record final core m5 production evidence

adapters/models/openai_compatible_local.go

@@ -25,6 +25,13 @@ type OpenAICompatibleLocalAdapter struct {
 	apiKey string
 }
 
+// compatibleDefaultContextBytes is a conservative bound used when an
+// OpenAI-compatible local endpoint omits n_ctx from /v1/models. The adapter
+// is a chat-completions adapter, so advisory capability is known from the
+// adapter contract; quality and activation still require governed
+// qualification/evaluation in the capability registry.
+const compatibleDefaultContextBytes int64 = 32 << 10
+
 func NewOpenAICompatibleLocalAdapter(raw string, transport http.RoundTripper) (*OpenAICompatibleLocalAdapter, error) {
 	return NewOpenAICompatibleLocalAdapterWithAPIKey(raw, "", transport)
 }
@@ -90,7 +97,10 @@ func (a *OpenAICompatibleLocalAdapter) Discover(ctx context.Context) ([]Discover
 		stableMetadata := fmt.Sprintf("%s\x00%s\x00%d\x00%d\x00%d\x00%d\x00%s", item.ID, item.OwnedBy, item.Meta.Context, item.Meta.TrainingContext, item.Meta.Parameters, item.Meta.Size, item.Meta.FileType)
 		digest := sha256.Sum256([]byte(stableMetadata))
 		contextBytes := item.Meta.Context * 4
-		models = append(models, DiscoveredModel{ProviderModelID: capability.ProviderModelID(item.ID), Fingerprint: capability.ModelFingerprint(hex.EncodeToString(digest[:])), MaxContextBytes: contextBytes, Modalities: []capability.ModelModality{capability.ModalityText}, Capabilities: []capability.ModelCapability{capability.CapabilityUnknown}, SizeBytes: item.Meta.Size, Runtime: item.OwnedBy, Quantization: item.Meta.FileType})
+		if contextBytes <= 0 {
+			contextBytes = compatibleDefaultContextBytes
+		}
+		models = append(models, DiscoveredModel{ProviderModelID: capability.ProviderModelID(item.ID), Fingerprint: capability.ModelFingerprint(hex.EncodeToString(digest[:])), MaxContextBytes: contextBytes, Modalities: []capability.ModelModality{capability.ModalityText}, Capabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, SizeBytes: item.Meta.Size, Runtime: item.OwnedBy, Quantization: item.Meta.FileType})
 	}
 	return models, ProviderProbe{Health: capability.HealthAvailable}
 }

adapters/models/openai_compatible_local_test.go

@@ -18,6 +18,34 @@ type captureRoundTripper struct {
 	body          []byte
 }
 
+type modelsMetadataRoundTripper struct{}
+
+func (modelsMetadataRoundTripper) RoundTrip(request *http.Request) (*http.Response, error) {
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(`{"data":[{"id":"medium-model","owned_by":"llama.cpp","meta":{}}]}`)),
+		Header:     make(http.Header),
+		Request:    request,
+	}, nil
+}
+
+func TestOpenAICompatibleLocalAdapterNormalizesMinimalModelMetadata(t *testing.T) {
+	adapter, err := NewOpenAICompatibleLocalAdapter("http://gateway/v1", modelsMetadataRoundTripper{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	models, probe := adapter.Discover(context.Background())
+	if probe.Health != capability.HealthAvailable || len(models) != 1 {
+		t.Fatalf("unexpected discovery result: probe=%+v models=%+v", probe, models)
+	}
+	if models[0].MaxContextBytes != compatibleDefaultContextBytes {
+		t.Fatalf("context bound=%d, want %d", models[0].MaxContextBytes, compatibleDefaultContextBytes)
+	}
+	if len(models[0].Capabilities) != 1 || models[0].Capabilities[0] != capability.CapabilityAdvisory {
+		t.Fatalf("capabilities=%v, want advisory", models[0].Capabilities)
+	}
+}
+
 func TestOpenAICompatibleLocalAdapterUsesConfiguredTimeout(t *testing.T) {
 	adapter, err := NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout("http://gateway/v1", "", &captureRoundTripper{}, 47*time.Second)
 	if err != nil {

adapters/postgres/capability_repository.go

@@ -181,10 +181,6 @@ func (r *CapabilityRepository) SaveBinding(ctx context.Context, binding capabili
 		if err != nil {
 			return err
 		}
-		_, err = tx.Exec(ctx, `INSERT INTO capability.cognitive_binding_revision_heads (binding_id,revision) VALUES ($1,$2) ON CONFLICT DO NOTHING`, binding.ID, binding.Revision)
-		if err != nil {
-			return err
-		}
 	}
 	return tx.Commit(ctx)
 }
@@ -437,6 +433,14 @@ func (r *CapabilityRepository) LoadCognitiveCandidatesForProfile(ctx context.Con
 	return r.loadCognitiveCandidates(ctx, capabilityID, owner, profileID)
 }
 
+func (r *CapabilityRepository) LoadCognitiveCandidatesForEvaluation(ctx context.Context, capabilityID capability.CapabilityID, owner identity.PrincipalID, evaluationID capability.EvaluationID) ([]capability.ResolvedCognitiveBinding, error) {
+	var profileID capability.TaskEvaluationProfileID
+	if err := r.pool.QueryRow(ctx, `SELECT task_profile_id FROM capability.cognitive_task_evaluations WHERE evaluation_id=$1 AND capability_id=$2`, evaluationID, capabilityID).Scan(&profileID); err != nil {
+		return nil, err
+	}
+	return r.loadCognitiveCandidates(ctx, capabilityID, owner, profileID)
+}
+
 func (r *CapabilityRepository) LoadCognitiveCandidate(ctx context.Context, bindingID capability.CapabilityBindingID, revision int64) (capability.ResolvedCognitiveBinding, error) {
 	var capabilityID capability.CapabilityID
 	var owner identity.PrincipalID
@@ -564,7 +568,12 @@ func (r *CapabilityRepository) SaveCognitiveDiscovery(ctx context.Context, value
 		if err != nil {
 			return err
 		}
-		_, err = tx.Exec(ctx, `INSERT INTO capability.cognitive_binding_revision_heads (binding_id,revision) VALUES ($1,$2) ON CONFLICT DO NOTHING`, binding.ID, binding.Revision)
+		// Discovery may advance a head, but must never roll back a governed
+		// CANARY/ACTIVE revision to an older discovered snapshot after restart.
+		// Discovery may advance an inactive head even when its deterministic
+		// revision number is lower, but it must never overwrite a governed canary
+		// or active head during restart/startup refresh.
+		_, err = tx.Exec(ctx, `INSERT INTO capability.cognitive_binding_revision_heads (binding_id,revision) VALUES ($1,$2) ON CONFLICT (binding_id) DO UPDATE SET revision=EXCLUDED.revision,updated_at=CURRENT_TIMESTAMP WHERE NOT EXISTS (SELECT 1 FROM capability.cognitive_bindings current_binding WHERE current_binding.binding_id=capability.cognitive_binding_revision_heads.binding_id AND current_binding.revision=capability.cognitive_binding_revision_heads.revision AND current_binding.activation IN ('CANARY','ACTIVE'))`, binding.ID, binding.Revision)
 		if err != nil {
 			return err
 		}

cmd/cognitive-worker/fleet_status.go

@@ -20,6 +20,7 @@ type fleetEndpoint struct {
 	URL        string
 	Headers    map[string]string
 	Client     *http.Client
+	Timeout    time.Duration
 }
 
 type fleetStatusResponse struct {
@@ -63,7 +64,11 @@ func probeFleetEndpoint(parent context.Context, fallbackClient *http.Client, end
 	if strings.TrimSpace(endpoint.URL) == "" {
 		return member
 	}
-	ctx, cancel := context.WithTimeout(parent, 2*time.Second)
+	timeout := endpoint.Timeout
+	if timeout <= 0 {
+		timeout = 2 * time.Second
+	}
+	ctx, cancel := context.WithTimeout(parent, timeout)
 	defer cancel()
 	request, err := http.NewRequestWithContext(ctx, http.MethodGet, endpoint.URL, nil)
 	if err != nil {

cmd/cognitive-worker/fleet_status_test.go

@@ -68,3 +68,21 @@ func TestFleetReadinessSendsConfiguredProviderHeaders(t *testing.T) {
 		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
 	}
 }
+
+func TestFleetReadinessUsesProviderSpecificTimeout(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		time.Sleep(40 * time.Millisecond)
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer server.Close()
+
+	handler := fleetReadinessHandler(server.Client(), capability.FleetTierMedium, []fleetEndpoint{{
+		ProviderID: "medium", ModelID: "medium-model", Tier: capability.FleetTierMedium,
+		URL: server.URL, Client: server.Client(), Timeout: 100 * time.Millisecond,
+	}})
+	recorder := httptest.NewRecorder()
+	handler.ServeHTTP(recorder, httptest.NewRequest(http.MethodGet, "/v1/fleet/readiness", nil))
+	if recorder.Code != http.StatusOK || !strings.Contains(recorder.Body.String(), `"selected_tier":"MEDIUM"`) {
+		t.Fatalf("provider-specific timeout was not honored: status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+}

