feat(q8): invite login, settings hub, and logout

Confirm Q8 account/settings IA and ship the closed-beta gap: - Extend POST /auth/login with display_name (required for DEMO-YKAVU) - Entry screen tabs: signup vs already-registered login - Settings hub with tone, autonomy, sessions, data-flow, logout - Logout revokes current session and clears local token Co-authored-by: okuma <o0kuma@users.noreply.github.com>
Merge remote-tracking branch 'origin/main'
2026-08-03 07:46:06 +00:00 · 2026-08-03 07:40:48 +00:00 · 2026-08-03 07:39:55 +00:00
14 changed files with 624 additions and 23 deletions
--- a/core-backend/a1_a2_routes.go
+++ b/core-backend/a1_a2_routes.go
@ -278,15 +278,7 @@ func registerA1A2Routes(r *gin.Engine, db *gorm.DB) {
 		query.Order("id asc").Find(&messages)
 		out := make([]gin.H, 0, len(messages))
 		for _, m := range messages {
-			out = append(out, gin.H{
-				"id":              m.ID,
-				"conversation_id": m.ConversationID,
-				"sender_id":       m.SenderID,
-				"sender_mode":     m.SenderMode,
-				"text":            m.Text,
-				"retracted":       m.Retracted,
-				"created_at":      m.CreatedAt,
-			})
+			out = append(out, messageJSON(m))
 		}
 		c.JSON(http.StatusOK, gin.H{"messages": out})
 	})
--- a/core-backend/b_routes.go
+++ b/core-backend/b_routes.go
@ -21,6 +21,8 @@ func messageJSON(m Message) gin.H {
 		"sender_mode":        m.SenderMode,
 		"text":               m.Text,
 		"retracted":          m.Retracted,
+		"draft_edited":       m.DraftEdited,
+		"naturalness_rating": m.NaturalnessRating,
 		"created_at":         m.CreatedAt,
 	}
 }
@ -251,6 +253,8 @@ const adminDashboardHTML = `<!doctype html>
        ['twin msgs', data.messages_twin_total],
        ['escalations', data.escalations_total],
        ['peer veto rate', (data.peer_veto_rate || 0).toFixed(3)],
+        ['초안 무수정 발송률', (data.draft_unedited_rate || 0).toFixed(3) + ' (n=' + (data.draft_edited_tracked_total || 0) + ')'],
+        ['자연스러움 긍정률', (data.naturalness_positive_rate || 0).toFixed(3) + ' (n=' + (data.naturalness_ratings_total || 0) + ')'],
        ['draft req', data.draft_requests],
        ['draft err rate', (data.draft_error_rate || 0).toFixed(3)],
        ['draft avg ms', (data.draft_latency_avg_ms || 0).toFixed(1)],
--- a/core-backend/main.go
+++ b/core-backend/main.go
@ -31,6 +31,20 @@ type sendMessageRequest struct {
 	// L2 draft outside the whitelist (PRD.md §2.2). Ignored for
 	// human-authored messages.
 	Approved bool `json:"approved"`
+	// OriginalDraftText is the AI draft's text before any edits, sent by the
+	// client alongside the (possibly edited) final Text so the server can
+	// diff them into Message.DraftEdited (PRD.md §5 L1 approval-rate proxy,
+	// deploy-checklist.md N4-12). Optional -- omitted/empty means "not
+	// derived from a draft, or an older client", and DraftEdited stays nil
+	// rather than being guessed. Only consulted for SenderMode == twin.
+	OriginalDraftText string `json:"original_draft_text"`
+}
+
+// messageFeedbackRequest is the explicit naturalness signal (vision.md
+// metric, PRD.md §5, deploy-checklist.md N4-12) — "이 답장 나답아요?" one-tap
+// thumbs up/down on a twin-authored message.
+type messageFeedbackRequest struct {
+	Natural bool `json:"natural"`
 }

 type updateTwinSettingsRequest struct {
@ -125,6 +139,34 @@ func setupRouter(db *gorm.DB, relay *ConnectionManager, ai *AIServiceClient) *gi
 			peerVetoRate = float64(conversationsVetoed) / float64(conversationsTotal)
 		}

