현재 작업 상태 저장

.env.example

@@ -12,6 +12,8 @@ ENGINE_MODE=claude_cli
 ANTHROPIC_API_KEY=
 OPENAI_API_KEY=
 OPENAI_BASE_URL=https://api.openai.com/v1
+VIGNETTE_VOICE_POC_SAMPLE_TTS=false
+VIGNETTE_VOICE_POC_SAMPLE_TTS_DIR=
 
 # Auth/session. Local dev may enable dev-login; public/prod must not.
 SESSION_SECRET=dev-insecure-change-me
@@ -22,12 +24,14 @@ AUTH_ALLOWED_EMAIL_DOMAINS=["hs.ac.kr","twentyoz.kr"]
 AUTH_TEACHER_EMAILS=[]
 AUTH_ADMIN_EMAILS=[]
 AUTH_DEV_LOGIN_ENABLED=true
+AUTH_DEV_LOGIN_EXTRA_ORIGINS=[]
 DEFAULT_AFFILIATION=
 AUTO_SEED_PERSONAS=false
 ALLOW_SEED_PERSONA_FALLBACK=false
 EVALUATOR_GOLDEN_FEWSHOT_ENABLED=false
 FRONTEND_BASE_URL=http://localhost:5173
-CORS_ORIGINS=["https://vignette.chanpaca.net","https://vignette-b1q.pages.dev","http://localhost:5170","http://localhost:5171","http://localhost:5172","http://localhost:5173","http://localhost:5174","http://localhost:5175","http://localhost:5176","http://localhost:5177","http://localhost:5178","http://localhost:5179","http://localhost:5180","http://127.0.0.1:5170","http://127.0.0.1:5171","http://127.0.0.1:5172","http://127.0.0.1:5173","http://127.0.0.1:5174","http://127.0.0.1:5175","http://127.0.0.1:5176","http://127.0.0.1:5177","http://127.0.0.1:5178","http://127.0.0.1:5179","http://127.0.0.1:5180"]
+FRONTEND_ORIGIN_MAP={"api-vignette.chanpaca.net":"https://vignette.chanpaca.net","api-vnet.18ka.net":"https://vnet.18ka.net"}
+CORS_ORIGINS=["https://vignette.chanpaca.net","https://vnet.18ka.net","https://vignette-b1q.pages.dev","http://localhost:5170","http://localhost:5171","http://localhost:5172","http://localhost:5173","http://localhost:5174","http://localhost:5175","http://localhost:5176","http://localhost:5177","http://localhost:5178","http://localhost:5179","http://localhost:5180","http://127.0.0.1:5170","http://127.0.0.1:5171","http://127.0.0.1:5172","http://127.0.0.1:5173","http://127.0.0.1:5174","http://127.0.0.1:5175","http://127.0.0.1:5176","http://127.0.0.1:5177","http://127.0.0.1:5178","http://127.0.0.1:5179","http://127.0.0.1:5180"]
 
 # Live2D runtime. Models must be configured per persona via live2dModelUrl.
 VITE_LIVE2D_CUBISM_CORE=/live2d/live2dcubismcore.min.js