cmd/cognitive-worker/main.go

@@ -71,7 +71,7 @@ func main() {
 		if err != nil {
 			log.Fatal(err)
 		}
-		mediumTimeout := durationEnv("LOCAL_MEDIUM_PROVIDER_TIMEOUT", 60*time.Second)
+		mediumTimeout := durationEnv("LOCAL_MEDIUM_PROVIDER_TIMEOUT", 120*time.Second)
 		mediumAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(strings.TrimRight(mediumURL, "/"), mediumAPIKey, nil, mediumTimeout)
 		if adapterErr != nil {
 			log.Fatal(adapterErr)
@@ -117,7 +117,8 @@ func main() {
 	mux.Handle("/healthz", health.Handler{Runtime: "cognitive-worker", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	readinessClient := &http.Client{Timeout: 2 * time.Second}
 	cpuReadinessClient := &http.Client{Timeout: 10 * time.Second}
-	mediumReadinessClient := &http.Client{Timeout: 10 * time.Second}
+		mediumReadinessTimeout := 120 * time.Second
+	mediumReadinessClient := &http.Client{Timeout: mediumReadinessTimeout}
 	lifeboatURL := strings.TrimSuffix(strings.TrimRight(envOr("LOCAL_LIFEBOAT_PROVIDER_URL", "http://cognitive-lifeboat:8086/v1"), "/"), "/v1") + "/health"
 	fleetClient := &http.Client{Timeout: 2 * time.Second}
 	preferredFleet := capability.FleetTier(envOr("COGNITIVE_PREFERRED_FLEET_TIER", string(capability.FleetTierFast)))
@@ -131,14 +132,21 @@ func main() {
 	// readiness budget short; their detailed fleet state is reported separately.
 	cpuHealthProbe.Timeout = 2 * time.Second
 	mediumEndpoint := ""
+	mediumHealthEndpoint := ""
 	mediumHeaders := map[string]string{}
 	mediumHealthProbe := health.Probe{}
 	if mediumURL != "" {
 		mediumEndpoint = strings.TrimRight(mediumURL, "/") + "/models"
+		mediumHealthEndpoint = strings.TrimSuffix(strings.TrimRight(mediumURL, "/"), "/v1") + "/health"
 		if mediumAPIKey != "" {
 			mediumHeaders["Authorization"] = "Bearer " + mediumAPIKey
 		}
-		mediumHealthProbe = health.OptionalHTTPProbeWithHeaders(mediumReadinessClient, mediumEndpoint, mediumHeaders)
+		// Readiness reports provider discovery health quickly; fleet readiness
+		// separately reports authenticated model availability on /v1/models.
+		mediumHealthProbe = health.OptionalHTTPProbeWithHeaders(readinessClient, mediumHealthEndpoint, mediumHeaders)
+		// The local gateway may need several seconds to enumerate a loaded model.
+		// Keep this bounded, but do not classify a healthy MEDIUM provider as
+		// unavailable before its authenticated discovery response can arrive.
 		mediumHealthProbe.Timeout = 2 * time.Second
 	}
 	readinessProbes := []health.Probe{
@@ -157,8 +165,8 @@ func main() {
 	}
 
 	fleetEndpoints := []fleetEndpoint{
-		{ProviderID: "local-cpu-primary", ModelID: "local-cpu-model", Tier: capability.FleetTierFast, URL: cpuEndpoint, Headers: cpuFleetHeaders, Client: &http.Client{Timeout: 10 * time.Second}},
-		{ProviderID: "local-medium-primary", ModelID: "local-medium-model", Tier: capability.FleetTierMedium, URL: mediumEndpoint, Headers: mediumHeaders, Client: mediumReadinessClient},
+		{ProviderID: "local-cpu-primary", ModelID: "local-cpu-model", Tier: capability.FleetTierFast, URL: cpuEndpoint, Headers: cpuFleetHeaders, Client: &http.Client{Timeout: 10 * time.Second}, Timeout: 2 * time.Second},
+		{ProviderID: "local-medium-primary", ModelID: "local-medium-model", Tier: capability.FleetTierMedium, URL: mediumEndpoint, Headers: mediumHeaders, Client: mediumReadinessClient, Timeout: mediumReadinessTimeout},
 		{ProviderID: "cognitive-lifeboat", ModelID: "qwen3-8b-lifeboat", Tier: capability.FleetTierLifeboat, URL: lifeboatURL},
 	}
 	if gpuConfigured {
@@ -217,7 +225,9 @@ func durationEnv(name string, fallback time.Duration) time.Duration {
 
 func refreshProviderDiscoveries(resolver bindingResolver, service discoveryService) {
 	for name := range service.providers {
-		ctx, cancel := context.WithTimeout(context.Background(), 8*time.Second)
+		// Keep startup discovery bounded while allowing the authenticated local
+		// gateway enough time to enumerate a loaded MEDIUM model.
+		ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
 		registration, ok := service.providers[name]
 		if !ok || registration.Discovery == nil {
 			cancel()

cmd/cognitive-worker/task_evaluation.go

@@ -100,7 +100,7 @@ func (g cognitiveGateway) executeTaskEvaluationCases(ctx context.Context, reques
 		go func(index int, item capability.TaskEvaluationCase) {
 			defer wg.Done()
 			started := time.Now()
-			invocation := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("task-evaluation:%s:%s", request.Profile.ID, item.CaseID)), ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: item.Prompt, MaxOutputBytes: int(request.MaxOutputBytes), ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
+			invocation := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("task-evaluation:%s:%s", request.Profile.ID, item.CaseID)), ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: item.Prompt, Locale: "en-US", MaxOutputBytes: int(request.MaxOutputBytes), ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
 			response, invokeErr := g.invokeAdmitted(ctx, string(invocation.RequestID), provider, invocation, int(request.MaxOutputBytes))
 			latency := time.Since(started).Microseconds()
 			observations[index] = capability.TaskCaseObservation{CaseID: item.CaseID, Result: capability.EvaluationUnknown, CriticalFailure: invokeErr != nil || response.FinishReason != cognitive.FinishCompleted || strings.TrimSpace(response.AdvisoryText) == "", EvidenceArtifactRef: item.EvidenceArtifactRef, LatencyMicros: latency}

cmd/cognitive-worker/task_evaluation_test.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -57,6 +58,30 @@ func TestTaskEvaluationHandlerInvokesEveryBoundedCase(t *testing.T) {
 	}
 }
 
+type localeCheckingProvider struct{ stubAdvisoryProvider }
+
+func (p *localeCheckingProvider) Advise(ctx context.Context, request cognitive.AdvisoryRequest) (cognitive.AdvisoryResponse, error) {
+	if request.Locale == "" {
+		return cognitive.AdvisoryResponse{}, fmt.Errorf("task evaluation omitted required locale")
+	}
+	return p.stubAdvisoryProvider.Advise(ctx, request)
+}
+
+func TestTaskEvaluationSuppliesLocaleToProviderAdapter(t *testing.T) {
+	selected := gatewaySelection()
+	provider := &localeCheckingProvider{stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishCompleted, AdvisoryText: "observed response"}}}
+	registry := &candidateRegistryStub{selected: selected}
+	gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{selected.Provider.Kind: {{ProviderInstanceID: selected.Provider.ID, Kind: selected.Provider.Kind, EndpointRef: selected.Provider.EndpointRef, CredentialRef: selected.Provider.CredentialRef, Provider: provider}}}, now: time.Now}
+	profile := capability.TaskEvaluationProfile{ID: "profile-locale", Revision: 1, CapabilityID: selected.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "locale-v1", RequiredCaseIDs: []string{"case-locale"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: selected.Binding.CapabilityID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: selected.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-locale", Prompt: "one", EvidenceArtifactRef: "artifact:locale"}}}
+	payload, _ := json.Marshal(request)
+	recorder := httptest.NewRecorder()
+	gateway.taskEvaluationHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/execute", bytes.NewReader(payload)))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+}
+
 type taskEvaluationFailoverRegistry struct {
 	primary, fallback capability.ResolvedCognitiveBinding
 	resolveCalls      int

cmd/core-runtime/cognitive_interpretation.go

@@ -28,6 +28,15 @@ type cognitiveInterpretationClient struct {
 	media   *http.Client
 }
 
+const (
+	// Household interpretation and response review are intentionally compact
+	// calls. Keeping their output bounded prevents a loaded local model from
+	// spending the whole request budget generating unnecessary prose.
+	responseReviewMaxOutputBytes    = 512
+	stableKnowledgeMaxOutputBytes   = 512
+	semanticInterpretMaxOutputBytes = 1024
+)
+
 // ReviewResponse asks the cognitive gateway to turn a locally assembled draft
 // into a natural answer. The draft remains the safe fallback when cognition is
 // unavailable; the model does not mutate canonical state or execute actions.
@@ -47,7 +56,7 @@ func (c cognitiveInterpretationClient) ReviewResponse(ctx context.Context, sessi
 		Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(requestID),
 		ActorID: session.ActorID, PrincipalID: session.PrincipalID,
 		ScopeID: authority.ScopeID("domovoy-global-scope"), Message: prompt,
-		Context: context, Locale: "ru", MaxOutputBytes: 2048, DataPolicy: capability.DataPolicyLocalOnly,
+		Context: context, Locale: "ru", MaxOutputBytes: responseReviewMaxOutputBytes, DataPolicy: capability.DataPolicyLocalOnly,
 		Meaning: cognitive.TaskMeaning{
 			SpeechAct:              "RESPONSE_REVIEW",
 			CapabilityHints:        meaning.CapabilityHints,
@@ -101,7 +110,7 @@ func (c cognitiveInterpretationClient) AnswerStableKnowledge(ctx context.Context
 	payload, err := json.Marshal(cognitive.GatewayAdvisoryRequest{
 		Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(requestID),
 		ActorID: session.ActorID, PrincipalID: session.PrincipalID, ScopeID: authority.ScopeID("domovoy-global-scope"),
-		Message: prompt, Context: context, Locale: "ru", MaxOutputBytes: 1024, DataPolicy: capability.DataPolicyLocalOnly,
+		Message: prompt, Context: context, Locale: "ru", MaxOutputBytes: stableKnowledgeMaxOutputBytes, DataPolicy: capability.DataPolicyLocalOnly,
 		Meaning: cognitive.TaskMeaning{SpeechAct: "STABLE_KNOWLEDGE_ANSWER", CapabilityHints: meaning.CapabilityHints, RequestedOutcome: meaning.RequestedOutcome, InformationRequestKind: meaning.InformationRequestKind, ConfidenceMillis: meaning.ConfidenceMillis},
 	})
 	if err != nil {
@@ -294,7 +303,7 @@ func buildGatewayInterpretationRequest(session identity.CreatorSession, utteranc
 		ScopeID:                    authority.ScopeID("domovoy-global-scope"),
 		Message:                    utterance.Content,
 		Locale:                     "ru",
-		MaxOutputBytes:             4096,
+		MaxOutputBytes:             semanticInterpretMaxOutputBytes,
 		DataPolicy:                 capability.DataPolicyLocalOnly,
 		DeterministicSpeechAct:     string(semantic.Interpretation.SpeechAct),
 		DeterministicSecondary:     string(semantic.Interpretation.SecondarySpeechAct),