+		// L1 approval-rate proxy (PRD.md §5 "초안을 수정 없이 그대로 발송한
+		// 비율", vision.md 자연스러움 지표, deploy-checklist.md N4-12):
+		// draft_edited is only set (non-nil) for twin-mode messages sent
+		// through the draft-approval flow with an original draft to diff
+		// against -- messages where it's nil (human-authored, or no draft
+		// text sent) are excluded from both counts so they can't dilute
+		// the rate either way. Zero-tracked-messages guarded the same way
+		// peer_veto_rate guards zero conversations above.
+		var draftTrackedTotal, draftEditedTrue int64
+		db.Model(&Message{}).Where("sender_mode = ? AND draft_edited IS NOT NULL", SenderTwin).Count(&draftTrackedTotal)
+		db.Model(&Message{}).Where("sender_mode = ? AND draft_edited = ?", SenderTwin, true).Count(&draftEditedTrue)
+		var draftUneditedRate float64
+		if draftTrackedTotal > 0 {
+			draftUneditedRate = float64(draftTrackedTotal-draftEditedTrue) / float64(draftTrackedTotal)
+		}
+
+		// Explicit naturalness signal ("이 답장 나답아요?" 원탭, vision.md 지표,
+		// deploy-checklist.md N4-12): same nil-excluded-from-denominator
+		// shape as draft_unedited_rate above -- unrated messages don't
+		// count as either positive or negative.
+		var naturalnessTrackedTotal, naturalnessPositiveTrue int64
+		db.Model(&Message{}).Where("sender_mode = ? AND naturalness_rating IS NOT NULL", SenderTwin).Count(&naturalnessTrackedTotal)
+		db.Model(&Message{}).Where("sender_mode = ? AND naturalness_rating = ?", SenderTwin, true).Count(&naturalnessPositiveTrue)
+		var naturalnessPositiveRate float64
+		if naturalnessTrackedTotal > 0 {
+			naturalnessPositiveRate = float64(naturalnessPositiveTrue) / float64(naturalnessTrackedTotal)
+		}
+
 		rt := runtimeMetrics.snapshot()
 		c.JSON(http.StatusOK, gin.H{
 			"users_total":           usersTotal,
@ -139,6 +181,14 @@ func setupRouter(db *gorm.DB, relay *ConnectionManager, ai *AIServiceClient) *gi
 			// vision.md doesn't pin down the exact denominator, so treat
 			// this as a first approximation, not the final definition.
 			"peer_veto_rate": peerVetoRate,
+			// L1 approval-rate proxy / naturalness instrumentation (PRD.md §5,
+			// vision.md, deploy-checklist.md N4-12). This is capture
+			// infrastructure, not a PoC conclusion -- see docs/roadmap.md and
+			// docs/deploy-checklist.md N4-12 for that distinction.
+			"draft_edited_tracked_total": draftTrackedTotal,
+			"draft_unedited_rate":        draftUneditedRate,
+			"naturalness_ratings_total":  naturalnessTrackedTotal,
+			"naturalness_positive_rate":  naturalnessPositiveRate,
 			// Process-local draft/AI timings (roadmap B). Reset on restart.
 			"draft_requests":               rt.DraftRequests,
 			"draft_errors":                 rt.DraftErrors,
@ -371,11 +421,23 @@ func setupRouter(db *gorm.DB, relay *ConnectionManager, ai *AIServiceClient) *gi
 			}
 		}

+		// Implicit L1-approval-rate signal (PRD.md §5, deploy-checklist.md
+		// N4-12): only meaningful for twin-mode sends where the client told
+		// us what the original draft said. Absent/empty original text
+		// (human messages, or a client that isn't draft-derived) leaves
+		// DraftEdited nil rather than guessing true/false.
+		var draftEdited *bool
+		if req.SenderMode == SenderTwin && req.OriginalDraftText != "" {
+			edited := req.Text != req.OriginalDraftText
+			draftEdited = &edited
+		}
+
 		message := Message{
 			ConversationID: convID,
 			SenderID:       req.SenderID,
 			SenderMode:     req.SenderMode,
 			Text:           req.Text,
+			DraftEdited:    draftEdited,
 		}
 		if err := db.Create(&message).Error; err != nil {
 			c.JSON(http.StatusInternalServerError, gin.H{"detail": err.Error()})
@ -432,6 +494,46 @@ func setupRouter(db *gorm.DB, relay *ConnectionManager, ai *AIServiceClient) *gi
 		c.JSON(http.StatusOK, gin.H{"id": message.ID, "retracted": true})
 	})

+	// "이 답장 나답아요?" 자연스러움 피드백(vision.md 지표, PRD.md §5,
+	// deploy-checklist.md N4-12): 트윈이 실제로 보낸 메시지 위에 붙는 가벼운
+	// 원탭 피드백 -- 사람이 직접 쓴 메시지는 이 지표의 대상이 아니라 400.
+	// 재제출은 덮어쓰기로 처리한다(409로 막지 않음): 클라이언트가 이미 한
+	// 번 탭한 메시지는 다시 UI를 보여주지 않아 재제출 자체가 드물고, 네트워크
+	// 재시도 등으로 같은 요청이 두 번 가더라도 "평가를 덮어쓴다"는 동작이
+	// 사용자에게 해롭지 않기 때문 -- retract의 409(한 번 되돌린 걸 다시
+	// 되돌릴 수 없음, 되돌리기는 상태를 한 방향으로만 바꿈)와는 성격이 다름.
+	r.POST("/messages/:id/feedback", func(c *gin.Context) {
+		msgID, ok := parseUintParam(c, "id")
+		if !ok {
+			return
+		}
+
+		var message Message
+		if err := db.First(&message, msgID).Error; err != nil {
+			c.JSON(http.StatusNotFound, gin.H{"detail": "message not found"})
+			return
+		}
+		if message.SenderMode != SenderTwin {
+			c.JSON(http.StatusBadRequest, gin.H{"detail": "only twin-authored messages can be rated for naturalness"})
+			return
+		}
+
+		var req messageFeedbackRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"detail": err.Error()})
+			return
+		}
+
+		natural := req.Natural
+		message.NaturalnessRating = &natural
+		if err := db.Save(&message).Error; err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"detail": err.Error()})
+			return
+		}
+
+		c.JSON(http.StatusOK, messageJSON(message))
+	})
+
 	r.POST("/conversations/:id/draft", func(c *gin.Context) {
 		convID, ok := parseUintParam(c, "id")
 		if !ok {
--- a/core-backend/main_test.go
+++ b/core-backend/main_test.go
@ -937,6 +937,304 @@ func TestRetractMissingMessage(t *testing.T) {
 	}
 }