AGENTS.md

@@ -0,0 +1,101 @@
+# AGENTS.md — Vignette 프로젝트 작업 지침
+
+> Vignette = AI 심리상담 시뮬레이션 훈련 플랫폼 (한신대 산학협력).
+> 모노레포: `apps/api`(FastAPI/Python), `apps/web`(React 19/Vite/Playwright), `docs`, `infra`, `scripts`.
+
+## 📚 문서 맵 — 작업 전 해당 가이드를 먼저 읽어라
+
+| 목적 | 문서 |
+|---|---|
+| 저장소 개요·빠른 시작 | [`README.md`](./README.md) |
+| **로컬 서버 띄우기·테스트** | [`docs/guides/local-development.md`](./docs/guides/local-development.md) |
+| 시스템 아키텍처·데이터 흐름 | [`docs/guides/architecture.md`](./docs/guides/architecture.md) |
+| 테스트·검증 실행 | [`docs/guides/testing.md`](./docs/guides/testing.md) |
+| 원천문서·갭 로드맵 | [`docs/guides/source-docs-and-gaps.md`](./docs/guides/source-docs-and-gaps.md) |
+| **SSOT 상태판** | [`docs/dev_dashboard.html`](./docs/dev_dashboard.html) |
+| 백로그 | [`docs/ops/backlog-2026-06-26.md`](./docs/ops/backlog-2026-06-26.md) |
+| **서연 아바타 핸드오프** | [`docs/ops/handoff-avatar-seoyeon-2026-06-27.md`](./docs/ops/handoff-avatar-seoyeon-2026-06-27.md) |
+
+작업 결과로 동작/구조가 바뀌면 해당 가이드와 SSOT 대시보드를 함께 갱신한다.
+
+---
+
+## ⚠️ 0. 무조건 OS를 먼저 파악하고 시작한다 (최우선·필수)
+
+**어떤 작업이든 명령을 실행하기 전에 OS와 셸을 먼저 확정하라.** 이걸 건너뛰면
+경로/인코딩/도구 차이로 시간을 크게 낭비한다(실제로 그랬다).
+
+작업 시작 시 반드시 확인할 것:
+
+1. **OS / 셸**: 이 저장소의 주 개발 환경은 **Windows 11 + PowerShell**이다.
+   POSIX를 가정하지 마라. Bash 도구도 쓸 수 있으나 셸마다 문법이 다르다.
+2. **경로 규칙**: Windows 절대경로(`D:\...`, `C:\...`). 한글·공백 포함 경로가 흔하다
+   (예: OneDrive `문서\카카오톡 받은 파일`). `-LiteralPath`로 다루고, 외부 도구에
+   넘기기 전에 **ASCII 이름으로 로컬 복사**해 인코딩/공백 문제를 차단하라.
+3. **PowerShell 판(5.1) 주의**: 인라인 `if(){}else{}`를 식으로 못 쓴다(삼항 없음).
+   네이티브 exe stderr를 `2>&1`로 합치지 마라(ErrorRecord로 감싸짐).
+   기본 출력 인코딩은 UTF-16 — 다른 도구가 읽을 파일은 `-Encoding utf8`.
+4. **외부 CLI는 실제로 블로킹되는지 확인**: GUI 런처(`soffice.exe` 등)는 즉시
+   detach되어 `Start-Process -Wait`가 변환을 안 기다린다. 실제 작업 프로세스
+   (`soffice.bin`)를 직접 호출하라. 좀비 프로세스가 락을 잡으면 정리부터 한다.
+5. **도구 가용성 먼저 탐지**: 변환/처리 전에 LibreOffice·pandoc·python 라이브러리·
+   Playwright 브라우저 등 무엇이 설치돼 있는지 먼저 확인하고 경로를 잡아라.
+
+> 한 줄 요약: **"먼저 OS·셸·경로·도구를 확정한 뒤 실행한다."** 추정 금지.
+
+---
+
+## 1. 운영 원칙
+
+- **가짜 증거로 DONE 표기 금지.** 실증/외부 의존/소유자 결정이 필요한 항목은
+  `docs/ops/backlog-*.md`에 분류해 추적한다(B1 코스메틱 · B2 환경제약 · B3 소유자결정 · B4 외부거버넌스).
+- **`docs/dev_dashboard.html`이 SSOT(단일 진실 공급원)다.** 상태·검증 증거·결정 필요·로드맵의 권위 기준이며, 새 발견·작업 결과·상태 변경은 별도 문서로만 남기지 말고 대시보드에 반영/동기화한다. 백로그(`docs/ops/backlog-*.md`)는 대시보드와 일치시킨다(어긋나면 대시보드 기준).
+- 소유자(윤찬) 단독 결정 사안을 임의로 정하지 않는다(월권 금지).
+
+## 2. 검증 기준 (프론트 변경 시)
+
+- `cd apps/web && npm run typecheck`
+- 레이아웃 변경은 `e2e/layout-visual-gate.spec.ts`(7/7) + 레이아웃 포커스 E2E +