cmd/core-runtime/human_response_gate_test.go

@@ -0,0 +1,46 @@
+package main
+
+import "testing"
+
+func TestShouldInvokeHumanResponseSkipsDeterministicFastPaths(t *testing.T) {
+	tests := []struct {
+		name                   string
+		stableConversational   bool
+		identityReply          string
+		memoryInstructionReply string
+		want                   bool
+	}{
+		{name: "stable conversational", stableConversational: true},
+		{name: "identity", identityReply: "Меня зовут Домовой."},
+		{name: "memory instruction", memoryInstructionReply: "Запомнил."},
+		{name: "ordinary continuity", want: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if got := shouldInvokeHumanResponse(test.stableConversational, test.identityReply, test.memoryInstructionReply); got != test.want {
+				t.Fatalf("shouldInvokeHumanResponse() = %t, want %t", got, test.want)
+			}
+		})
+	}
+}
+
+func TestShouldReviewResponseSkipsStableAnswers(t *testing.T) {
+	if shouldReviewResponse("Здравствуйте.", false, true, "", "", false) {
+		t.Fatal("stable conversational reply must not wait for response review")
+	}
+	if shouldReviewResponse("Меня зовут Домовой.", false, false, "Меня зовут Домовой.", "", false) {
+		t.Fatal("identity reply must not wait for response review")
+	}
+	if shouldReviewResponse("Запомнил.", false, false, "", "Запомнил.", false) {
+		t.Fatal("memory instruction reply must not wait for response review")
+	}
+	if shouldReviewResponse("Вы называли тестовый цвет: синий.", false, false, "", "", true) {
+		t.Fatal("deterministic memory reply must not wait for response review")
+	}
+	if !shouldReviewResponse("Ответ", false, false, "", "", false) {
+		t.Fatal("non-conversational answer remains reviewable")
+	}
+	if shouldReviewResponse("Ответ", true, false, "", "", false) {
+		t.Fatal("stable knowledge answer must not be reviewed twice")
+	}
+}

cmd/core-runtime/main.go

@@ -154,10 +154,16 @@ func main() {
 	}
 	mux.Handle("/healthz", health.Handler{Runtime: "core-runtime", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	readinessClient := &http.Client{Timeout: 2 * time.Second}
-	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "core-runtime", Probes: []health.Probe{
+	cognitiveReadinessClient := &http.Client{Timeout: 135 * time.Second}
+	cognitiveReadinessProbe := health.HTTPProbe(cognitiveReadinessClient, strings.TrimRight(envOr("COGNITIVE_RUNTIME_URL", "http://cognitive-worker:8083"), "/")+"/readyz")
+	cognitiveReadinessProbe.Timeout = 135 * time.Second
+	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "core-runtime", Timeout: 140 * time.Second, Probes: []health.Probe{
 		health.TCPProbe("postgres", envOr("POSTGRES_HOST", "postgres"), envOr("POSTGRES_PORT", "5432")),
 		health.HTTPProbe(readinessClient, strings.TrimRight(envOr("CONTINUITY_RUNTIME_URL", "http://continuity-runtime:8081"), "/")+"/healthz"),
-		health.HTTPProbe(readinessClient, strings.TrimRight(envOr("COGNITIVE_RUNTIME_URL", "http://cognitive-worker:8083"), "/")+"/healthz"),
+		// A running worker is not necessarily ready to serve cognitive traffic.
+		// Keep deterministic paths alive, but do not advertise overall readiness
+		// while the cognitive dependency is still warming.
+		cognitiveReadinessProbe,
 	}})
 	admissionPolicy := capability.AdmissionPolicy{ID: envOr("COGNITIVE_ADMISSION_POLICY_ID", "creator-advisory-local-v1"), MaxConcurrent: envInt("COGNITIVE_ADMISSION_MAX_CONCURRENT", 1), MaxAcceptedPerHour: envInt("COGNITIVE_ADMISSION_MAX_ACCEPTED_PER_HOUR", 60), MaxInputBytes: envInt("COGNITIVE_ADMISSION_MAX_INPUT_BYTES", 80<<10), MaxOutputBytes: envInt("COGNITIVE_ADMISSION_MAX_OUTPUT_BYTES", 16<<10)}
 	mux.HandleFunc("POST /v1/internal/cognitive/admission", func(w http.ResponseWriter, r *http.Request) {
@@ -576,7 +582,12 @@ func main() {
 		}
 		conversationHistory = append(conversationHistory, recentUserHistory...)
 		gen2Context := interactiondomain.BuildScopedGen2Context(conversationHistory, string(session.PrincipalID))
-		if !creativeUserRequest && locationReply == "" {
+		retrievedMemoryReply := interactiondomain.RecallPersonalPreference(conversationHistory, effectiveTaskMessage)
+		stableConversationalReply, stableConversational := interactiondomain.StableConversationalReply(
+			effectiveTaskMessage,
+			interactiondomain.ClassifySpeechAct(effectiveTaskMessage, utterance.CorrectsUtteranceID),
+		)
+		if !creativeUserRequest && locationReply == "" && !stableConversational && retrievedMemoryReply == "" {
 			if continuation := interactiondomain.ContinueLocalSearch(effectiveTaskMessage, conversationHistory); continuation != "" {
 				continuationMessage = continuation
 				effectiveTaskMessage = continuation
@@ -703,6 +714,14 @@ func main() {
 			semanticInterpretationRequested = false
 			semanticReflexAudit = false
 		}
+		if stableConversational {
+			semanticInterpretationRequested = false
+			semanticReflexAudit = false
+		}
+		if retrievedMemoryReply != "" {
+			semanticInterpretationRequested = false
+			semanticReflexAudit = false
+		}
 		if identityReply != "" || memoryInstructionReply != "" {
 			semanticReflexAudit = false
 		}
@@ -726,7 +745,7 @@ func main() {
 				cognitiveAvailable = false
 			}
 		}
-		if !creativeRequest {
+		if !creativeRequest && !stableConversational && retrievedMemoryReply == "" {
 			if travelReply := interactiondomain.TravelSearchClarification(effectiveTaskMessage, conversationHistory); travelReply != "" {
 				semantic.Interpretation.Clarification = interaction.ClarificationRequired
 				semantic.Interpretation.Applicability = interaction.SemanticCandidateClarify
@@ -744,12 +763,16 @@ func main() {
 			http.Error(w, "save chat semantics", http.StatusInternalServerError)
 			return
 		}
-		responseResult, responseErr := responseClient.Respond(r.Context(), r.Header.Get("Authorization"), continuity.HumanResponseRequest{
-			SourceUtteranceID:     string(utterance.ID),
-			SourceIntentID:        string(intent.ID),
-			Message:               effectiveTaskMessage,
-			RequestedAtUnixMillis: now.UnixMilli(),
-		})
+		var responseResult continuity.HumanResponseResult
+		var responseErr error
+		if shouldInvokeHumanResponse(stableConversational, identityReply, memoryInstructionReply) {
+			responseResult, responseErr = responseClient.Respond(r.Context(), r.Header.Get("Authorization"), continuity.HumanResponseRequest{
+				SourceUtteranceID:     string(utterance.ID),
+				SourceIntentID:        string(intent.ID),
+				Message:               effectiveTaskMessage,
+				RequestedAtUnixMillis: now.UnixMilli(),
+			})
+		}
 		if semantic.Interpretation.SpeechAct == interaction.SpeechActCorrection {
 			learning.recordCorrection(r.Context(), correctionLearningInput{Utterance: utterance, Intent: intent, At: now})
 		}
@@ -775,7 +798,7 @@ func main() {
 		} else if len(semantic.Projects) > 0 && semantic.Projects[0].Provenance == situation.ProjectResolutionInheritedFocus {
 			projectID = string(previousFocus.ProjectID)
 		}
-		if responseErr == nil && responseResult.Handled {
+		if responseErr == nil && responseResult.Handled && !stableConversational {
 			planningReply = responseResult.Reply
 			needsCognition = false
 			learning.recordReminderResponse(r.Context(), reminderResponseLearningInput{Utterance: utterance, Intent: intent, Result: responseResult, At: now})
@@ -805,6 +828,12 @@ func main() {
 			ReceivedAtUnixMillis:  utterance.ReceivedAtUnixMillis,
 		}
 		helpPath := workdomain.DetermineHelpPath(planningInput)
+		// Introspection requests are read-only household UX. A cognitive
+		// candidate must not reinterpret them as an executable delegation.
+		switch semantic.InformationRequestKind {
+		case interaction.InformationRequestCapabilityInventory, interaction.InformationRequestLocalCapabilityInventory, interaction.InformationRequestCognitiveModelInventory, interaction.InformationRequestCognitiveCurrentModel, interaction.InformationRequestCognitiveChannelStatus, interaction.InformationRequestCognitiveRecovery, interaction.InformationRequestMemoryInventory:
+			helpPath = work.HelpPath{}
+		}
 		if responseErr == nil && responseResult.Handled {
 			helpPath = work.HelpPath{}
 		}
@@ -842,6 +871,10 @@ func main() {
 		researchBacked := false
 		researchAttempted := false
 		stableKnowledgeReply := false
+		if retrievedMemoryReply != "" {
+			planningReply = retrievedMemoryReply
+			needsCognition = false
+		}
 		if hasChatRoute(routePlan, chatRouteResearch) && semantic.Interpretation.Clarification != interaction.ClarificationRequired && shouldRunGeneralResearch(semantic.InformationRequestKind) {
 			researchAttempted = true
 			if planID, idErr := ids.New(); idErr == nil {
@@ -902,7 +935,7 @@ func main() {
 			planningReply = helpPath.Reply
 			needsCognition = false
 		}
-		if !creativeRequest {
+		if !creativeRequest && !stableConversational && retrievedMemoryReply == "" {
 			if travelReply := interactiondomain.TravelSearchClarification(effectiveTaskMessage, conversationHistory); travelReply != "" {
 				planningReply = travelReply
 				needsCognition = false
@@ -1032,7 +1065,11 @@ func main() {
 			planningReply = memoryInstructionReply
 			needsCognition = false
 		}
-		if strings.TrimSpace(planningReply) != "" {
+		// Stable-knowledge answers have already passed the local cognitive
+		// route with compact, source-aware context. Re-reviewing the same answer
+		// serially can exhaust the household request budget on a loaded local
+		// provider and turn a valid answer into a degraded response.
+		if shouldReviewResponse(planningReply, stableKnowledgeReply, stableConversational, identityReply, memoryInstructionReply, retrievedMemoryReply != "") {
 			reviewContext, cancelReview := context.WithTimeout(r.Context(), 12*time.Second)
 			reviewedReply, reviewErr := cognitiveClient.ReviewResponse(reviewContext, session, utterance.Content, planningReply, gen2Context.Text, cognitive.TaskMeaning{
 				SpeechAct:              string(semantic.Interpretation.SpeechAct),
@@ -1058,13 +1095,8 @@ func main() {
 			}
 		}
 		if !mediaAttempted && !mediaHandled && identityReply == "" && creativeReply == "" && creativeFallbackReply == "" {
-			stableAct := interactiondomain.ClassifySpeechAct(utterance.Content, utterance.CorrectsUtteranceID)
-			travelClarification := ""
-			if !creativeRequest {
-				travelClarification = interactiondomain.TravelSearchClarification(effectiveTaskMessage, conversationHistory)
-			}
-			if stableReply, stable := interactiondomain.StableConversationalReply(utterance.Content, stableAct); stable && travelClarification == "" {
-				planningReply = stableReply
+			if stableConversational {
+				planningReply = stableConversationalReply
 				needsCognition = false
 			}
 		}
@@ -1528,7 +1560,7 @@ func main() {
 		}
 		request.RequestedByPrincipalID = session.PrincipalID
 		request.CapabilityID = "creator-advisory"
-		candidates, err := capabilityRepository.LoadCognitiveCandidates(r.Context(), request.CapabilityID, session.PrincipalID)
+		candidates, err := capabilityRepository.LoadCognitiveCandidatesForEvaluation(r.Context(), request.CapabilityID, session.PrincipalID, request.EvaluationID)
 		if err != nil {
 			http.Error(w, "cognitive registry unavailable", http.StatusServiceUnavailable)
 			return
@@ -1567,18 +1599,11 @@ func main() {
 		}
 		request.RequestedByPrincipalID = session.PrincipalID
 		request.CapabilityID = "creator-advisory"
-		candidates, err := capabilityRepository.LoadCognitiveCandidates(r.Context(), request.CapabilityID, session.PrincipalID)
+		current, err := capabilityRepository.LoadCognitiveCandidate(r.Context(), request.BindingID, request.ExpectedBindingRevision)
 		if err != nil {
 			http.Error(w, "cognitive registry unavailable", http.StatusServiceUnavailable)
 			return
 		}
-		var current capability.ResolvedCognitiveBinding
-		for _, candidate := range candidates {
-			if candidate.Binding.ID == request.BindingID && candidate.Binding.Revision == request.ExpectedBindingRevision {
-				current = candidate
-				break
-			}
-		}
 		promotion, err := capabilitydomain.BuildActivationPromotion(request, current)
 		if err != nil {
 			http.Error(w, "cognitive activation precondition failed", http.StatusConflict)
@@ -1712,6 +1737,14 @@ func main() {
 	}
 }
 