+// --- Draft-edited implicit signal (PRD.md §5, deploy-checklist.md N4-12) ---
+
+func TestDraftEditedTrueWhenSentTextDiffersFromOriginalDraft(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, token := mustSignup(t, server.URL, "수정함")
+	setAutonomyLevel(t, server.URL, senderID, token, AutonomyL1)
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	resp := postJSON(t, convBase+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "ㅇㅇ 좋아 이따 보자", SenderMode: SenderTwin, Approved: true,
+		OriginalDraftText: "ㅇㅇ 좋아",
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200, got %d", resp.StatusCode)
+	}
+	var out map[string]interface{}
+	json.NewDecoder(resp.Body).Decode(&out)
+	msgID := uint(out["id"].(float64))
+
+	var message Message
+	db.First(&message, msgID)
+	if message.DraftEdited == nil || *message.DraftEdited != true {
+		t.Fatalf("expected DraftEdited=true, got %v", message.DraftEdited)
+	}
+}
+
+func TestDraftEditedFalseWhenSentTextMatchesOriginalDraft(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, token := mustSignup(t, server.URL, "그대로")
+	setAutonomyLevel(t, server.URL, senderID, token, AutonomyL1)
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	resp := postJSON(t, convBase+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "ㅇㅇ 좋아", SenderMode: SenderTwin, Approved: true,
+		OriginalDraftText: "ㅇㅇ 좋아",
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200, got %d", resp.StatusCode)
+	}
+	var out map[string]interface{}
+	json.NewDecoder(resp.Body).Decode(&out)
+	msgID := uint(out["id"].(float64))
+
+	var message Message
+	db.First(&message, msgID)
+	if message.DraftEdited == nil || *message.DraftEdited != false {
+		t.Fatalf("expected DraftEdited=false, got %v", message.DraftEdited)
+	}
+}
+
+func TestDraftEditedNilWhenNoOriginalDraftTextSent(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, token := mustSignup(t, server.URL, "구버전클라")
+	setAutonomyLevel(t, server.URL, senderID, token, AutonomyL1)
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	// No OriginalDraftText field at all (e.g. an older client) -- must not
+	// be guessed as edited or unedited, stays nil.
+	resp := postJSON(t, convBase+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "ㅇㅇ 알겠음", SenderMode: SenderTwin, Approved: true,
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200, got %d", resp.StatusCode)
+	}
+	var out map[string]interface{}
+	json.NewDecoder(resp.Body).Decode(&out)
+	msgID := uint(out["id"].(float64))
+	if _, present := out["draft_edited"]; !present || out["draft_edited"] != nil {
+		t.Fatalf("expected draft_edited to be null in response, got %v", out["draft_edited"])
+	}
+
+	var message Message
+	db.First(&message, msgID)
+	if message.DraftEdited != nil {
+		t.Fatalf("expected DraftEdited nil, got %v", *message.DraftEdited)
+	}
+}
+
+func TestDraftEditedNotSetForHumanMessages(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, _ := mustSignup(t, server.URL, "사람메시지")
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	// Even if a caller weirdly sent original_draft_text on a human message,
+	// it must be ignored -- DraftEdited is a twin-only signal.
+	resp := postJSON(t, convBase+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "안녕", SenderMode: SenderHuman,
+		OriginalDraftText: "다른 텍스트",
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200, got %d", resp.StatusCode)
+	}
+	var out map[string]interface{}
+	json.NewDecoder(resp.Body).Decode(&out)
+	msgID := uint(out["id"].(float64))
+
+	var message Message
+	db.First(&message, msgID)
+	if message.DraftEdited != nil {
+		t.Fatalf("expected DraftEdited nil for a human message, got %v", *message.DraftEdited)
+	}
+}
+
+// --- Explicit naturalness feedback ("이 답장 나답아요?", vision.md, N4-12) ---
+
+func TestMessageFeedbackOnTwinMessageReflectedOnRefetch(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, token := mustSignup(t, server.URL, "피드백")
+	setAutonomyLevel(t, server.URL, senderID, token, AutonomyL1)
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	if err := db.Create(&ConversationParticipant{ConversationID: conv.ID, UserID: senderID}).Error; err != nil {
+		t.Fatalf("create participant: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	sendResp := postJSON(t, convBase+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "ㅇㅇ 좋아", SenderMode: SenderTwin, Approved: true,
+	})
+	var sent map[string]interface{}
+	json.NewDecoder(sendResp.Body).Decode(&sent)
+	msgID := uint(sent["id"].(float64))
+
+	feedbackResp := postJSON(t, server.URL+"/messages/"+strconv.FormatUint(uint64(msgID), 10)+"/feedback", messageFeedbackRequest{Natural: true})
+	if feedbackResp.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200 submitting feedback, got %d", feedbackResp.StatusCode)
+	}
+
+	// Re-fetch via the conversation history endpoint (authenticated) and
+	// confirm the rating stuck.
+	req, _ := http.NewRequest(http.MethodGet, convBase+"/messages", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	listResp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatalf("list messages: %v", err)
+	}
+	var listOut map[string]interface{}
+	json.NewDecoder(listResp.Body).Decode(&listOut)
+	messages := listOut["messages"].([]interface{})
+	var found map[string]interface{}
+	for _, m := range messages {
+		mm := m.(map[string]interface{})
+		if uint(mm["id"].(float64)) == msgID {
+			found = mm
+		}
+	}
+	if found == nil {
+		t.Fatalf("sent message not found in refetch: %v", messages)
+	}
+	if found["naturalness_rating"] != true {
+		t.Fatalf("expected naturalness_rating=true on refetch, got %v", found["naturalness_rating"])
+	}
+
+	// Resubmission overwrites rather than 409ing (documented choice: a
+	// bounded rating change isn't unsafe the way re-retracting would be).
+	overwrite := postJSON(t, server.URL+"/messages/"+strconv.FormatUint(uint64(msgID), 10)+"/feedback", messageFeedbackRequest{Natural: false})
+	if overwrite.StatusCode != http.StatusOK {
+		t.Fatalf("expected 200 overwriting feedback, got %d", overwrite.StatusCode)
+	}
+	var overwriteOut map[string]interface{}
+	json.NewDecoder(overwrite.Body).Decode(&overwriteOut)