+  `e2e/session-layout.spec.ts`(8/8) 무회귀. E2E는 web+api(+DB) 스택이 떠 있어야 한다.
+
+## 3. 커뮤니케이션
+
+- 모든 대화·주석·커밋 메시지는 한글.
+- git 커밋 메시지에 Co-Authored-By / Codex 관련 문구 추가 금지.
+
+---
+
+## 4. 이미지 생성(gpt-image-2 = imagegen2) · Live2D식 아바타
+
+> "이미지 생성해/만들어/그려줘" 요청 → `~/.Codex/skills/codex-image` 스킬이 아래 래퍼를 자동 사용.
+
+### 4.1 gpt-image-2 호출 (반드시 래퍼)
+```bash
+bash ~/.codex/imagegen-headless/codex_imagegen.sh \
+  --out <경로.png> [--size WxH] [--quality low|medium|high|auto] \
+  [-i <참조이미지> ...] [--all] "<프롬프트>"
+```
+- 인증: ChatGPT 구독 OAuth(`~/.codex/auth.json`의 `auth_mode=="chatgpt"`). **API 키 사용 금지**(과금).
+- codex 0.140+는 생성 이미지를 세션 rollout JSONL에 **base64로 인라인 반환** → 래퍼의 `extract_imagegen.py` 추출만 결정적. stdout의 "저장 경로"는 **환각**(직접 codex exec 금지).
+- 프롬프트는 **stdin 파이프**로(인자 전달 시 멈춤). 변주 생성 시 base를 `-i` 참조로 넘겨 아이덴티티·프레이밍 고정.
+- 투명배경 미지원 → 단색 평면 배경으로 생성 후 누끼. 한글 텍스트 렌더 가능(stdin 파이프라 인코딩 문제 없음).
+
+### 4.2 누끼(컷아웃)
+- `object-separation` 스킬(BiRefNet): `~/.venvs/object-separation/Scripts/python.exe ~/.agents/skills/object-separation/scripts/separate_object.py <in> <out> --model birefnet-general`
+- 알파 정제(잔류 헤이즈 제거): 임계치 `<35→0, >205→255` + 페더(`docs/avatar-art/seoyeon/publish.py` 참조).
+
+### 4.3 Live2D식 아바타(파츠 분리 리깅)
+아바타는 기본 **SVG 파라미터 리그** 또는 **래스터 파츠 분리 리깅**(`apps/web/src/components/avatar/RasterBust.tsx`)으로 렌더. 후자는 `persona.rasterArtSet` 지정 시 활성.
+- 레이어(각각 독립 opacity/교체 → 표정 중에도 깜빡임·입술싱크가 따로 움직임):
+  `base(neutral 전신)` + `upperface-<표정>(눈썹+눈)` + `eyelid-closed(깜빡임, 표정 무관)` + `mouth-<표정>` + `mouth-open(립싱크)`.
+- 파이프라인(재현 스크립트는 `docs/avatar-art/seoyeon/`):
+  1. `codex_imagegen.sh`로 base + 표정 변주(sad/tired/anxious/warm/startled/eyes-closed/speaking) 생성. 변주는 base를 `-i` 참조로, 동일 평면 배경.
+  2. BiRefNet 누끼 → `publish.py`(표준 캔버스 900×1125 정규화 + 알파 정제)로 `apps/web/public/avatar/<artSet>/` 게시.
+  3. `make-parts.py`로 특징 영역(upperface/eyelid/mouth) 크롭+페더 파츠를 `parts/` 생성(영역 상수 튜너 블럭).
+- 연결: `persona.ts`의 `AvatarPersona.rasterArtSet` / `Session.tsx`의 `PERSONA_AVATAR_LOOKS[<code>].rasterArtSet` / `RasterBust.tsx`(28표정→클러스터 매핑 포함).
+- dev 미리보기(인증 없음): `/dev/avatar-preview`(`AvatarPreview.tsx`). 스크린샷: `node apps/web/scripts/avatar-shot.mjs`(BASE_URL 환경변수로 포트 지정).
+- 실제 Live2D Cubism(`.moc3`)은 편집기 저작이 필요해 자동화 불가 → 위 레이어 합성이 실용적 대안.
+
+### 4.4 진행 중인 아바타 작업 핸드오프
+서연(P1) 아바타 작업은 **별도 세션에서 진행**. 현재 상태·남은 작업(Image #2=짧은 보브 기준 재생성 등)은
+[`docs/ops/handoff-avatar-seoyeon-2026-06-27.md`](./docs/ops/handoff-avatar-seoyeon-2026-06-27.md) 참조.

