControlledWatchdogAdapter("run-stale-expiry")
--- a/docs/edge-local-dev-guide.md
+++ b/docs/edge-local-dev-guide.md
@ -72,6 +72,18 @@ nodes:

 `models[]`를 사용할 경우 `openai.adapter`/`openai.target`은 하향 호환 fallback이다. 실제 routing은 `nodes[].providers[].id`를 기준으로 provider-pool에서 선택된다.

+### Provider response-stall timeout ownership
+
+dev-runtime에서 긴 응답의 정지 판정은 아래 순서를 유지한다.
+
+```text
+Node provider response_stall_timeout_ms = 120000
+external caller boundary               ≈ 180000
+Edge request hard timeout              > 180000
+```
+
+`response_stall_timeout_ms` 변경은 restart-required다. `config check`와 `config refresh --mode dry-run`에서 이를 확인한 뒤 Edge와 Node를 같은 source ref로 rebuild/restart하고, 각 binary의 build identity와 실행 중인 process identity를 다시 대조한다. tracked 검증 근거에는 source/build/config 식별자, 단조 시간, terminal 개수와 결과 분류만 남기며 prompt, output, token, credential 원문은 기록하지 않는다.
+
 확인:

 ```bash
--- a/scripts/e2e-openai-lemonade.sh
+++ b/scripts/e2e-openai-lemonade.sh
@ -43,8 +43,14 @@ cleanup() {
        exec 3>&- 2>/dev/null || true
    fi
    if [ -n "$EDGE_PID" ]; then wait "$EDGE_PID" 2>/dev/null || true; fi
-    if [ -n "$NODE_PID" ]; then kill "$NODE_PID" 2>/dev/null || true; fi
-    if [ -n "$LEMONADE_PID" ]; then kill "$LEMONADE_PID" 2>/dev/null || true; fi
+    if [ -n "$NODE_PID" ]; then
+        kill "$NODE_PID" 2>/dev/null || true
+        wait "$NODE_PID" 2>/dev/null || true
+    fi
+    if [ -n "$LEMONADE_PID" ]; then
+        kill "$LEMONADE_PID" 2>/dev/null || true
+        wait "$LEMONADE_PID" 2>/dev/null || true
+    fi
    if [ "${IOP_LEMONADE_KEEP_TMP:-0}" = "1" ]; then
        echo "[openai-lemonade] keeping temp dir: $TMP_DIR"
        return
@ -116,6 +122,8 @@ authorization := os.Args[3]
 var modelRequests atomic.Int64
 var chatRequests atomic.Int64
 var responseRequests atomic.Int64
+var hangRequests atomic.Int64
+var hangCancellations atomic.Int64
 mux := http.NewServeMux()
 mux.HandleFunc("/v1/models", func(w http.ResponseWriter, r *http.Request) {
 if r.Method != http.MethodGet || r.RequestURI != "/v1/models" {
@ -128,7 +136,7 @@ return
 }
 modelRequests.Add(1)
 w.Header().Set("Content-Type", "application/json")
-_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"` + model + `"}]}`))
+_, _ = w.Write([]byte(`{"object":"list","data":[{"id":"` + model + `"},{"id":"lemonade-profile-alias"}]}`))
 })
 mux.HandleFunc("/v1/chat/completions", func(w http.ResponseWriter, r *http.Request) {
 if r.Method != http.MethodPost || r.RequestURI != "/v1/chat/completions" {
@ -160,6 +168,15 @@ if f, ok := w.(http.Flusher); ok {
 f.Flush()
 }
 }
+if len(req.Messages) == 1 && req.Messages[0].Content == "IOP_HANG_AFTER_FINISH" {
+hangRequests.Add(1)
+_, _ = w.Write([]byte("data: " + `{"choices":[{"delta":{"content":"IOP_HANG_VISIBLE_PREFIX"},"finish_reason":"stop"}]}` + "\n\n"))
+flush()
+<-r.Context().Done()
+hangCancellations.Add(1)
+log.Print("IOP_HANG_PROVIDER_RAW_DIAGNOSTIC")
+return
+}
 _, _ = w.Write([]byte("data: " + `{"choices":[{"delta":{"reasoning_content":"thinking..."}}]}` + "\n\n"))
 flush()
 _, _ = w.Write([]byte("data: " + `{"choices":[{"delta":{"content":"IOP_OPENAI_"}}]}` + "\n\n"))
@ -199,19 +216,23 @@ return
 models := modelRequests.Load()
 chats := chatRequests.Load()
 responses := responseRequests.Load()
-if models < 1 || chats != 2 || responses != 1 {