+	if overwriteOut["naturalness_rating"] != false {
+		t.Fatalf("expected naturalness_rating=false after overwrite, got %v", overwriteOut["naturalness_rating"])
+	}
+}
+
+func TestMessageFeedbackRejectsHumanMessage(t *testing.T) {
+	server, db := setupTestServer(t)
+	senderID, _ := mustSignup(t, server.URL, "사람피드백")
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	sendResp := postJSON(t, server.URL+"/conversations/"+strconv.FormatUint(uint64(conv.ID), 10)+"/messages", sendMessageRequest{
+		SenderID: senderID, Text: "안녕", SenderMode: SenderHuman,
+	})
+	var sent map[string]interface{}
+	json.NewDecoder(sendResp.Body).Decode(&sent)
+	msgID := uint(sent["id"].(float64))
+
+	resp := postJSON(t, server.URL+"/messages/"+strconv.FormatUint(uint64(msgID), 10)+"/feedback", messageFeedbackRequest{Natural: true})
+	if resp.StatusCode != http.StatusBadRequest {
+		t.Fatalf("expected 400 rating a human message, got %d", resp.StatusCode)
+	}
+}
+
+func TestMessageFeedbackMissingMessage(t *testing.T) {
+	server, _ := setupTestServer(t)
+	resp := postJSON(t, server.URL+"/messages/9999/feedback", messageFeedbackRequest{Natural: true})
+	if resp.StatusCode != http.StatusNotFound {
+		t.Fatalf("expected 404, got %d", resp.StatusCode)
+	}
+}
+
+// --- /admin/metrics rate calculations for the two signals above ---
+
+func TestAdminMetricsDraftUneditedAndNaturalnessRates(t *testing.T) {
+	runtimeMetrics = &RuntimeMetrics{} // see comment in TestAdminMetricsCountsMessagesEscalationsAndVeto
+	server, db := setupTestServer(t)
+	senderID, token := mustSignup(t, server.URL, "지표계산")
+	setAutonomyLevel(t, server.URL, senderID, token, AutonomyL1)
+
+	conv := Conversation{IsGroup: false}
+	if err := db.Create(&conv).Error; err != nil {
+		t.Fatalf("create conversation: %v", err)
+	}
+	convBase := server.URL + "/conversations/" + strconv.FormatUint(uint64(conv.ID), 10)
+
+	// 2 edited + 3 unedited twin sends -> draft_unedited_rate should be
+	// 3/5 = 0.6. A plain human message and a twin send with no
+	// original_draft_text must not affect the denominator either way.
+	postJSON(t, convBase+"/messages", sendMessageRequest{SenderID: senderID, Text: "인간이 직접", SenderMode: SenderHuman})
+	postJSON(t, convBase+"/messages", sendMessageRequest{SenderID: senderID, Text: "초안없음", SenderMode: SenderTwin, Approved: true})
+	for i := 0; i < 2; i++ {
+		postJSON(t, convBase+"/messages", sendMessageRequest{
+			SenderID: senderID, Text: "수정된 문구", SenderMode: SenderTwin, Approved: true,
+			OriginalDraftText: "원래 초안",
+		})
+	}
+	for i := 0; i < 3; i++ {
+		postJSON(t, convBase+"/messages", sendMessageRequest{
+			SenderID: senderID, Text: "그대로", SenderMode: SenderTwin, Approved: true,
+			OriginalDraftText: "그대로",
+		})
+	}
+
+	mreq, _ := http.NewRequest(http.MethodGet, server.URL+"/admin/metrics", nil)
+	mreq.Header.Set("Authorization", "Bearer test-admin-token")
+	resp, err := http.DefaultClient.Do(mreq)
+	if err != nil {
+		t.Fatalf("get metrics: %v", err)
+	}
+	var out map[string]interface{}
+	json.NewDecoder(resp.Body).Decode(&out)
+
+	if out["draft_edited_tracked_total"].(float64) != 5 {
+		t.Fatalf("expected 5 draft-tracked messages, got %v", out["draft_edited_tracked_total"])
+	}
+	if got := out["draft_unedited_rate"].(float64); got != 0.6 {
+		t.Fatalf("expected draft_unedited_rate 0.6, got %v", got)
+	}
+
+	// No naturalness ratings submitted yet -- must not divide by zero.
+	if out["naturalness_ratings_total"].(float64) != 0 {
+		t.Fatalf("expected 0 naturalness ratings, got %v", out["naturalness_ratings_total"])
+	}
+	if out["naturalness_positive_rate"].(float64) != 0 {
+		t.Fatalf("expected naturalness_positive_rate 0 (zero-guarded), got %v", out["naturalness_positive_rate"])
+	}
+
+	// Rate a couple of the twin messages we just sent and check the rate
+	// updates too (1 positive out of 2 rated -> 0.5).
+	var twinMessages []Message
+	db.Where("sender_mode = ?", SenderTwin).Order("id asc").Find(&twinMessages)
+	if len(twinMessages) < 2 {
+		t.Fatalf("expected at least 2 twin messages, got %d", len(twinMessages))
+	}
+	postJSON(t, server.URL+"/messages/"+strconv.FormatUint(uint64(twinMessages[0].ID), 10)+"/feedback", messageFeedbackRequest{Natural: true})
+	postJSON(t, server.URL+"/messages/"+strconv.FormatUint(uint64(twinMessages[1].ID), 10)+"/feedback", messageFeedbackRequest{Natural: false})
+
+	resp2, err := http.DefaultClient.Do(mreq)
+	if err != nil {
+		t.Fatalf("get metrics again: %v", err)
+	}
+	var out2 map[string]interface{}
+	json.NewDecoder(resp2.Body).Decode(&out2)
+	if out2["naturalness_ratings_total"].(float64) != 2 {
+		t.Fatalf("expected 2 naturalness ratings, got %v", out2["naturalness_ratings_total"])
+	}
+	if got := out2["naturalness_positive_rate"].(float64); got != 0.5 {
+		t.Fatalf("expected naturalness_positive_rate 0.5, got %v", got)
+	}
+}
+
 func TestDraftMissingConversation(t *testing.T) {
 	server, _ := setupTestServer(t)
 	resp := postJSON(t, server.URL+"/conversations/9999/draft", draftMessageRequest{
--- a/core-backend/models.go
+++ b/core-backend/models.go
@ -129,6 +129,21 @@ type Message struct {
 	// one-tap undo") -- set via POST /messages/:id/retract, twin-authored
 	// messages only.
 	Retracted bool `gorm:"not null;default:false"`
+	// DraftEdited is the implicit L1-approval-rate signal (PRD.md §5's
+	// "초안을 수정 없이 그대로 발송한 비율", vision.md's naturalness metric,
+	// deploy-checklist.md N4-12). Nil for human-authored messages and for
+	// any message not sent through the draft-approval flow (old rows, or a
+	// client that didn't send original_draft_text) -- only set to a
+	// concrete true/false for twin-mode messages where the client told us
+	// what the original AI draft text was, so we could diff it against
+	// what actually got sent. Never guessed.
+	DraftEdited *bool
+	// NaturalnessRating is the explicit "이 답장 나답아요?" signal (vision.md
+	// naturalness metric, PRD.md §5, deploy-checklist.md N4-12): true = 👍,
+	// false = 👎, nil = not rated yet. Set via POST /messages/:id/feedback,
+	// twin-authored messages only (rating a human's own words doesn't make
+	// sense for this metric).
+	NaturalnessRating *bool
 	CreatedAt         time.Time
 }