apps/api/app/config.py

@@ -40,6 +40,10 @@ def _is_allowed_local_dev_cors_origin(value: str) -> bool:
     )
 
 
+def _is_local_or_forbidden_non_dev_origin(value: str) -> bool:
+    return _is_local_url(value) and not _is_allowed_local_dev_cors_origin(value)
+
+
 class Settings(BaseSettings):
     model_config = SettingsConfigDict(
         env_file=".env",
@@ -84,6 +88,14 @@ class Settings(BaseSettings):
         default="https://api.openai.com/v1",
         validation_alias="OPENAI_BASE_URL",
     )
+    voice_poc_sample_tts_enabled: bool = Field(
+        default=False,
+        validation_alias="VIGNETTE_VOICE_POC_SAMPLE_TTS",
+    )
+    voice_poc_sample_tts_dir: str = Field(
+        default="",
+        validation_alias="VIGNETTE_VOICE_POC_SAMPLE_TTS_DIR",
+    )
 
     # ── 세션/인증 (BFF OAuth 2.1, 토큰 서버 보관) ────────
     session_secret: str = Field(
@@ -143,15 +155,27 @@ class Settings(BaseSettings):
         default="http://localhost:5173",
         validation_alias="FRONTEND_BASE_URL",
     )
+    frontend_origin_map: dict[str, str] = Field(
+        default_factory=lambda: {
+            "api-vignette.chanpaca.net": "https://vignette.chanpaca.net",
+            "api-vnet.18ka.net": "https://vnet.18ka.net",
+        },
+        validation_alias="FRONTEND_ORIGIN_MAP",
+    )
 
     # ── CORS (정적 프론트 + SSE 분리경로) ────────────────
     cors_origins: list[str] = Field(
         default=[
             "https://vignette.chanpaca.net",
+            "https://vnet.18ka.net",
             "https://vignette-b1q.pages.dev",
         ],
         validation_alias="CORS_ORIGINS",
     )
+    auth_dev_login_extra_origins: list[str] = Field(
+        default=[],
+        validation_alias="AUTH_DEV_LOGIN_EXTRA_ORIGINS",
+    )
 
     # Built-in personas are developer/bootstrap fixtures, not runtime truth.
     # Production should use approved rows from app.persona_card only.
@@ -181,6 +205,8 @@ class Settings(BaseSettings):
                 forbidden.append("AUTO_SEED_PERSONAS")
             if self.allow_seed_persona_fallback:
                 forbidden.append("ALLOW_SEED_PERSONA_FALLBACK")
+            if self.voice_poc_sample_tts_enabled:
+                forbidden.append("VIGNETTE_VOICE_POC_SAMPLE_TTS")
             if not self.oauth_google_client_id.strip():
                 forbidden.append("OAUTH_GOOGLE_CLIENT_ID")
             if not self.oauth_google_client_secret.strip():
@@ -189,11 +215,13 @@ class Settings(BaseSettings):
                 forbidden.append("SESSION_SECRET")
             if _is_local_url(self.frontend_base_url):
                 forbidden.append("FRONTEND_BASE_URL")
-            if any(
+            if any(_is_local_or_forbidden_non_dev_origin(origin) for origin in self.cors_origins):
-                _is_local_url(origin) and not _is_allowed_local_dev_cors_origin(origin)
-                for origin in self.cors_origins
-            ):
                 forbidden.append("CORS_ORIGINS")
+            if any(
+                _is_local_or_forbidden_non_dev_origin(origin)
+                for origin in self.frontend_origin_map.values()
+            ):
+                forbidden.append("FRONTEND_ORIGIN_MAP")
             if forbidden:
                 joined = ", ".join(forbidden)
                 raise ValueError(f"{joined} must be production-safe when ENVIRONMENT={self.environment}")

apps/api/app/routes/auth.py

@@ -221,9 +221,40 @@ def _is_local_origin(origin: str) -> bool:
     return host in {"localhost", "127.0.0.1", "::1"}
 
