Bound provider terminal labels in evaluation records

Why: A provider-controlled finish reason could leak arbitrary content into the supposed content-free audit file.

Evidence: Red then green terminal-label regression and full Backend vet/Go test in Dev receipts.

Impact: Only evaluation records map unknown terminal labels to a finite class; tool execution rules are unchanged.

Refs: U3 r0.5 independent review finding

internal/server/work_checkpoint_eval_budget_test.go

@@ -74,6 +74,41 @@ func TestEvaluationRecordsContentFreeTerminal(t *testing.T) {
 	}
 }
 
+func TestEvaluationRedactsProviderControlledTerminalLabels(t *testing.T) {
+	dir, _ := newEvaluationEventStore(t)
+	provider := func(context.Context, string, []agentllm.Message, []agentllm.ToolDecl, agentllm.StreamOptions) <-chan agentllm.Event {
+		out := make(chan agentllm.Event, 1)
+		out <- agentllm.Event{Kind: agentllm.KindMessageEnd,
+			Message:      agentllm.Message{Role: agentllm.RoleAssistant, Content: "safe", StopReason: agentllm.StopEnd},
+			Usage:        agentllm.Usage{PromptTokens: 12, CompletionTokens: 5, TotalTokens: 17},
+			FinishReason: "PRIVATE_REASON_R05", StreamComplete: true,
+			TerminalObservation: &agentllm.TerminalObservation{ReadOutcome: "PRIVATE_OUTCOME_R05", UsagePresent: true},
+		}
+		close(out)
+		return out
+	}
+	for range evalProviderStreamWithProvider(provider, dir)(evalTestContext(t), "fixture", nil, nil, agentllm.StreamOptions{MaxTokens: 4000}) {
+	}
+	raw, err := os.ReadFile(filepath.Join(dir, "provider-records.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(raw), "PRIVATE_REASON_R05") || strings.Contains(string(raw), "PRIVATE_OUTCOME_R05") {
+		t.Fatal("provider-controlled terminal label escaped into a content-free record")
+	}
+	var record struct {
+		Calls []struct {
+			FinishReason string                        `json:"finishReason"`
+			Observation  *agentllm.TerminalObservation `json:"terminalObservation"`
+		} `json:"calls"`
+	}
+	if err := json.Unmarshal(raw, &record); err != nil || len(record.Calls) != 1 ||
+		record.Calls[0].FinishReason != "unknown" || record.Calls[0].Observation == nil ||
+		record.Calls[0].Observation.ReadOutcome != "unknown" {
+		t.Fatalf("unexpected sanitized terminal labels: %+v err=%v", record, err)
+	}
+}
+
 func TestEvaluationRedactsErrorTerminalAndBlocksTool(t *testing.T) {
 	dir, _ := newEvaluationEventStore(t)
 	var calls atomic.Int32

internal/server/work_checkpoint_eval_test.go

@@ -413,6 +413,28 @@ type evalCallRecord struct {
 	FailureClass        string                        `json:"failureClass,omitempty"`
 }
 
+func evalFinishReasonClass(reason string) string {
+	switch reason {
+	case "", "stop", "length", "tool_calls", "function_call", "content_filter":
+		return reason
+	default:
+		return "unknown"
+	}
+}
+
+func evalTerminalObservation(observation *agentllm.TerminalObservation) *agentllm.TerminalObservation {
+	if observation == nil {
+		return nil
+	}
+	copy := *observation
+	switch copy.ReadOutcome {
+	case "done_marker", "clean_eof", "read_error", "cancelled", "invalid_frame", "provider_error", "transport_error":
+	default:
+		copy.ReadOutcome = "unknown"
+	}
+	return &copy
+}
+
 func evalProviderStreamWithProvider(provider agentllm.StreamFunc, dir string) agentllm.StreamFunc {
 	batchLimit := 120000
 	if raw := os.Getenv("WORK_CHECKPOINT_REMAINING_TOKENS"); raw != "" {
@@ -565,12 +587,12 @@ func evalProviderStreamWithProvider(provider agentllm.StreamFunc, dir string) ag
 			var terminalObservation *agentllm.TerminalObservation
 			httpStatus := 0
 			if terminal != nil {
-				finishReason, streamComplete = terminal.FinishReason, terminal.StreamComplete
+				finishReason, streamComplete = evalFinishReasonClass(terminal.FinishReason), terminal.StreamComplete
 				returnedUsage = terminal.Usage
-				terminalObservation = terminal.TerminalObservation
+				terminalObservation = evalTerminalObservation(terminal.TerminalObservation)
 				httpStatus = terminal.HTTPStatus
 			} else if streamError != nil {
-				terminalObservation = streamError.TerminalObservation
+				terminalObservation = evalTerminalObservation(streamError.TerminalObservation)
 				httpStatus = streamError.HTTPStatus
 			}
 			failureClass := ""