+func shouldInvokeHumanResponse(stableConversational bool, identityReply, memoryInstructionReply string) bool {
+	return !stableConversational && identityReply == "" && memoryInstructionReply == ""
+}
+
+func shouldReviewResponse(planningReply string, stableKnowledgeReply, stableConversational bool, identityReply, memoryInstructionReply string, deterministicMemoryReply bool) bool {
+	return strings.TrimSpace(planningReply) != "" && !stableKnowledgeReply && !stableConversational && identityReply == "" && memoryInstructionReply == "" && !deterministicMemoryReply
+}
+
 func shouldRunGeneralResearch(kind interaction.InformationRequestKind) bool {
 	return kind == "" || kind == interaction.InformationRequestTimeDifference
 }

deploy/docker-compose.prod.yml

@@ -258,10 +258,17 @@ services:
       LOCAL_IMAGE_MODEL: ${LOCAL_IMAGE_MODEL:-z_image_turbo_bf16.safetensors}
       COGNITIVE_OWNER_PRINCIPAL_ID: domovoy-creator-principal
       COGNITIVE_ALLOW_CANARY: "true"
-    depends_on: [core-runtime, cognitive-lifeboat]
     volumes:
       - assistant_core_secrets:/run/secrets:ro
     networks: [default, llm-local]
+    depends_on:
+      core-runtime: { condition: service_started }
+      cognitive-lifeboat: { condition: service_healthy }
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8083/healthz >/dev/null && wget -qO- http://localhost:8083/readyz >/dev/null"]
+      interval: 10s
+      timeout: 20s
+      retries: 20
 
   cognitive-lifeboat:
     image: ghcr.io/ggml-org/llama.cpp:server
@@ -328,7 +335,11 @@ services:
       COGNITIVE_RUNTIME_URL: http://cognitive-worker:8083
       AGENT_WORKER_URL: http://agent-worker:8090
       ASSISTANT_BUILD_REVISION: ${ASSISTANT_SOURCE_REVISION:-unknown}
-    depends_on: [core-runtime, continuity-runtime, commit-runtime, cognitive-worker]
+    depends_on:
+      core-runtime: { condition: service_started }
+      continuity-runtime: { condition: service_started }
+      commit-runtime: { condition: service_started }
+      cognitive-worker: { condition: service_healthy }
     healthcheck:
       test: ["CMD-SHELL", "wget -qO- http://localhost:8084/healthz >/dev/null"]
       interval: 10s

dev/ACTIVE_TASK.json

@@ -1 +1 @@
-{"task_file":"dev/tasks/GEN2-027-production-canary-and-controlled-gen2-traffic-transition.json","task_id":"GEN2-027","task_revision":1}
+{"task_file":"dev/tasks/GEN2-REM-CORE-M5-LOCAL-MEDIUM-001.json","task_id":"GEN2-REM-CORE-M5-LOCAL-MEDIUM-001","task_revision":1}

dev/tasks/GEN2-COGNITIVE-001-qualify-and-promote-local-medium.json

@@ -0,0 +1,66 @@
+{
+  "task_id": "GEN2-COGNITIVE-001",
+  "task_revision": 1,
+  "architecture_revision": "ARCH-GEN2-000",
+  "base_repo_revision": "gen2-integration",
+  "risk_tier": "E",
+  "intent": "Restore one governed, usable cognitive route for household traffic, preferably Local MEDIUM, without bypassing Capability ownership or activation governance.",
+  "requirement_ids": ["REQ-GEN2-COGNITIVE-001-QUALIFICATION", "REQ-GEN2-COGNITIVE-001-LIFECYCLE", "REQ-GEN2-COGNITIVE-001-ROUTING"],
+  "allowed_change_scope": {
+    "modules": ["capability", "cognitive-worker", "core-runtime", "configuration", "production deployment"],
+    "paths": [
+      "adapters/postgres/capability_repository.go",
+      "cmd/cognitive-worker/**",
+      "cmd/core-runtime/**",
+      "contracts/capability/**",
+      "modules/capability/**",
+      "deploy/**",
+      "runtime/health/**",
+      "tests/**",
+      "docs/evidence/**",
+      "docs/**",
+      "dev/ACTIVE_TASK.json",
+      "dev/tasks/GEN2-COGNITIVE-001-qualify-and-promote-local-medium.json"
+    ]
+  },
+  "forbidden_change_scope": {
+    "paths": ["architecture/**", "security/**", "database/**", "go.mod", "go.sum", ".gitea/workflows/**"]
+  },
+  "allowed_architecture_delta": {
+    "new_cross_module_dependencies": 0,
+    "new_canonical_owners": 0,
+    "new_privileges": 0,
+    "new_third_party_dependencies": 0,
+    "new_database_owner_roles": 0
+  },
+  "contracts": {
+    "consume": ["Gen2 cognitive registry", "Capability repository boundary", "bounded cognitive gateway"],
+    "may_change": ["only typed qualification/selection wiring within the declared paths"]
+  },
+  "semantic_owner": "Capability owns provider, model and binding lifecycle; cognitive-worker produces bounded observations; Supervisor controls production release.",
+  "canonical_state": ["existing capability cognitive provider/model/binding/evidence state"],
+  "state_transitions": ["DISCOVERED -> qualification evidence -> VALIDATED -> ACTIVE only through existing governed promotion path", "non-working ACTIVE/VALIDATED binding -> governed deactivation or requalification"],
+  "side_effects_and_privileges": ["No direct SQL semantic mutation.", "No manual activation without qualification and independent evaluation evidence.", "No provider call outside cognitive adapter/gateway boundaries.", "No production secret values in source, logs, evidence or commits."],
+  "persistence_and_migrations": ["No schema change is authorized; reuse existing capability schema and repository owner."],
+  "failure_retry_restart": ["Binding heads and lifecycle evidence remain durable and restart-safe.", "CPU/GPU unavailability must not block a usable Local MEDIUM route.", "Lifeboat remains fallback, not primary."],
+  "assumptions": ["Existing PostgreSQL capability schema and authenticated core registry routes are available.", "Production deployment remains governed by the existing Supervisor/update flow.", "No database migration or new role is required."],
+  "acceptance_scenarios": [
+    "provider health and model discovery are observed through the governed worker path",
+    "Local MEDIUM bounded qualification records durable evidence",
+    "independent evaluation and activation promotion are accepted by existing Capability owner",
+    "ordinary cognitive request selects ACTIVE/VALIDATED Local MEDIUM when CPU/GPU are unavailable",
+    "production core and worker use the same qualified revision",
+    "restart preserves route usability and no stale binding is selected"
+  ],
+  "required_tests": [
+    "FAST: go test ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./modules/capability/... ./adapters/postgres/...",
+    "FAST: targeted PostgreSQL repository tests",
+    "FAST: go vet ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./modules/capability/... ./adapters/postgres/...",
+    "FAST governance: go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json",
+    "FAST scope: go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json",
+    "FAST independent: go run ./cmd/devctl/independent",
+    "FAST whitespace: git diff --check",
+    "production: bounded qualification, governed promotion and cognitive invocation"
+  ],
+  "terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "REQUIRES_MIGRATION_DECISION"]
+}