 
+def _frontend_origin_map() -> dict[str, str]:
+    mapped: dict[str, str] = {}
+    for api_host, frontend_origin in settings.frontend_origin_map.items():
+        host = (api_host or "").strip().lower().rstrip(".")
+        origin = _url_origin(frontend_origin)
+        if host and origin:
+            mapped[host] = origin
+    return mapped
+
+
+def _dev_login_extra_origins() -> set[str]:
+    return {
+        origin
+        for value in settings.auth_dev_login_extra_origins
+        if (origin := _url_origin(value))
+    }
+
+
+def _dev_login_extra_hosts() -> set[str]:
+    return {
+        host
+        for origin in _dev_login_extra_origins()
+        if (host := (urlsplit(origin).hostname or "").lower())
+    }
+
+
+def _is_dev_login_allowed_origin(origin: str) -> bool:
+    host = (urlsplit(origin).hostname or "").lower()
+    return _is_local_origin(origin) or origin in _dev_login_extra_origins() or host in _dev_login_extra_hosts()
+
+
 def _configured_frontend_origins() -> list[str]:
     origins: list[str] = []
-    for value in [settings.frontend_base_url, *settings.cors_origins]:
+    for value in [settings.frontend_base_url, *settings.frontend_origin_map.values(), *settings.cors_origins]:
         origin = _url_origin(value)
         if origin and origin not in origins:
             origins.append(origin)
@@ -243,8 +274,8 @@ def _frontend_origin_for_request(request: Request | None = None) -> str:
         forwarded_host = request.headers.get("x-forwarded-host")
         host = (forwarded_host or request.headers.get("host") or "").split(",", 1)[0].strip()
         hostname = host.rsplit(":", 1)[0].lower() if host else ""
-        if hostname == "api-vignette.chanpaca.net":
+        if mapped_origin := _frontend_origin_map().get(hostname):
-            return "https://vignette.chanpaca.net"
+            return mapped_origin
         if hostname in {"localhost", "127.0.0.1", "::1"}:
             return fallback
 
@@ -348,15 +379,24 @@ def _dev_login_available(request: Request) -> bool:
     if settings.environment != "dev" or not settings.auth_dev_login_enabled:
         return False
 
+    if _dev_login_extra_origins():
+        return True
+
+    saw_browser_origin = False
     for header_name in ("origin", "referer"):
         origin = _url_origin(request.headers.get(header_name))
-        if origin and not _is_local_origin(origin):
+        if origin:
-            return False
+            saw_browser_origin = True
+            if not _is_dev_login_allowed_origin(origin):
+                return False
+
+    if not saw_browser_origin and _dev_login_extra_origins():
+        return True
 
     forwarded_host = request.headers.get("x-forwarded-host")
     host = (forwarded_host or request.headers.get("host") or "").split(",", 1)[0].strip()
     origin = _url_origin(f"http://{host}") if host else None
-    return bool(origin and _is_local_origin(origin))
+    return bool(origin and _is_dev_login_allowed_origin(origin))
 
 
 @router.get("/config", response_model=AuthConfigResponse)

apps/api/app/routes/sessions.py

@@ -381,8 +381,14 @@ async def _load_session_or_404(
     principal: Principal,
     *,
     allow_ended: bool = False,
+    include_turn_evaluation: bool = False,
 ) -> InProcSession:
-    sess = await session_persistence.load_session(session_id, principal, allow_ended=True)
+    sess = await session_persistence.load_session(
+        session_id,
+        principal,
+        allow_ended=True,
+        include_turn_evaluation=include_turn_evaluation,
+    )
     if sess is not None:
         store.put(sess)
     elif runtime_fallback_allowed():
@@ -874,7 +880,11 @@ async def get_session_detail(
 ) -> SessionDetailResponse:
     """Return a learner-owned session with transcript for resume/history."""
     _ensure_learner(principal)
-    sess = await _load_session_or_404(session_id, principal, allow_ended=True)
+    sess = await _load_session_or_404(
+        session_id,
+        principal,
+        allow_ended=True,
+    )
     return _session_detail(sess, review_ready=await _review_ready(sess, principal))
 
 
@@ -951,7 +961,12 @@ async def get_session_review(
 ) -> SessionReviewResponse:
     """Return a learner-safe review built only from the stored session transcript."""
     _ensure_learner(principal)
-    sess = await _load_session_or_404(session_id, principal, allow_ended=True)
+    sess = await _load_session_or_404(
+        session_id,
+        principal,
+        allow_ended=True,
+        include_turn_evaluation=True,
+    )
     visible_turns = _learner_visible_turns(sess)
     hidden_turns = len(visible_turns) != len(sess.turns)
 