-http.Error(w, fmt.Sprintf("unexpected request counts: models=%d chats=%d responses=%d", models, chats, responses), http.StatusInternalServerError)
+hangs := hangRequests.Load()
+cancellations := hangCancellations.Load()
+if models < 1 || chats != 4 || responses != 1 || hangs != 1 || cancellations != 1 {
+http.Error(w, fmt.Sprintf("unexpected request counts: models=%d chats=%d responses=%d hangs=%d cancellations=%d", models, chats, responses, hangs, cancellations), http.StatusInternalServerError)
 return
 }
 w.Header().Set("Content-Type", "application/json")
-_, _ = fmt.Fprintf(w, `{"models":%d,"chats":%d,"responses":%d}`, models, chats, responses)
+_, _ = fmt.Fprintf(w, `{"models":%d,"chats":%d,"responses":%d,"hangs":%d,"cancellations":%d}`, models, chats, responses, hangs, cancellations)
 })
 log.Fatal(http.ListenAndServe(os.Args[1], mux))
 }
 EOF

    FAKE_AUTHORIZATION="Bearer fake-lemonade-token"
-    go run "$FAKE_LEMONADE" "127.0.0.1:$LEMONADE_PORT" "$MODEL" "$FAKE_AUTHORIZATION" > "$TMP_DIR/fake_lemonade.out" 2>&1 &
+    FAKE_LEMONADE_BIN="$TMP_DIR/fake-lemonade"
+    go build -o "$FAKE_LEMONADE_BIN" "$FAKE_LEMONADE"
+    "$FAKE_LEMONADE_BIN" "127.0.0.1:$LEMONADE_PORT" "$MODEL" "$FAKE_AUTHORIZATION" > "$TMP_DIR/fake_lemonade.out" 2>&1 &
    LEMONADE_PID=$!
    wait_port 127.0.0.1 "$LEMONADE_PORT" "fake lemonade"
 fi
@ -282,6 +303,7 @@ $(cat "$PROVIDER_HEADER_BLOCK")
          - "lemonade-profile-alias"
        health: available
        capacity: 4
+        response_stall_timeout_ms: 200
 EOF

 cat > "$NODE_CONFIG" <<EOF
@ -359,6 +381,45 @@ else
 fi
 grep -q 'data: \[DONE\]' "$STREAM_OUT"

+if [ "$MODE" = "fake" ]; then
+    # Flush a logical finish and deliberately keep the provider response open.
+    # Node's 200ms typed watchdog owns the failure before Edge's 30s request
+    # timeout. The committed stream must receive one sanitized terminal only.
+    HANG_OUT="$TMP_DIR/hang-after-finish.txt"
+    HANG_STARTED=$SECONDS
+    curl -fsS -N \
+        -H "Content-Type: application/json" \
+        -d '{"model":"client-request-model","stream":true,"max_tokens":32,"messages":[{"role":"user","content":"IOP_HANG_AFTER_FINISH"}]}' \
+        "http://127.0.0.1:$OPENAI_PORT/v1/chat/completions" > "$HANG_OUT"
+    HANG_ELAPSED=$((SECONDS - HANG_STARTED))
+    if (( HANG_ELAPSED >= 10 )); then
+        echo "[openai-lemonade] hang fixture exceeded the bounded IOP liveness window"
+        exit 1
+    fi
+    HANG_PREFIX_COUNT=$(grep -c 'IOP_HANG_VISIBLE_PREFIX' "$HANG_OUT" || true)
+    HANG_FINISH_COUNT=$(grep -c '"finish_reason":"stop"' "$HANG_OUT" || true)
+    HANG_ERROR_COUNT=$(grep -c '"type":"run_error"' "$HANG_OUT" || true)
+    HANG_DONE_COUNT=$(grep -c '^data: \[DONE\]$' "$HANG_OUT" || true)
+    if [ "$HANG_PREFIX_COUNT" -ne 1 ] || [ "$HANG_FINISH_COUNT" -ne 1 ] || [ "$HANG_ERROR_COUNT" -ne 1 ] || [ "$HANG_DONE_COUNT" -ne 1 ]; then
+        echo "[openai-lemonade] hang fixture terminal counts prefix=$HANG_PREFIX_COUNT finish=$HANG_FINISH_COUNT run_error=$HANG_ERROR_COUNT done=$HANG_DONE_COUNT"
+        exit 1
+    fi
+    if grep -q 'IOP_HANG_PROVIDER_RAW_DIAGNOSTIC' "$HANG_OUT"; then
+        echo "[openai-lemonade] raw provider diagnostic leaked to the client"
+        exit 1
+    fi
+
+    # A normal request after cancellation proves the provider handler and the
+    # Node/Edge path remain usable after the fenced attempt.