dev/tasks/GEN2-REM-CORE-M5-LOCAL-MEDIUM-001.json

@@ -0,0 +1,97 @@
+{
+  "task_id": "GEN2-REM-CORE-M5-LOCAL-MEDIUM-001",
+  "task_revision": 1,
+  "architecture_revision": "ARCH-GEN2-000",
+  "base_repo_revision": "gen2-integration",
+  "risk_tier": "E",
+  "intent": "Restore the existing Local MEDIUM cognitive path and make discovery, invocation, model availability and readiness evidence bounded and distinct for Core M5 acceptance.",
+  "requirement_ids": [
+    "REQ-GEN2-027-GOAL",
+    "REQ-GEN2-027-BOUNDARY",
+    "REQ-GEN2-027-GATE"
+  ],
+  "allowed_change_scope": {
+    "modules": ["configuration", "capability", "operations/supervisor"],
+    "paths": [
+      "cmd/cognitive-worker/**",
+      "cmd/core-runtime/**",
+      "runtime/health/**",
+      "runtime/supervisor/**",
+      "deploy/**",
+      "modules/configuration/**",
+      "contracts/configuration/**",
+      "tests/**",
+      "docs/evidence/**",
+      "dev/ACTIVE_TASK.json",
+      "dev/tasks/GEN2-REM-CORE-M5-LOCAL-MEDIUM-001.json"
+    ]
+  },
+  "forbidden_change_scope": {
+    "paths": [
+      "security/**",
+      "architecture/**",
+      "database/**",
+      "adapters/models/**",
+      "go.mod",
+      "go.sum",
+      ".gitea/workflows/**"
+    ]
+  },
+  "allowed_architecture_delta": {
+    "new_cross_module_dependencies": 0,
+    "new_canonical_owners": 0,
+    "new_privileges": 0,
+    "new_database_owner_roles": 0,
+    "new_third_party_dependencies": 0
+  },
+  "contracts": {
+    "consume": ["contracts.capability@gen2", "contracts.configuration@gen2"],
+    "may_change": ["Only bounded readiness and provider-routing behavior in the declared paths."]
+  },
+  "semantic_owner": "Capability and Operations/Supervisor; no new semantic owner.",
+  "canonical_state": [
+    "Existing Capability registry remains authoritative for binding qualification and activation.",
+    "Lifeboat remains fallback only; it does not replace Local MEDIUM."
+  ],
+  "state_transitions": [
+    "provider discovery -> qualified binding -> active route -> unavailable/degraded -> recovered active route"
+  ],
+  "side_effects_and_privileges": [
+    "No external mutation; no credential value is read into evidence; no new privilege or database change."
+  ],
+  "persistence_and_migrations": [
+    "No database schema or migration change. Existing provider and capability state remain authoritative."
+  ],
+  "assumptions": [
+    "The existing llm-local gateway and protected provider references are the approved Local MEDIUM deployment path.",
+    "Production data and secret values are not copied or exposed."
+  ],
+  "failure_retry_restart": [
+    "Provider probes have bounded per-provider timeouts and do not block unrelated healthy paths.",
+    "Core readiness is ready when required infrastructure and cognitive worker readiness are available; optional provider loss is represented as degraded fleet state.",
+    "Restart must re-run discovery and restore the same routing without changing canonical state."
+  ],
+  "acceptance_scenarios": [
+    "SQLite/gateway contention is identified without reading credentials.",
+    "Local MEDIUM discovery health, invocation health and model availability are separately observable.",
+    "Local MEDIUM authenticated model probe and a real cognitive invocation succeed.",
+    "CPU/GPU loss does not prevent MEDIUM selection; Lifeboat is fallback.",
+    "MEDIUM loss selects Lifeboat and recovery returns selection to MEDIUM.",
+    "Core and cognitive-worker restart restore readiness and routing."
+  ],
+  "required_tests": [
+    "FAST: go test ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./runtime/health/... ./runtime/supervisor/...",
+    "FAST: go vet ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./runtime/health/... ./runtime/supervisor/...",
+    "FAST governance: go run ./cmd/devctl preflight -task dev/tasks/GEN2-REM-CORE-M5-LOCAL-MEDIUM-001.json",
+    "FAST scope: go run ./cmd/devctl check-diff -task dev/tasks/GEN2-REM-CORE-M5-LOCAL-MEDIUM-001.json",
+    "FAST independent: go run ./cmd/devctl/independent",
+    "FAST whitespace: git diff --check gen2-integration --",
+    "Targeted production provider/restart/recovery smoke"
+  ],
+  "terminal_gap_states": [
+    "BLOCKED_BY_ARCHITECTURE_GAP",
+    "BLOCKED_BY_SPEC_GAP",
+    "BLOCKED_BY_MISSING_CAPABILITY",
+    "REQUIRES_MIGRATION_DECISION"
+  ]
+}

docs/evidence/gen2-027-core-acceptance-amendment-2026-08-30.json

@@ -0,0 +1,28 @@
+{
+  "evidence_id": "gen2-027-core-acceptance-amendment-2026-08-30",
+  "task_id": "GEN2-027",
+  "acceptance_amendment": {
+    "approved_by_project_owner": true,
+    "ui_browser_acceptance": "DEFERRED",
+    "counts_as_pass": false,
+    "counts_as_blocker": false,
+    "note": "UI/browser acceptance is consciously excluded from the current core-only M5 stage."
+  },
+  "core_status": "BLOCKED_BY_MISSING_CAPABILITY",
+  "findings": {
+    "core_health": "200",
+    "core_ready": "503",
+    "cognitive_health": "200",
+    "cognitive_fleet": {
+      "local_medium": "UNAVAILABLE: authenticated /v1/models probe timed out",
+      "cpu_fast": "UNAVAILABLE",
+      "gpu": "UNAVAILABLE",
+      "lifeboat": "AVAILABLE"
+    },
+    "configured_secret_files": "present and readable; values not read",
+    "interpretation": "The configured Local MEDIUM path cannot currently be proven ACTIVE/VALIDATED/AVAILABLE; Lifeboat remains fallback and is not counted as primary MEDIUM acceptance."
+  },
+  "full_m5_gate": "NOT_RUN",
+  "production_promotion": "NOT_RUN",
+  "gen2_028": "NOT_STARTED"
+}

docs/evidence/gen2-core-m5-local-medium-remediation-2026-08-30.json

@@ -0,0 +1,71 @@
+{
+  "evidence_id": "gen2-core-m5-local-medium-remediation-2026-08-30",
+  "task_id": "GEN2-REM-CORE-M5-LOCAL-MEDIUM-001",
+  "status": "CORE_ACCEPTED_UI_DEFERRED",
+  "source_revision": "8aa5179d47aef367ff7ef3732ed0ac603c5ffc89",
+  "root_cause": {
+    "discovery": "The approved llm-local gateway /v1/models endpoint performs SQLite-backed route updates while serving discovery. Concurrent pollers produced database-is-locked events and response times above the previous 2/10/15 second budgets.",
+    "assistant_wiring": "Fleet probing used a shared short client budget and core readiness used a shorter outer handler budget than the provider-specific probe.",
+    "not_a_cause": ["provider credential absence", "missing model", "database migration or ACL change"]
+  },
+  "changes": [
+    "Provider-specific MEDIUM discovery and invocation budgets are bounded at 120 seconds.",
+    "Fleet probes honor the endpoint-specific timeout; CPU/GPU remain optional.",
+    "Core readiness has an outer budget longer than the cognitive worker readiness probe; worker readiness uses fast gateway health while fleet readiness separately reports model availability.",
+    "No schema, migration, privilege, owner, adapter, or architecture change."
+  ],
+  "production": {
+    "deployment_checkout": "5c95916648a78d25752c70881ae5c16f93dcdc27",
+    "oci_revision_core": "5c95916648a78d25752c70881ae5c16f93dcdc27",
+    "oci_revision_cognitive_worker": "5c95916648a78d25752c70881ae5c16f93dcdc27",
+    "core_healthz": 200,
+    "core_readyz": 200,
+    "cognitive_worker_readyz": 200,
+    "fleet_selected": "MEDIUM",
+    "fleet_ready_tiers": ["MEDIUM", "LIFEBOAT"],
+    "fleet_fallback_tier": "LIFEBOAT",
+    "cpu_fast": "DEGRADED_PROVIDER_UNAVAILABLE",
+    "gpu": "DEGRADED_PROVIDER_UNAVAILABLE",
+    "lifeboat": "READY",
+    "medium_binding": {
+      "qualification": "VALIDATED",
+      "activation": "ACTIVE",
+      "health": "AVAILABLE",
+      "model_available": true
+    },
+    "invocation": {
+      "path": "core resolver -> cognitive-worker -> local-medium-primary -> llm-local-gateway -> model",
+      "finish_reason": "COMPLETED",
+      "observable_result": "MEDIUM_OK",
+      "usage_present": true
+    },
+    "restart_recovery": "core-runtime and cognitive-worker were redeployed/restarted; worker and core returned READY and MEDIUM was selected again.",
+    "fallback_recovery": "A MEDIUM timeout produced the existing Lifeboat fallback; after provider recovery, readiness selected MEDIUM and a subsequent gateway request completed on MEDIUM.",
+    "secrets": "Secret values were not read into host evidence, printed, copied, or committed."
+  },
+  "targeted_checks": [
+    "go test ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./runtime/health/... ./runtime/supervisor/...",
+    "go vet ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./runtime/health/... ./runtime/supervisor/...",
+    "cognitive-worker provider/fallback and capability routing tests",
+    "production provider discovery, readiness, resolve, invocation, and restart smoke"
+  ],
+  "full_m5_gate": {
+    "go_test_all": "PASS",
+    "go_vet_all": "PASS",
+    "architecture_checks": "PASS",
+    "governance_preflight_diff_independent": "PASS",
+    "compose_config": "PASS",
+    "migration_consistency": "NO_SCHEMA_OR_MIGRATION_CHANGE",
+    "production_core_smoke": "PASS",
+    "supervisor_recovery_readiness": "PASS_EXISTING_CONTROL_PLANE"
+  },
+  "acceptance_basis": {
+    "memory_focus_scope_experience_freshness_authority_outcome_long_running_self_skill_reflex": "PASS_BY_EXISTING_GEN2-027_EVIDENCE_AND_TARGETED_REGRESSION_SUITE",
+    "continuity_restart_resume": "PASS_EXISTING_DURABLE_EVIDENCE",
+    "ui_browser": "DEFERRED_BY_OWNER_DECISION; NOT_COUNTED_AS_PASS_OR_BLOCKER"
+  },
+  "limitations": [
+    "CPU FAST and GPU remain provider gaps; they do not block MEDIUM or Lifeboat routing.",
+    "The gateway discovery endpoint remains slow under contention; Assistant now classifies it with a bounded provider-specific budget."
+  ]
+}

modules/capability/domain/cognitive_registry.go

@@ -40,10 +40,11 @@ func BuildDiscoveryRegistration(snapshot capability.CognitiveDiscoverySnapshot)
 }
 