--- a/docs/deploy-checklist.md
+++ b/docs/deploy-checklist.md
@ -232,7 +232,7 @@ API 계층은 프로덕션에서 검증됨 (`CORE_API_BASE=https://msn.iykyka.co
 | ID | 작업 | Status | 비고 |
 |----|------|--------|------|
 | **N4-11** | 오프라인 메시지 큐 | **done (부분 검증)** (2026-08-03) | 서버는 이미 모든 메시지를 DB에 durable하게 저장 — 별도 큐를 새로 만들지 않고 `GET /conversations/:id/messages?since_id=`(신규, 파라미터 없으면 기존 전체 히스토리 그대로) + 클라이언트 WS 재연결(backoff 1s→2s→...→30s 캡, 성공 시 리셋, `mobile/lib/services/ws_client.dart`) + 재연결/앱 포그라운드 복귀 시 since_id 캐치업(`mobile/lib/screens/chat_screen.dart`)으로 gap을 메움. `core-backend/a1_a2_test.go`(`TestListMessagesSinceID`)와 `mobile/test/message_sync_test.dart`(backoff 계산 + 중복 없는 병합을 순수 함수로 뽑아 검증)로 확인한 부분과, **실제 소켓 재연결 타이밍·앱 백그라운드/포그라운드 전환에서 OS가 소켓을 어떻게 처리하는지는 단위 테스트로 증명 불가 — 실기기 Android QA 남음**(N4-C6b와 같은 프레이밍) |
-| **N4-12** | 자연스러움 피드백 UI | todo | vision 지표 |
+| **N4-12** | 자연스러움 피드백 UI | **done (계측만)** (2026-08-03) | vision 지표. **PoC 실행/결론이 아니라 캡처 장치** — 실제 %는 N5 이후 실 사용 데이터로. 암묵 신호: `core-backend/models.go` `Message.DraftEdited`(nullable — 사람 메시지/구버전 클라는 nil, 트윈 승인-발송 경로에서 클라이언트가 `original_draft_text`를 같이 보내면 서버가 diff해 true/false) + `mobile/lib/screens/chat_screen.dart` `_sendTwinApproved`가 `_pendingDraft.text`(편집 전 원본)를 함께 전송. 명시 신호: `Message.NaturalnessRating`(nullable bool) + `POST /messages/:id/feedback`(트윈 메시지만, 400 아니면 재제출은 덮어씀 — 409 아님) + `chat_screen.dart`/`message_bubble.dart`의 "이 답장 나답아요?" 원탭 👍/👎(한 번 탭하면 그 세션에서 다시 안 보여줌, `_ratedMessageIds` + `message_sync.dart`의 `ratedMessageIdsFrom()`). `/admin/metrics`에 `draft_unedited_rate`/`naturalness_positive_rate`(둘 다 분모 0일 때 0으로 zero-guard, `peer_veto_rate`와 동일 패턴) 추가, `adminDashboardHTML`에 카드 2개 추가. 백엔드 테스트 8개(`core-backend/main_test.go`) + 모바일 테스트 5개(`mobile/test/`) 추가, 전부 통과 |
 | **N4-13** | `prototype.md` `SHARE_URL` | todo | Master 기입 |
 | **N4-14** | 내부 release APK | todo | `docs/android-release.md` |
 | **N4-15** | roadmap/`[~]` 동기화 | todo | 완료 시 체크 |
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@ -121,9 +121,12 @@
  409). 계정 삭제 시 코드는 "사용됨" 상태를 유지한 채 유저 참조만 지움. 발급자 인증은
  `ADMIN_API_TOKEN`(A2). 운영 절차·만료/회수는 C (`docs/invite-ops.md`)
 - [~] `vision.md` 성공 지표(자연스러움·거부율·안전선 위반) 계측용 분석/피드백 수집 — 거부율은
-  `/admin/metrics`의 `peer_veto_rate`로 1차 근사 가능해짐(대화방 단위, 확정 정의 아님). 자연스러움
-  피드백 수집 UI는 Flutter 클라이언트 책임이라 보류. 안전선 위반 0건은 런타임에 "수집"하는 지표라기
-  보다 지금까지의 하드게이트 테스트들이 이미 보증하는 것 — 별도 계측 불필요
+  `/admin/metrics`의 `peer_veto_rate`로 1차 근사 가능해짐(대화방 단위, 확정 정의 아님). **2026-08-03
+  자연스러움 계측 인프라 추가(deploy-checklist.md N4-12, 아래 참고)**: 암묵 신호(초안 무수정
+  발송률)와 명시 신호(원탭 "이 답장 나답아요?" 피드백) 둘 다 캡처되어 `/admin/metrics`에
+  집계된다. **주의 — 이건 PoC 실행이나 실제 수치 결론이 아니라 계측/캡처 장치일 뿐이다.**
+  실제 자연스러움 %는 실 베타 사용 데이터가 쌓여야 나온다(N5 이후). 안전선 위반 0건은 런타임에
+  "수집"하는 지표라기보다 지금까지의 하드게이트 테스트들이 이미 보증하는 것 — 별도 계측 불필요
 - [x] 모니터링 대시보드 (에스컬레이션 트리거율, 생성 지연시간, 오류율) — **2026-08-03 재확인**:
  이 항목의 `[~]` 상태와 "대시보드 UI 자체와 생성 지연시간·오류율은 아직 없음" 설명은 실제로는
  낡은 기록이었음(§B의 "생성 지연시간·오류율 계측"/"모니터링 대시보드(최소)" `[x]` 항목과 서로
--- a/mobile/lib/models/models.dart
+++ b/mobile/lib/models/models.dart
@ -174,6 +174,8 @@ class ChatMessage {
    required this.text,
    required this.retracted,
    required this.createdAt,
+    this.draftEdited,
+    this.naturalnessRating,
  });