+    POST_CANCEL_OUT="$TMP_DIR/post-cancel-stream.txt"
+    curl -fsS -N \
+        -H "Content-Type: application/json" \
+        -d '{"model":"client-request-model","stream":true,"max_tokens":32,"messages":[{"role":"user","content":"Reply after cancellation"}]}' \
+        "http://127.0.0.1:$OPENAI_PORT/v1/chat/completions" > "$POST_CANCEL_OUT"
+    grep -q '"content":"IOP_OPENAI_' "$POST_CANCEL_OUT"
+    [ "$(grep -c '^data: \[DONE\]$' "$POST_CANCEL_OUT")" -eq 1 ]
+fi
+
 # Command-path smoke via the edge CLI.
 SMOKE_OUT="$TMP_DIR/iop-edge-smoke-openai.txt"
 (cd "$REPO_ROOT" && go run ./apps/edge/cmd/edge smoke openai \
@ -371,15 +432,35 @@ grep -q "IOP Edge OpenAI Smoke Test SUCCESS!" "$SMOKE_OUT"
 if [ "$MODE" = "fake" ]; then
    FAKE_ASSERT_OUT="$TMP_DIR/fake_assert.json"
    curl -fsS "http://127.0.0.1:$LEMONADE_PORT/assert" > "$FAKE_ASSERT_OUT"
-    grep -q '"chats":2' "$FAKE_ASSERT_OUT"
+    grep -q '"chats":4' "$FAKE_ASSERT_OUT"
    grep -q '"responses":1' "$FAKE_ASSERT_OUT"
+    grep -q '"hangs":1' "$FAKE_ASSERT_OUT"
+    grep -q '"cancellations":1' "$FAKE_ASSERT_OUT"
 fi

-if grep -i -E "node reported error|error run_id=|\[[^]]+-evt\] error|panic:" "$EDGE_OUT" "$NODE_OUT" >/dev/null; then
+FAILURE_MATCHES="$TMP_DIR/failure-matches.txt"
+grep -i -E "node reported error|error run_id=|\[[^]]+-evt\] error|panic:" "$EDGE_OUT" "$NODE_OUT" > "$FAILURE_MATCHES" || true
+if [ "$MODE" = "fake" ]; then
+    UNEXPECTED_FAILURES="$TMP_DIR/unexpected-failures.txt"
+    grep -vi -E "response_stalled|provider response stalled" "$FAILURE_MATCHES" > "$UNEXPECTED_FAILURES" || true
+else
+    UNEXPECTED_FAILURES="$FAILURE_MATCHES"
+fi
+if [ -s "$UNEXPECTED_FAILURES" ]; then
    echo "[openai-lemonade] detected failure marker"
-    echo "=== EDGE OUTPUT ==="; cat "$EDGE_OUT"
-    echo "=== NODE OUTPUT ==="; cat "$NODE_OUT"
+    echo "[openai-lemonade] raw runtime output retained only in the temporary artifact directory"
    exit 1
 fi

+if [ "$MODE" = "fake" ]; then
+    grep -qi -E "response_stalled|provider response stalled" "$EDGE_OUT" "$NODE_OUT"
+    kill "$LEMONADE_PID"
+    wait "$LEMONADE_PID" 2>/dev/null || true
+    if kill -0 "$LEMONADE_PID" 2>/dev/null; then
+        echo "[openai-lemonade] fake provider survived teardown"
+        exit 1
+    fi
+    LEMONADE_PID=""
+fi
+
 echo "[openai-lemonade] OpenAI-compatible Lemonade serving test PASSED (mode=$MODE)."