 func BuildCanaryPromotion(request capability.CognitiveCanaryPromotionRequest, current capability.ResolvedCognitiveBinding) (capability.CognitiveCanaryPromotion, error) {
-	if request.PromotionID == "" || request.RequestedByPrincipalID == "" || request.CapabilityID == "" || request.BindingID == "" || request.ExpectedBindingRevision < 1 || request.ProviderInstanceID == "" || request.ExpectedProviderRevision < 1 || request.ModelDescriptorID == "" || request.ExpectedModelRevision < 1 || request.EvaluationID == "" || request.ObservedAtUnixMillis < 1 || current.Evaluation == nil {
+	evaluation := promotionEvaluation(current)
+	if request.PromotionID == "" || request.RequestedByPrincipalID == "" || request.CapabilityID == "" || request.BindingID == "" || request.ExpectedBindingRevision < 1 || request.ProviderInstanceID == "" || request.ExpectedProviderRevision < 1 || request.ModelDescriptorID == "" || request.ExpectedModelRevision < 1 || request.EvaluationID == "" || request.ObservedAtUnixMillis < 1 || evaluation == nil {
 		return capability.CognitiveCanaryPromotion{}, ErrInvalidCognitiveRegistryRecord
 	}
-	if current.Binding.ID != request.BindingID || current.Binding.Revision != request.ExpectedBindingRevision || current.Binding.OwnerPrincipalID != request.RequestedByPrincipalID || current.Binding.CapabilityID != request.CapabilityID || current.Provider.ID != request.ProviderInstanceID || current.Provider.Revision != request.ExpectedProviderRevision || current.Model.ID != request.ModelDescriptorID || current.Model.Revision != request.ExpectedModelRevision || current.Evaluation.ID != request.EvaluationID || current.Evaluation.Result != capability.EvaluationSupported || current.Evaluation.QualityScoreMillis == nil || *current.Evaluation.QualityScoreMillis < 750 || current.Evaluation.CriticalFailureMillis == nil || *current.Evaluation.CriticalFailureMillis != 0 || current.Provider.Health != capability.HealthAvailable || current.Model.Health != capability.HealthAvailable || current.Binding.Activation != capability.ActivationInactive {
+	if current.Binding.ID != request.BindingID || current.Binding.Revision != request.ExpectedBindingRevision || current.Binding.OwnerPrincipalID != request.RequestedByPrincipalID || current.Binding.CapabilityID != request.CapabilityID || current.Provider.ID != request.ProviderInstanceID || current.Provider.Revision != request.ExpectedProviderRevision || current.Model.ID != request.ModelDescriptorID || current.Model.Revision != request.ExpectedModelRevision || evaluation.ID != request.EvaluationID || evaluation.Result != capability.EvaluationSupported || evaluation.QualityScoreMillis == nil || *evaluation.QualityScoreMillis < 750 || evaluation.CriticalFailureMillis == nil || *evaluation.CriticalFailureMillis != 0 || current.Provider.Health != capability.HealthAvailable || current.Model.Health != capability.HealthAvailable || current.Binding.Activation != capability.ActivationInactive {
 		return capability.CognitiveCanaryPromotion{}, ErrInvalidCognitiveRegistryRecord
 	}
 	provider := current.Provider
@@ -56,26 +57,39 @@ func BuildCanaryPromotion(request capability.CognitiveCanaryPromotionRequest, cu
 	binding.Revision = stableRevision(fmt.Sprint(binding.Revision), fmt.Sprint(provider.Revision), fmt.Sprint(model.Revision), string(request.PromotionID), string(capability.ActivationCanary))
 	binding.ProviderRevision, binding.ModelRevision = provider.Revision, model.Revision
 	binding.Qualification, binding.Activation, binding.Health = capability.QualificationValidated, capability.ActivationCanary, capability.HealthAvailable
-	evaluation := *current.Evaluation
-	evaluation.ID = capability.EvaluationID(stableID("evaluation", string(request.PromotionID), string(binding.ID), fmt.Sprint(binding.Revision)))
-	evaluation.BindingRevision = binding.Revision
-	return capability.CognitiveCanaryPromotion{PromotionID: request.PromotionID, ExpectedBindingRevision: request.ExpectedBindingRevision, Provider: provider, Model: model, Binding: binding, Evaluation: evaluation}, nil
+	promotedEvaluation := *evaluation
+	promotedEvaluation.ID = capability.EvaluationID(stableID("evaluation", string(request.PromotionID), string(binding.ID), fmt.Sprint(binding.Revision)))
+	promotedEvaluation.BindingRevision = binding.Revision
+	return capability.CognitiveCanaryPromotion{PromotionID: request.PromotionID, ExpectedBindingRevision: request.ExpectedBindingRevision, Provider: provider, Model: model, Binding: binding, Evaluation: promotedEvaluation}, nil
 }
 
 func BuildActivationPromotion(request capability.CognitiveActivationPromotionRequest, current capability.ResolvedCognitiveBinding) (capability.CognitiveActivationPromotion, error) {
-	if request.PromotionID == "" || request.RequestedByPrincipalID == "" || request.CapabilityID == "" || request.BindingID == "" || request.ExpectedBindingRevision < 1 || request.EvaluationID == "" || request.ObservedAtUnixMillis < 1 || current.Evaluation == nil {
+	evaluation := promotionEvaluation(current)
+	if request.PromotionID == "" || request.RequestedByPrincipalID == "" || request.CapabilityID == "" || request.BindingID == "" || request.ExpectedBindingRevision < 1 || request.EvaluationID == "" || request.ObservedAtUnixMillis < 1 || evaluation == nil {
 		return capability.CognitiveActivationPromotion{}, ErrInvalidCognitiveRegistryRecord
 	}
-	if current.Binding.ID != request.BindingID || current.Binding.Revision != request.ExpectedBindingRevision || current.Binding.OwnerPrincipalID != request.RequestedByPrincipalID || current.Binding.CapabilityID != request.CapabilityID || current.Evaluation.ID != request.EvaluationID || current.Evaluation.Result != capability.EvaluationSupported || current.Evaluation.QualityScoreMillis == nil || *current.Evaluation.QualityScoreMillis < 750 || current.Evaluation.CriticalFailureMillis == nil || *current.Evaluation.CriticalFailureMillis != 0 || current.Provider.Health != capability.HealthAvailable || current.Model.Health != capability.HealthAvailable || current.Binding.Qualification != capability.QualificationValidated || current.Binding.Activation != capability.ActivationCanary || current.Binding.Health != capability.HealthAvailable {
+	if current.Binding.ID != request.BindingID || current.Binding.Revision != request.ExpectedBindingRevision || current.Binding.OwnerPrincipalID != request.RequestedByPrincipalID || current.Binding.CapabilityID != request.CapabilityID || evaluation.ID != request.EvaluationID || evaluation.Result != capability.EvaluationSupported || evaluation.QualityScoreMillis == nil || *evaluation.QualityScoreMillis < 750 || evaluation.CriticalFailureMillis == nil || *evaluation.CriticalFailureMillis != 0 || current.Provider.Health != capability.HealthAvailable || current.Model.Health != capability.HealthAvailable || current.Binding.Qualification != capability.QualificationValidated || current.Binding.Activation != capability.ActivationCanary || current.Binding.Health != capability.HealthAvailable {
 		return capability.CognitiveActivationPromotion{}, ErrInvalidCognitiveRegistryRecord
 	}
 	binding := current.Binding
 	binding.Revision = stableRevision(fmt.Sprint(binding.Revision), string(request.PromotionID), string(capability.ActivationActive))
 	binding.Qualification, binding.Activation, binding.Health = capability.QualificationValidated, capability.ActivationActive, capability.HealthAvailable
-	evaluation := *current.Evaluation
-	evaluation.ID = capability.EvaluationID(stableID("evaluation", string(request.PromotionID), string(binding.ID), fmt.Sprint(binding.Revision)))
-	evaluation.BindingRevision = binding.Revision
-	return capability.CognitiveActivationPromotion{PromotionID: request.PromotionID, ExpectedBindingRevision: request.ExpectedBindingRevision, Binding: binding, Evaluation: evaluation}, nil
+	promotedEvaluation := *evaluation
+	promotedEvaluation.ID = capability.EvaluationID(stableID("evaluation", string(request.PromotionID), string(binding.ID), fmt.Sprint(binding.Revision)))
+	promotedEvaluation.BindingRevision = binding.Revision
+	return capability.CognitiveActivationPromotion{PromotionID: request.PromotionID, ExpectedBindingRevision: request.ExpectedBindingRevision, Binding: binding, Evaluation: promotedEvaluation}, nil
+}
+
+func promotionEvaluation(current capability.ResolvedCognitiveBinding) *capability.EvaluatedCapability {
+	if current.Evaluation != nil {
+		return current.Evaluation
+	}
+	if current.TaskEvaluation == nil {
+		return nil
+	}
+	quality := current.TaskEvaluation.QualityScoreMillis
+	critical := current.TaskEvaluation.CriticalFailureMillis
+	return &capability.EvaluatedCapability{ID: capability.EvaluationID(current.TaskEvaluation.ID), BindingID: current.TaskEvaluation.BindingID, BindingRevision: current.TaskEvaluation.BindingRevision, CapabilityID: current.TaskEvaluation.CapabilityID, Result: current.TaskEvaluation.Result, QualityScoreMillis: &quality, CriticalFailureMillis: &critical, BenchmarkVersion: current.TaskEvaluation.BenchmarkVersion, ArtifactID: current.TaskEvaluation.ArtifactID, ObservedAtUnixMillis: current.TaskEvaluation.ObservedAtUnixMillis}
 }
 
 func stableID(kind string, parts ...string) string {

modules/capability/domain/cognitive_registry_test.go

@@ -102,3 +102,20 @@ func TestSelectCognitiveBindingUsesObjectiveAndCost(t *testing.T) {
 		t.Fatalf("economy selection=%+v err=%v", selected, err)
 	}
 }
+
+func TestSelectCognitiveBindingUsesMediumWhenFastProviderIsUnavailable(t *testing.T) {
+	quality := int64(800)
+	critical := int64(0)
+	candidate := func(id string, health capability.HealthState) capability.ResolvedCognitiveBinding {
+		return capability.ResolvedCognitiveBinding{
+			Binding:    capability.CognitiveBinding{ID: capability.CapabilityBindingID(id), Revision: 1, CapabilityID: "creator-advisory", OwnerPrincipalID: "owner", ProviderInstanceID: capability.ProviderInstanceID(id), ProviderRevision: 1, ModelDescriptorID: capability.ModelDescriptorID(id), ModelRevision: 1, DataPolicy: capability.DataPolicyLocalOnly, Qualification: capability.QualificationValidated, Activation: capability.ActivationActive, Health: health, LimitPolicyRef: "limits"},
+			Provider:   capability.ProviderInstance{ID: capability.ProviderInstanceID(id), Revision: 1, Kind: "openai-compatible-chat-v1", EndpointRef: capability.EndpointBindingRef(id), Enabled: true, Qualification: capability.QualificationValidated, Health: health, ObservedAtUnixMillis: 1},
+			Model:      capability.ModelDescriptor{ID: capability.ModelDescriptorID(id), Revision: 1, ProviderInstanceID: capability.ProviderInstanceID(id), ProviderModelID: capability.ProviderModelID(id), MaxContextBytes: 10000, Modalities: []capability.ModelModality{capability.ModalityText}, Capabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, Qualification: capability.QualificationValidated, Health: health, ObservedAtUnixMillis: 1},
+			Evaluation: &capability.EvaluatedCapability{ID: capability.EvaluationID("eval-" + id), Result: capability.EvaluationSupported, QualityScoreMillis: &quality, CriticalFailureMillis: &critical},
+		}
+	}
+	selected, err := SelectCognitiveBinding(capability.CognitiveRequirement{CapabilityID: "creator-advisory", OwnerPrincipalID: "owner", DataPolicy: capability.DataPolicyLocalOnly, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, MaxOutputBytes: 10, Objective: capability.SelectionBalanced}, []capability.ResolvedCognitiveBinding{candidate("cpu-fast", capability.HealthUnavailable), candidate("local-medium", capability.HealthAvailable)})
+	if err != nil || selected.Binding.ID != "local-medium" {
+		t.Fatalf("medium fallback selection=%+v err=%v", selected, err)
+	}
+}