  final int id;
@ -183,6 +185,15 @@ class ChatMessage {
  final String text;
  final bool retracted;
  final DateTime createdAt;
+  /// L1 approval-rate proxy (PRD.md §5, deploy-checklist.md N4-12) — null
+  /// for human messages and any twin message the server couldn't diff
+  /// against an original draft; otherwise true = edited before send,
+  /// false = sent verbatim. Purely informational on the client; the server
+  /// is the source of truth (computed once at send time).
+  final bool? draftEdited;
+  /// "이 답장 나답아요?" explicit feedback (vision.md metric, N4-12) — null
+  /// = not rated yet, true = 👍, false = 👎.
+  final bool? naturalnessRating;

  bool get isTwin => senderMode == SenderMode.twin;

@ -196,6 +207,8 @@ class ChatMessage {
      text: json['text'] as String? ?? '',
      retracted: json['retracted'] as bool? ?? false,
      createdAt: DateTime.tryParse(json['created_at'] as String? ?? '') ?? DateTime.now(),
+      draftEdited: json['draft_edited'] as bool?,
+      naturalnessRating: json['naturalness_rating'] as bool?,
    );
  }
 }
--- a/mobile/lib/screens/chat_screen.dart
+++ b/mobile/lib/screens/chat_screen.dart
@ -52,6 +52,10 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
  bool _loadingHistory = true;
  late bool _floodBlocked;
  late final ApiClient _api;
+  // "이 답장 나답아요?" 피드백(vision.md 지표, deploy-checklist.md N4-12):
+  // 이번 세션에서(또는 서버에서 이미) 평가된 메시지 id들 — 한 번 탭한 메시지엔
+  // 화면이 다시 그려져도(리로드 없이) 뱃지를 다시 보여주지 않는다.
+  final _ratedMessageIds = <int>{};
  // 답장 마감 알림(roadmap.md §2.7-F) — 이 대화방에 걸린 "이따 답장" 스누즈 시각.
  // 서버에는 존재하지 않는 순수 온디바이스 상태라 화면에 들어올 때 로컬 DB에서 로드한다.
  DateTime? _snoozedUntil;
@ -97,6 +101,7 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
        _messages
          ..clear()
          ..addAll(merged);
+        _syncRatedMessageIds(merged);
      });
      _scrollToEnd();
      _markLatestRead();
@ -194,6 +199,7 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
          ..clear()
          ..addAll(history);
        _loadingHistory = false;
+        _syncRatedMessageIds(history);
      });
      _scrollToEnd();
      // 읽음 마커는 화면에 들어올 때가 아니라 "나갈 때"(dispose) 찍는다 --
@ -232,6 +238,28 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
    });
  }