apps/api/app/routes/voice.py

@@ -49,11 +49,13 @@ _MAX_AUDIO_BYTES = 10 * 1024 * 1024
 async def voice_health() -> JSONResponse:
     """Return voice service readiness."""
     available = voice_service.is_available()
+    tts_provider = voice_service.tts_provider()
     body = {
         "status": "ok" if available else "degraded",
         "available": available,
         "stt_model": voice_svc.STT_MODEL,
         "tts_model": voice_svc.TTS_MODEL,
+        "tts_provider": tts_provider,
         "reason": None if available else "OPENAI_API_KEY is not configured",
     }
     return JSONResponse(body, status_code=200 if available else 503)
@@ -108,6 +110,7 @@ async def voice_ws(websocket: WebSocket) -> None:
             "session_id": session_id,
             "voice": voice_preset.openai_voice,
             "preset": voice_preset.preset,
+            "tts_provider": voice_service.tts_provider(),
             "state": "idle",
             **bind_meta,
         },
@@ -374,7 +377,12 @@ async def _run_turn_and_speak(
     # TTS speaking state comes before chunk metadata and binary audio.
     await _safe_send_json(
         websocket,
-        {"type": "state", "state": "speaking", "voice": voice_preset.openai_voice},
+        {
+            "type": "state",
+            "state": "speaking",
+            "voice": voice_preset.openai_voice,
+            "tts_provider": voice_service.tts_provider(),
+        },
     )
     try:
         n = 0

apps/api/app/services/voice.py

@@ -20,6 +20,7 @@ from __future__ import annotations
 
 import re
 from dataclasses import dataclass
+from pathlib import Path
 from typing import AsyncIterator, Optional
 
 import httpx
@@ -49,6 +50,32 @@ TTS_RESPONSE_FORMAT = "mp3"
 # End-of-turn readiness default for cascaded STT providers.
 EOT_SILENCE_THRESHOLD_MS = 1200
 
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+POC_SAMPLE_TTS_PRESET = "soft-young-fem"
+POC_SAMPLE_TTS_DEFAULT_DIR = (
+    _REPO_ROOT / "docs" / "voice-art" / "p1-seoyeon-higgs-v3-20260627"
+)
+POC_SAMPLE_TTS_CHUNK_SIZE = 4096
+_POC_SAMPLE_TTS_DEFAULT_SAMPLE = "p1_seoyeon_01_depressed_slow"
+_POC_SAMPLE_TTS_KEYWORDS: tuple[tuple[str, tuple[str, ...]], ...] = (
+    (
+        "p1_seoyeon_03_anxious_guarded",
+        ("엄마", "비밀", "말하지", "불안", "무서", "걱정", "들키", "싫", "갈래"),
+    ),
+    (
+        "p1_seoyeon_02_tired_flat",
+        ("잠", "피곤", "무거", "아무것도", "지쳐", "힘들", "에너지"),
+    ),
+    (
+        "p1_seoyeon_05_recovered_lively",
+        ("오늘은", "친구", "웃", "괜찮았", "좋았", "해냈"),
+    ),
+    (
+        "p1_seoyeon_04_rapport_relief",
+        ("괜찮", "들어", "고마", "선생님", "편해", "조금", "말해"),
+    ),
+)
+
 # OpenAI 공식 voice 풀(2026 기준): alloy, ash, ballad, coral, echo, fable,
 #   nova, onyx, sage, shimmer, verse. 페르소나 톤별로 골라 매핑한다.
 _OPENAI_VOICES = {
@@ -240,9 +267,32 @@ def assess_end_of_turn(
 class VoiceService:
     """OpenAI STT/TTS 어댑터. 앱 수명주기 동안 1 인스턴스 재사용(httpx 풀 공유)."""
 
-    def __init__(self, api_key: Optional[str] = None, base_url: Optional[str] = None) -> None:
+    def __init__(
+        self,
+        api_key: Optional[str] = None,
+        base_url: Optional[str] = None,
+        *,
+        poc_sample_tts_enabled: Optional[bool] = None,
+        environment: Optional[str] = None,
+        poc_sample_tts_dir: Optional[str | Path] = None,
+    ) -> None:
         self._api_key = (api_key if api_key is not None else settings.openai_api_key) or ""
         self._base_url = (base_url or settings.openai_base_url or OPENAI_BASE_URL).rstrip("/")
+        self._environment = environment if environment is not None else settings.environment
+        self._poc_sample_tts_enabled = (
+            bool(settings.voice_poc_sample_tts_enabled)
+            if poc_sample_tts_enabled is None
+            else bool(poc_sample_tts_enabled)
+        )
+        sample_dir_value: str | Path = (
+            poc_sample_tts_dir
+            if poc_sample_tts_dir is not None
+            else (settings.voice_poc_sample_tts_dir or POC_SAMPLE_TTS_DEFAULT_DIR)
+        )
+        sample_dir = Path(sample_dir_value)
+        if not sample_dir.is_absolute():
+            sample_dir = _REPO_ROOT / sample_dir
+        self._poc_sample_tts_dir = sample_dir
         self._client: Optional[httpx.AsyncClient] = None
 