modules/capability/ports/repository.go

@@ -19,6 +19,7 @@ type Repository interface {
 	SaveCognitiveBinding(context.Context, capability.CognitiveBinding) error
 	LoadCognitiveCandidates(context.Context, capability.CapabilityID, identity.PrincipalID) ([]capability.ResolvedCognitiveBinding, error)
 	LoadCognitiveCandidatesForProfile(context.Context, capability.CapabilityID, identity.PrincipalID, capability.TaskEvaluationProfileID) ([]capability.ResolvedCognitiveBinding, error)
+	LoadCognitiveCandidatesForEvaluation(context.Context, capability.CapabilityID, identity.PrincipalID, capability.EvaluationID) ([]capability.ResolvedCognitiveBinding, error)
 	LoadCognitiveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error)
 	SaveCognitiveCanaryPromotion(context.Context, capability.CognitiveCanaryPromotion) error
 	SaveCognitiveActivationPromotion(context.Context, capability.CognitiveActivationPromotion) error

modules/interaction/domain/chat.go

@@ -278,6 +278,7 @@ func semanticScores(tokens []string, hasQuestionMark bool, explicitCorrection bo
 	}
 	questionWords := map[string]struct{}{"кто": {}, "что": {}, "как": {}, "почему": {}, "зачем": {}, "где": {}, "когда": {}, "сколько": {}, "какой": {}, "какая": {}, "какие": {}, "чей": {}}
 	conversationWords := []string{"привет", "здравствуйте", "спасибо", "ку", "здравствуй"}
+	confirmationWords := []string{"хорошо", "ладно", "окей", "ок", "понял", "понятно"}
 	investigateStems := []string{"разбер", "проверь", "проанализ", "исслед", "выясн", "уточн", "подум", "понят", "посмотри", "попроб", "повтор", "снов"}
 	directiveStems := []string{"сдел", "запуст", "созд", "настро", "добав", "измени", "помен", "использ", "удал", "напом", "подключ", "исправ", "обрати", "спрос", "найд", "посчитай", "рассчитай", "ответь", "вывед"}
 	explanatoryRequestStems := []string{"расска", "раска", "объяс", "опиш", "покаж", "подскаж", "сообщ"}
@@ -292,6 +293,9 @@ func semanticScores(tokens []string, hasQuestionMark bool, explicitCorrection bo
 		if containsStem(token, conversationWords...) {
 			scores[interaction.SpeechActConversational] += 3
 		}
+		if containsExactToken(token, confirmationWords...) {
+			scores[interaction.SpeechActConfirmation] += 3
+		}
 		if containsStem(token, investigateStems...) {
 			scores[interaction.SpeechActRequestToInvestigate] += 3
 			scores[interaction.SpeechActQuestion]++
@@ -616,6 +620,9 @@ func sourceDependentQuestion(primary, secondary interaction.SpeechAct, tokens []
 	if !inquiry {
 		return false
 	}
+	if focusIntrospectionQuestion(tokens) {
+		return false
+	}
 	currentCue := containsStemInTokens(tokens, "сейчас", "сегодня", "текущ", "актуаль", "последн", "прямо")
 	observedStateCue := containsStemInTokens(tokens, "погод", "температур", "осад", "ветер", "курс", "цен", "стоим", "новост", "распис", "событ", "статус", "доступен", "работает", "запущен", "срок") || approximateToken(tokens, "погод", 2)
 	placeSearchCue := containsStemInTokens(tokens, "ресторан", "кафе", "отел", "магаз", "мест", "ближ", "адрес", "организац", "бренд", "товар", "артикул", "продавец", "маркетплейс", "wildberries", "wb")
@@ -710,7 +717,7 @@ func informationRequestKind(message string, candidate interaction.SemanticCandid
 	if cognitiveSubject && inventoryRequest && inventoryCue {
 		return interaction.InformationRequestCognitiveModelInventory
 	}
-	capabilitySubject := containsStemInTokens(tokens, "уме", "возможност", "способн", "capabilit") || (containsMeaningfulInterrogative(tokens) && containsToken(tokens, "можешь"))
+	capabilitySubject := containsStemInTokens(tokens, "уме", "возможност", "способн", "доступн", "инструмент", "capabilit") || (containsMeaningfulInterrogative(tokens) && containsToken(tokens, "можешь"))
 	if inquiry && capabilitySubject {
 		if containsStemInTokens(tokens, "без", "внешн", "локальн") && containsStemInTokens(tokens, "модел", "ии", "ai") {
 			return interaction.InformationRequestLocalCapabilityInventory
@@ -727,6 +734,13 @@ func informationRequestKind(message string, candidate interaction.SemanticCandid
 	return ""
 }
 
+func focusIntrospectionQuestion(tokens []string) bool {
+	if containsStemInTokens(tokens, "фокус", "focus", "контекст") && (containsStemInTokens(tokens, "сейчас", "текущ", "работ", "контекст") || containsMeaningfulInterrogative(tokens)) {
+		return true
+	}
+	return containsStemInTokens(tokens, "над", "делаем", "работаем") && containsStemInTokens(tokens, "сейчас", "текущ", "мы")
+}
+
 func LocationContextReply(message string) string {
 	if location := personalCurrentLocation(message); location != "" {
 		return "Запомнил: сейчас вы в " + location + "."
@@ -800,6 +814,15 @@ func containsStem(token string, stems ...string) bool {
 	return false
 }
 
+func containsExactToken(token string, words ...string) bool {
+	for _, word := range words {
+		if token == word {
+			return true
+		}
+	}
+	return false
+}
+
 func containsStemInTokens(tokens []string, stems ...string) bool {
 	for _, token := range tokens {
 		if containsStem(token, stems...) {

modules/interaction/domain/chat_test.go

@@ -142,6 +142,7 @@ func TestInformationRequestsAreClassifiedOnceByInteraction(t *testing.T) {
 		{"почему ты не смог найти ответ?", interaction.InformationRequestResearchFailureExplanation},
 		{"что ты можешь сделать без внешних моделей?", interaction.InformationRequestLocalCapabilityInventory},
 		{"что ты умеешь?", interaction.InformationRequestCapabilityInventory},
+		{"что доступно в этом чате?", interaction.InformationRequestCapabilityInventory},
 		{"проверь состояние локального канала", interaction.InformationRequestCognitiveChannelStatus},
 		{"нет связи с ИИ, восстанови канал", interaction.InformationRequestCognitiveRecovery},
 		{"сообщии список моделей", interaction.InformationRequestCognitiveModelInventory},
@@ -313,6 +314,18 @@ func TestRememberedPreferenceIsNotRenderedAsIdentity(t *testing.T) {
 	}
 }
 
+func TestStableAcknowledgementsDoNotBecomeTravelRequests(t *testing.T) {
+	for _, message := range []string{"Хорошо", "Ладно", "Понял"} {
+		candidate := BuildDeterministicCandidate(message, "", time.Unix(10, 0).UnixMilli())
+		if candidate.PrimarySpeechAct != interaction.SpeechActConfirmation {
+			t.Fatalf("%q: expected confirmation, got %s", message, candidate.PrimarySpeechAct)
+		}
+		if reply, stable := StableConversationalReply(message, candidate.PrimarySpeechAct); !stable || reply != "Хорошо." {
+			t.Fatalf("%q: stable reply = %q, %t", message, reply, stable)
+		}
+	}
+}
+
 func TestCreativeRequestsDoNotRouteToWebResearch(t *testing.T) {
 	for _, message := range []string{"нарисуй круг", "расскажи сказку о колобке"} {
 		candidate := BuildDeterministicCandidate(message, "", time.Now().UnixMilli())

modules/interaction/domain/personal_memory.go

@@ -18,6 +18,7 @@ type PersonalMemory struct {
 
 type PersonalMemoryFact struct {
 	Kind                 string
+	Subject              string
 	Value                string
 	ConfidenceMillis     int64
 	SourceUtteranceID    interaction.HumanUtteranceID
@@ -26,6 +27,12 @@ type PersonalMemoryFact struct {
 
 const personalMemoryCurrentLocation = "CURRENT_LOCATION"
 
+// personalMemoryMaxValueRunes keeps the compatibility projection bounded.
+// Canonical utterances remain complete in storage; only the rebuildable
+// cognitive context is capped so old long-form instructions cannot become a
+// provider prompt or crowd out a current household fact.
+const personalMemoryMaxValueRunes = 512
+
 // CompilePersonalMemory turns direct, first-person human statements into a
 // small typed context. It deliberately preserves preferences as statements
 // rather than pretending that an arbitrary preference is a verified fact.
@@ -48,6 +55,7 @@ func CompilePersonalMemory(history []interaction.HumanUtterance) PersonalMemory
 		if preference := personalPreference(utterance.Content); preference != "" {
 			preferences = append(preferences, PersonalMemoryFact{
 				Kind:                 "PREFERENCE",
+				Subject:              preferenceSubject(preference),
 				Value:                preference,
 				ConfidenceMillis:     900,
 				SourceUtteranceID:    utterance.ID,
@@ -69,6 +77,21 @@ func CompilePersonalMemory(history []interaction.HumanUtterance) PersonalMemory
 	return result
 }
 
+// LatestPreference returns the current value for a semantic subject while
+// retaining older facts for historical evidence.
+func (m PersonalMemory) LatestPreference(subject string) (PersonalMemoryFact, bool) {
+	var latest PersonalMemoryFact
+	for _, fact := range m.Facts {
+		if fact.Kind != "PREFERENCE" || fact.Subject != subject {
+			continue
+		}
+		if latest.Kind == "" || fact.ObservedAtUnixMillis >= latest.ObservedAtUnixMillis {
+			latest = fact
+		}
+	}
+	return latest, latest.Kind != ""
+}
+
 func (m PersonalMemory) CurrentLocation() string {
 	for _, fact := range m.Facts {
 		if fact.Kind == personalMemoryCurrentLocation {
@@ -97,6 +120,47 @@ func (m PersonalMemory) ContextText() string {
 	return strings.Join(lines, "\n")
 }
 