+  /// 서버에서 이미 평가된 메시지(다른 세션/기기에서 탭했거나, 새로고침 전에
+  /// 이미 평가된 경우)를 _ratedMessageIds에 반영 — UI를 다시 보여주지 않기
+  /// 위함. 실제 판정은 message_sync.dart의 순수 함수(ratedMessageIdsFrom)로
+  /// 분리해 위젯 없이도 테스트할 수 있게 했다.
+  void _syncRatedMessageIds(List<ChatMessage> messages) {
+    _ratedMessageIds.addAll(ratedMessageIdsFrom(messages));
+  }
+
+  /// "이 답장 나답아요?" 원탭 피드백(vision.md 지표, deploy-checklist.md
+  /// N4-12) — 트윈이 보낸 메시지 위에서만 노출(MessageBubble이 이미 twin +
+  /// isMine + !retracted로 걸러서 호출). 한 번 탭하면 다시 평가를 받지 않음.
+  Future<void> _submitFeedback(ChatMessage message, bool natural) async {
+    setState(() => _ratedMessageIds.add(message.id));
+    try {
+      await _api.submitMessageFeedback(message.id, natural);
+    } on ApiException {
+      // Best-effort — this is a soft signal, not a safety-critical action;
+      // if it fails to save server-side we don't re-offer the tap (that
+      // would look like a broken button), we just quietly drop it.
+    }
+  }
+
  void _onEvent(Map<String, dynamic> event) {
    final retractionId = _socket!.parseRetractionId(event);
    if (retractionId != null) {
@ -247,6 +275,8 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
            text: m.text,
            retracted: true,
            createdAt: m.createdAt,
+            draftEdited: m.draftEdited,
+            naturalnessRating: m.naturalnessRating,
          );
        }
      });
@ -368,6 +398,10 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
        text: text,
        senderMode: SenderMode.twin,
        approved: true,
+        // 초안 무수정 발송률(PRD.md §5, N4-12): draft.text는 사용자가 편집
+        // 컨트롤러(_draftEdit)를 건드리기 전 AI 초안 원문 — 서버가 이걸
+        // 최종 text와 diff해 draft_edited를 계산한다.
+        originalDraftText: draft.text,
      );
      setState(() {
        _pendingDraft = null;
@ -675,10 +709,18 @@ class _ChatScreenState extends State<ChatScreen> with WidgetsBindingObserver {
                        itemCount: _messages.length,
                        itemBuilder: (context, i) {
                          final m = _messages[i];
+                          final isMine = me != null && m.senderId == me;
                          return MessageBubble(
                            message: m,
-                            isMine: me != null && m.senderId == me,
+                            isMine: isMine,
                            onRetract: m.isTwin ? () => _retract(m) : null,
+                            // "이 답장 나답아요?" (vision.md 지표, N4-12): 트윈이
+                            // 실제로 보낸(=내가 보낸) 메시지에만 노출, 이미
+                            // 평가된 메시지는 다시 보여주지 않는다.
+                            alreadyRated: _ratedMessageIds.contains(m.id),
+                            onFeedback: m.isTwin && isMine && !m.retracted
+                                ? (natural) => _submitFeedback(m, natural)
+                                : null,
                          );
                        },
                      ),
--- a/mobile/lib/services/api_client.dart
+++ b/mobile/lib/services/api_client.dart
@ -194,12 +194,28 @@ class ApiClient {
    required String text,
    SenderMode senderMode = SenderMode.human,
    bool approved = false,
+    // 초안 무수정 발송률(PRD.md §5, deploy-checklist.md N4-12): 트윈
+    // 승인-발송 경로에서만 AI 초안 원문을 같이 보내 서버가 diff할 수 있게
+    // 한다. null이면(사람 메시지, 또는 초안 기반이 아닌 발송) 아예 필드를
+    // 넣지 않음 — relationshipTier/autonomyLevel 등과 같은 "값 있을 때만
+    // 포함" 관례.
+    String? originalDraftText,
  }) async {
    final json = await _json('POST', '/conversations/$conversationId/messages', body: {
      'sender_id': senderId,
      'text': text,
      'sender_mode': senderMode == SenderMode.twin ? 'twin' : 'human',
      if (approved) 'approved': true,
+      if (originalDraftText != null) 'original_draft_text': originalDraftText,
+    });
+    return ChatMessage.fromJson(json);
+  }
+
+  /// "이 답장 나답아요?" 자연스러움 피드백(vision.md 지표, N4-12) — 트윈
+  /// 메시지에만 허용. 재제출 시 서버가 덮어쓴다(에러 아님).
+  Future<ChatMessage> submitMessageFeedback(int messageId, bool natural) async {
+    final json = await _json('POST', '/messages/$messageId/feedback', body: {
+      'natural': natural,
    });
    return ChatMessage.fromJson(json);
  }
--- a/mobile/lib/services/message_sync.dart
+++ b/mobile/lib/services/message_sync.dart
@ -53,3 +53,12 @@ List<ChatMessage> mergeNewMessages(
  if (toAdd.isEmpty) return List.of(existing);
  return [...existing, ...toAdd];
 }
+
+/// "이 답장 나답아요?" 피드백(vision.md 지표, deploy-checklist.md N4-12) 순수
+/// 로직: [messages] 중 이미 평가된(naturalness_rating != null) 메시지 id들만
+/// 뽑아낸다. chat_screen.dart는 히스토리를 새로 불러올 때마다(또는 캐치업
+/// 때마다) 이 결과를 세션 상태의 "이미 평가됨" 집합에 합쳐 — 서버가 이미
+/// 알고 있는 평가는 화면이 다시 그려져도 뱅지/버튼을 다시 보여주지 않는다.
+Set<int> ratedMessageIdsFrom(List<ChatMessage> messages) {
+  return messages.where((m) => m.naturalnessRating != null).map((m) => m.id).toSet();
+}
--- a/mobile/lib/widgets/message_bubble.dart
+++ b/mobile/lib/widgets/message_bubble.dart
@ -10,11 +10,18 @@ class MessageBubble extends StatelessWidget {
    required this.message,
    required this.isMine,
    this.onRetract,
+    this.alreadyRated = false,
+    this.onFeedback,
  });

  final ChatMessage message;
  final bool isMine;
  final VoidCallback? onRetract;
+  // "이 답장 나답아요?" 피드백(vision.md 지표, deploy-checklist.md N4-12).
+  // alreadyRated면(이번 세션에 이미 탭했거나 서버에 기록이 있으면) 버튼 대신
+  // 조용한 완료 표시만 보여주고 다시 탭할 수 없게 한다 — 재평가 없음.
+  final bool alreadyRated;
+  final void Function(bool natural)? onFeedback;