     # ── 수명주기 ──────────────────────────────────────────
@@ -264,6 +314,35 @@ class VoiceService:
         """음성 기능 가용 여부(키 설정됨). 라우트가 핸드셰이크에서 검사."""
         return bool(self._api_key)
 
+    def tts_provider(self) -> str:
+        if self._poc_sample_tts_available():
+            return "p1-sample-poc"
+        if self._api_key:
+            return "openai"
+        if self._poc_sample_tts_enabled and self._environment != "dev":
+            return "disabled-non-dev"
+        return "unavailable"
+
+    def poc_sample_tts_available(self) -> bool:
+        return self._poc_sample_tts_available()
+
+    def _poc_sample_tts_available(self) -> bool:
+        return (
+            self._poc_sample_tts_enabled
+            and self._environment == "dev"
+            and self._poc_sample_path(_POC_SAMPLE_TTS_DEFAULT_SAMPLE).is_file()
+        )
+
+    def _should_use_poc_sample_tts(self, voice: VoicePreset) -> bool:
+        return (
+            self._poc_sample_tts_enabled
+            and self._environment == "dev"
+            and voice.preset == POC_SAMPLE_TTS_PRESET
+        )
+
+    def _poc_sample_path(self, sample_id: str) -> Path:
+        return self._poc_sample_tts_dir / f"{sample_id}.mp3"
+
     @property
     def _http(self) -> httpx.AsyncClient:
         if not self._api_key:
@@ -341,6 +420,10 @@ class VoiceService:
         text = speakable_text(text)
         if not text:
             return
+        if self._should_use_poc_sample_tts(voice):
+            async for chunk in self._synthesize_poc_sample_tts(text):
+                yield chunk
+            return
         payload = build_tts_payload(
             text,
             voice,
@@ -375,6 +458,25 @@ class VoiceService:
         except httpx.HTTPError as e:
             raise RuntimeError(f"TTS transport error: {e}") from e
 
+    async def _synthesize_poc_sample_tts(self, text: str) -> AsyncIterator[TTSChunk]:
+        sample_id = self._select_poc_sample_id(text)
+        sample_path = self._poc_sample_path(sample_id)
+        try:
+            data = sample_path.read_bytes()
+        except OSError as e:
+            raise VoiceUnavailable(f"P1 sample TTS asset is missing: {sample_path}") from e
+        for i in range(0, len(data), POC_SAMPLE_TTS_CHUNK_SIZE):
+            chunk = data[i : i + POC_SAMPLE_TTS_CHUNK_SIZE]
+            if chunk:
+                yield TTSChunk(audio=chunk)
+
+    def _select_poc_sample_id(self, text: str) -> str:
+        normalized = text.casefold()
+        for sample_id, keywords in _POC_SAMPLE_TTS_KEYWORDS:
+            if any(keyword.casefold() in normalized for keyword in keywords):
+                return sample_id
+        return _POC_SAMPLE_TTS_DEFAULT_SAMPLE
+
     async def _synthesize_fallback(self, payload: dict[str, object]) -> AsyncIterator[TTSChunk]:
         """tts-1 폴백(비스트림 POST → 전체 바이트를 청크로 분할)."""
         try:

apps/api/app/session_persistence.py

@@ -18,6 +18,11 @@ from .store import DEFAULT_TURN_VISIBLE_TO, InProcSession, TurnRecord
 
 _EVALUATION_CACHE: dict[str, dict[str, Any]] = {}
 _SESSION_AUDIT_ROLES = {"teacher", "admin"}
+_APPROPRIATENESS_SCORE = {
+    "warn": 1.0,
+    "neutral": 3.0,
+    "pos": 5.0,
+}
 
 
 _JOINED_CARD_COLUMNS = (
@@ -71,6 +76,248 @@ def _stage(stage: object) -> str:
     return getattr(stage, "value", str(stage))
 