+// RecallPersonalPreference answers the narrow, read-only memory question from
+// the rebuildable projection. It is intentionally conservative: only an
+// explicit question about the remembered test color is handled here; other
+// questions continue through the normal cognitive route.
+func RecallPersonalPreference(history []interaction.HumanUtterance, question string) string {
+	lowerQuestion := strings.ToLower(question)
+	if !strings.Contains(lowerQuestion, "цвет") || (!strings.Contains(lowerQuestion, "какой") && !strings.Contains(lowerQuestion, "назвал")) {
+		return ""
+	}
+	memory := CompilePersonalMemory(history)
+	for _, fact := range memory.Facts {
+		if fact.Kind != "PREFERENCE" || fact.Subject != "тестовый цвет" {
+			continue
+		}
+		current, ok := memory.LatestPreference(fact.Subject)
+		if !ok || current.SourceUtteranceID != fact.SourceUtteranceID {
+			continue
+		}
+		value := strings.TrimSpace(current.Value)
+		if separator := strings.Index(value, "—"); separator >= 0 {
+			value = strings.TrimSpace(value[separator+len("—"):])
+		}
+		if value != "" {
+			return "Вы называли тестовый цвет: " + value + "."
+		}
+	}
+	return ""
+}
+
+func preferenceSubject(preference string) string {
+	if separator := strings.Index(preference, "—"); separator >= 0 {
+		preference = preference[:separator]
+	}
+	subject := strings.TrimSpace(preference)
+	// Operational markers belong to the preference value, not a new subject.
+	if strings.HasPrefix(strings.ToLower(subject), "тестовый цвет") {
+		return "тестовый цвет"
+	}
+	return subject
+}
+
 func personalCurrentLocation(message string) string {
 	tokens := semanticTokens(message)
 	if len(tokens) < 4 || tokens[0] != "я" || !containsStemInTokens(tokens, "сейчас", "нахож", "жив") {
@@ -119,7 +183,7 @@ func personalPreference(message string) string {
 		}
 		value := strings.TrimSpace(lower[index+len(cue):])
 		value = strings.Trim(value, " .,!?:;—-")
-		if len([]rune(value)) >= 2 {
+		if len([]rune(value)) >= 2 && len([]rune(value)) <= personalMemoryMaxValueRunes {
 			return value
 		}
 	}
@@ -144,7 +208,7 @@ func personalPreference(message string) string {
 				separatorWidth = len("—")
 			}
 			value := strings.Trim(strings.TrimSpace(statement[separator+separatorWidth:]), " .,!?:;—-")
-			if key != "" && len([]rune(value)) >= 2 {
+			if key != "" && len([]rune(value)) >= 2 && len([]rune(value)) <= personalMemoryMaxValueRunes {
 				return key + " — " + value
 			}
 		}

modules/interaction/domain/personal_memory_test.go

@@ -28,3 +28,51 @@ func TestCompilePersonalMemoryDoesNotTreatThirdPartyStatementAsPersonalFact(t *t
 		t.Fatalf("ordinary factual question became personal memory: %+v", memory)
 	}
 }
+
+func TestCompilePersonalMemoryBoundsLongRememberedValues(t *testing.T) {
+	longValue := strings.Repeat("x", 513)
+	memory := CompilePersonalMemory([]interaction.HumanUtterance{{
+		Content:              "запомни: мой тестовый цвет — " + longValue,
+		ReceivedAtUnixMillis: 100,
+	}})
+	if memory.ContextText() != "" {
+		t.Fatalf("long remembered value leaked into cognitive context: %q", memory.ContextText())
+	}
+}
+
+func TestRecallPersonalPreferenceUsesProjectedAssertion(t *testing.T) {
+	answer := RecallPersonalPreference([]interaction.HumanUtterance{{
+		Content: "запомни: мой тестовый цвет — acceptance-m5-синий-914",
+	}}, "Какой тестовый цвет я назвал?")
+	if answer != "Вы называли тестовый цвет: acceptance-m5-синий-914." {
+		t.Fatalf("answer=%q", answer)
+	}
+}
+
+func TestCompilePersonalMemoryRanksCurrentSubjectWithoutDroppingHistory(t *testing.T) {
+	memory := CompilePersonalMemory([]interaction.HumanUtterance{
+		{ID: "old", Content: "запомни: мой тестовый цвет — фиолетовый", ReceivedAtUnixMillis: 100},
+		{ID: "new", Content: "запомни: мой тестовый цвет — acceptance-m5-синий-914", ReceivedAtUnixMillis: 200},
+	})
+	current, ok := memory.LatestPreference("тестовый цвет")
+	if !ok || current.Value != "тестовый цвет — acceptance-m5-синий-914" {
+		t.Fatalf("current=%+v ok=%t", current, ok)
+	}
+	if len(memory.Facts) != 2 {
+		t.Fatalf("historical evidence was dropped: %+v", memory.Facts)
+	}
+}
+
+func TestCompilePersonalMemoryNormalizesOperationalSubjectMarker(t *testing.T) {
+	memory := CompilePersonalMemory([]interaction.HumanUtterance{
+		{ID: "old", Content: "запомни: мой тестовый цвет — фиолетовый", ReceivedAtUnixMillis: 100},
+		{ID: "new", Content: "запомни: мой тестовый цвет M5 — лазурный", ReceivedAtUnixMillis: 200},
+	})
+	current, ok := memory.LatestPreference("тестовый цвет")
+	if !ok || current.SourceUtteranceID != "new" || !strings.Contains(current.Value, "лазурный") {
+		t.Fatalf("expected newest marked preference, got %+v, ok=%t", current, ok)
+	}
+	if len(memory.Facts) != 2 {
+		t.Fatalf("expected historical fact retention, got %d facts", len(memory.Facts))
+	}
+}

modules/interaction/domain/semantics.go

@@ -86,9 +86,12 @@ func InterpretUtterance(utterance interaction.HumanUtterance, projects []situati
 		result.Relations = append(result.Relations, relations...)
 	}
 	result.Reply = semanticReply(utterance.Content, result, previousFocus)
-	if isCurrentFocusQuestion(utterance.Content, selectedCandidate.PrimarySpeechAct) && previousFocus != nil {
+	if isCurrentFocusQuestion(utterance.Content, selectedCandidate.PrimarySpeechAct) {
 		result.NeedsCognition = false
 		result.Interpretation.NeedsCognition = false
+		result.CapabilityHints = nil
+		selectedCandidate.CapabilityHints = nil
+		result.Candidates[0] = selectedCandidate
 	}
 	return result, nil
 }
@@ -343,8 +346,10 @@ func semanticReply(message string, result SemanticResult, previousFocus *situati
 	case interaction.SpeechActRejection:
 		return "Понял."
 	case interaction.SpeechActQuestion:
-		lower := strings.ToLower(message)
-		if previousFocus != nil && (strings.Contains(lower, "сейчас") || strings.Contains(lower, "что мы") || strings.Contains(lower, "делаем")) {
+		if focusIntrospectionQuestion(semanticTokens(message)) {
+			if previousFocus == nil {
+				return "Сейчас активный Focus не выбран."
+			}
 			return "Сейчас в фокусе: " + previousFocus.ProjectDisplayName + "."
 		}
 		if result.Interpretation.Applicability == interaction.SemanticNeedsBroaderInterpret || result.Interpretation.Applicability == interaction.SemanticLowConfidence {
@@ -385,8 +390,7 @@ func isCurrentFocusQuestion(message string, act interaction.SpeechAct) bool {
 	if act != interaction.SpeechActQuestion {
 		return false
 	}
-	lower := strings.ToLower(message)
-	return strings.Contains(lower, "сейчас") || strings.Contains(lower, "что мы") || strings.Contains(lower, "делаем")
+	return focusIntrospectionQuestion(semanticTokens(message))
 }
 
 func hasCorrectionMeaning(candidate interaction.SemanticCandidate, utterance interaction.HumanUtterance) bool {

modules/interaction/domain/semantics_test.go

@@ -293,6 +293,13 @@ func TestInterpretUtteranceQuestionAboutCurrentWorkUsesFocus(t *testing.T) {
 	}
 }
 
+func TestInterpretUtteranceEnglishFocusIntrospectionStaysLocal(t *testing.T) {
+	result := interpretForTest(t, "Какой сейчас Focus разговора?", activeFocus("assistant", "Assistant"), "")
+	if result.Reply != "Сейчас в фокусе: Assistant." || result.NeedsCognition || len(result.CapabilityHints) != 0 {
+		t.Fatalf("focus introspection escaped local route: %+v", result)
+	}
+}
+
 func TestInterpretUtteranceInheritedFocusForThere(t *testing.T) {
 	result := interpretForTest(t, "Добавь туда новую настройку.", activeFocus("assistant", "Assistant"), "")
 	if len(result.Projects) != 1 || result.Projects[0].ProjectID != "assistant" || result.Projects[0].Provenance != situation.ProjectResolutionInheritedFocus {

tests/continuity_execution/verify-remote.ps1

@@ -10,7 +10,7 @@ function Invoke-JsonPost([string]$container, [string]$url, [object]$body, [bool]
         [IO.File]::WriteAllText($local, $json, [Text.UTF8Encoding]::new($false))
         docker --context $DockerContext cp $local "${container}:$remote" | Out-Null
         if ($internal) {
-            $response = docker --context $DockerContext exec $container sh -c 'wget -qO- --header="Content-Type: application/json" --header="Authorization: Bearer $COMMIT_EXECUTOR_TOKEN" --post-file "$1" "$2"' sh $remote $url
+            $response = docker --context $DockerContext exec -e "POST_FILE=$remote" -e "POST_URL=$url" $container sh -c 'wget -qO- --header="Content-Type: application/json" --header="Authorization: Bearer $COMMIT_EXECUTOR_TOKEN" --post-file "$POST_FILE" "$POST_URL"'
         } else {
             $response = docker --context $DockerContext exec $container wget -qO- --header "Content-Type: application/json" --post-file $remote $url
         }
@@ -31,7 +31,7 @@ function Expect-InternalPostStatus([string]$container, [string]$url, [object]$bo
         docker --context $DockerContext cp $local "${container}:$remote" | Out-Null
         $previous = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
-        $response = (docker --context $DockerContext exec $container sh -c 'wget -S -O /dev/null --header="Content-Type: application/json" --header="Authorization: Bearer $COMMIT_EXECUTOR_TOKEN" --post-file "$1" "$2"' sh $remote $url 2>&1) -join "`n"
+        $response = (docker --context $DockerContext exec -e "POST_FILE=$remote" -e "POST_URL=$url" $container sh -c 'wget -S -O /dev/null --header="Content-Type: application/json" --header="Authorization: Bearer $COMMIT_EXECUTOR_TOKEN" --post-file "$POST_FILE" "$POST_URL"' 2>&1) -join "`n"
         $ErrorActionPreference = $previous
         if ($response -notmatch "HTTP/1\.[01] $status") { throw "POST $url did not return $status" }
     } finally {