  static String _time(DateTime dt) {
    final local = dt.toLocal();
@ -131,6 +138,49 @@ class MessageBubble extends StatelessWidget {
              label: const Text('되돌리기'),
            ),
          ),
+        if (twin && !retracted && onFeedback != null)
+          Padding(
+            padding: const EdgeInsets.only(top: 2, right: 4, left: 4),
+            child: alreadyRated
+                ? Text(
+                    '평가 고마워요',
+                    style: theme.textTheme.labelSmall?.copyWith(
+                      color: scheme.onSurfaceVariant.withValues(alpha: 0.6),
+                      fontSize: 11,
+                    ),
+                  )
+                : Row(
+                    mainAxisSize: MainAxisSize.min,
+                    children: [
+                      Text(
+                        '이 답장 나답아요?',
+                        style: theme.textTheme.labelSmall?.copyWith(
+                          color: scheme.onSurfaceVariant,
+                          fontSize: 11,
+                        ),
+                      ),
+                      const SizedBox(width: 2),
+                      IconButton(
+                        tooltip: '나다워요',
+                        onPressed: () => onFeedback!(true),
+                        padding: EdgeInsets.zero,
+                        constraints: const BoxConstraints(minWidth: 28, minHeight: 28),
+                        iconSize: 16,
+                        color: scheme.onSurfaceVariant,
+                        icon: const Icon(Icons.thumb_up_outlined),
+                      ),
+                      IconButton(
+                        tooltip: '나답지 않아요',
+                        onPressed: () => onFeedback!(false),
+                        padding: EdgeInsets.zero,
+                        constraints: const BoxConstraints(minWidth: 28, minHeight: 28),
+                        iconSize: 16,
+                        color: scheme.onSurfaceVariant,
+                        icon: const Icon(Icons.thumb_down_outlined),
+                      ),
+                    ],
+                  ),
+          ),
      ],
    );

--- a/mobile/test/message_sync_test.dart
+++ b/mobile/test/message_sync_test.dart
@ -2,7 +2,7 @@ import 'package:flutter_test/flutter_test.dart';
 import 'package:ykavu_mobile/models/models.dart';
 import 'package:ykavu_mobile/services/message_sync.dart';

-ChatMessage _msg(int id, {bool retracted = false}) => ChatMessage(
+ChatMessage _msg(int id, {bool retracted = false, bool? naturalnessRating}) => ChatMessage(
      id: id,
      conversationId: 1,
      senderId: 1,
@ -10,6 +10,7 @@ ChatMessage _msg(int id, {bool retracted = false}) => ChatMessage(
      text: 'msg-$id',
      retracted: retracted,
      createdAt: DateTime(2026, 8, 3),
+      naturalnessRating: naturalnessRating,
    );

 void main() {
@ -82,4 +83,27 @@ void main() {
      expect(merged.length, 2);
    });
  });
+
+  // "이 답장 나답아요?" 피드백(vision.md 지표, deploy-checklist.md N4-12):
+  // chat_screen.dart는 이 순수 함수의 결과를 세션 상태의 "이미 평가됨" id
+  // 집합에 합쳐서, 히스토리를 다시 불러와도(화면 재진입, 캐치업 등) 이미
+  // 평가된 메시지 위에 뱃지/버튼을 다시 보여주지 않는다.
+  group('ratedMessageIdsFrom', () {
+    test('picks up ids with a non-null naturalness rating (true or false)', () {
+      final messages = [
+        _msg(1, naturalnessRating: true),
+        _msg(2, naturalnessRating: false),
+        _msg(3),
+      ];
+      expect(ratedMessageIdsFrom(messages), {1, 2});
+    });
+
+    test('empty list -> empty set', () {
+      expect(ratedMessageIdsFrom(const []), isEmpty);
+    });
+
+    test('no rated messages -> empty set', () {
+      expect(ratedMessageIdsFrom([_msg(1), _msg(2)]), isEmpty);
+    });
+  });
 }
--- a/mobile/test/models_test.dart
+++ b/mobile/test/models_test.dart
@ -74,4 +74,37 @@ void main() {
    });
    expect(withoutOverride.autonomyLevel, isNull);
  });
+
+  // 초안 무수정 발송률 + 자연스러움 피드백(PRD.md §5, deploy-checklist.md
+  // N4-12): 서버가 null을 보낼 수 있는 두 필드가 각각 null/true/false를
+  // 그대로 통과시키는지 확인.
+  test('ChatMessage parses draft_edited and naturalness_rating when present', () {
+    final withBoth = ChatMessage.fromJson({
+      'id': 10,
+      'conversation_id': 1,
+      'sender_id': 1,
+      'sender_mode': 'twin',
+      'text': 'ㅇㅇ 좋아',
+      'retracted': false,
+      'draft_edited': true,
+      'naturalness_rating': false,
+      'created_at': '2026-08-03T00:00:00Z',
+    });
+    expect(withBoth.draftEdited, isTrue);
+    expect(withBoth.naturalnessRating, isFalse);
+  });
+
+  test('ChatMessage defaults draft_edited and naturalness_rating to null when absent', () {
+    final withoutEither = ChatMessage.fromJson({
+      'id': 11,
+      'conversation_id': 1,
+      'sender_id': 1,
+      'sender_mode': 'human',
+      'text': '안녕',
+      'retracted': false,
+      'created_at': '2026-08-03T00:00:00Z',
+    });
+    expect(withoutEither.draftEdited, isNull);
+    expect(withoutEither.naturalnessRating, isNull);
+  });
 }