 
+def _clean_text(value: Any) -> str | None:
+    if value is None:
+        return None
+    text = str(value).strip()
+    return text or None
+
+
+def _safe_float(value: Any) -> float | None:
+    if isinstance(value, (int, float)):
+        return float(value)
+    return None
+
+
+def _evaluation_loop(evaluation: dict[str, Any]) -> str:
+    loop = _clean_text(evaluation.get("loop")) or "fast"
+    return loop if loop in {"fast", "deep"} else "fast"
+
+
+def _dict_items(value: Any) -> list[dict[str, Any]]:
+    if not isinstance(value, list):
+        return []
+    return [item for item in value if isinstance(item, dict)]
+
+
+def _evaluation_feedback_rows(evaluation: dict[str, Any] | None) -> list[dict[str, Any]]:
+    """Normalize scalar/rationale turn-evaluation fields into feedback_scores rows."""
+    if not isinstance(evaluation, dict):
+        return []
+    loop = _evaluation_loop(evaluation)
+    rows: list[dict[str, Any]] = []
+
+    def add(
+        dimension: str,
+        *,
+        score: float | None = None,
+        rationale: str | None = None,
+        top1_score: float | None = None,
+    ) -> None:
+        dim = _clean_text(dimension)
+        if not dim:
+            return
+        rows.append(
+            {
+                "dimension": dim,
+                "score": score,
+                "rationale": rationale,
+                "top1_score": top1_score,
+                "loop": loop,
+            }
+        )
+
+    appropriateness = _clean_text(evaluation.get("appropriateness")) or "neutral"
+    if appropriateness not in _APPROPRIATENESS_SCORE:
+        appropriateness = "neutral"
+    add(
+        "appropriateness",
+        score=_APPROPRIATENESS_SCORE[appropriateness],
+        rationale=_clean_text(evaluation.get("appropriateness_note")),
+    )
+
+    rapport = _safe_float(evaluation.get("rapport_signal"))
+    if rapport is not None:
+        add("rapport_signal", score=max(-1.0, min(1.0, rapport)))
+
+    theory_mode = _clean_text(evaluation.get("theory_mode"))
+    if theory_mode:
+        add("theory_mode", rationale=theory_mode)
+
+    error = _clean_text(evaluation.get("error"))
+    if error:
+        add("error", rationale=error)
+
+    for tag in _dict_items(evaluation.get("techniques")):
+        code = _clean_text(tag.get("code"))
+        rationale = _clean_text(tag.get("rationale"))
+        if code and rationale:
+            add(f"technique:{code}", rationale=rationale)
+
+    for state in _dict_items(evaluation.get("client_state_read")):
+        code = _clean_text(state.get("code"))
+        rationale = _clean_text(state.get("rationale"))
+        if code and rationale:
+            add(f"client_state:{code}", rationale=rationale)
+
+    return rows
+
+
+def _evaluation_technique_rows(evaluation: dict[str, Any] | None) -> list[dict[str, str]]:
+    if not isinstance(evaluation, dict):
+        return []
+    rows: list[dict[str, str]] = []
+    for tag in _dict_items(evaluation.get("techniques")):
+        code = _clean_text(tag.get("code"))
+        if not code:
+            continue
+        rows.append(
+            {
+                "code": code,
+                "label_ko": _clean_text(tag.get("label_ko")) or code,
+                "category": _clean_text(tag.get("category")) or "",
+            }
+        )
+    return rows
+
+
+def _evaluation_client_state_rows(evaluation: dict[str, Any] | None) -> list[dict[str, str]]:
+    if not isinstance(evaluation, dict):
+        return []
+    rows: list[dict[str, str]] = []
+    for state in _dict_items(evaluation.get("client_state_read")):
+        code = _clean_text(state.get("code"))
+        if not code:
+            continue
+        rows.append(
+            {
+                "code": code,
+                "label_ko": _clean_text(state.get("label_ko")) or code,
+            }
+        )
+    return rows
+
+
+def _evaluation_comment_rows(evaluation: dict[str, Any] | None) -> list[dict[str, Any]]:
+    if not isinstance(evaluation, dict):
+        return []
+    deviation = evaluation.get("intent_deviation")
+    if not isinstance(deviation, dict):
+        return []
+    note = _clean_text(evaluation.get("appropriateness_note")) or "의도와 다른 부분"
+    return [{"kind": "critique", "text": note, "intent_deviation": deviation}]
+
+
+def _appropriateness_from_score(score: Any) -> str:
+    value = _safe_float(score)
+    if value is None:
+        return "neutral"
+    if value >= 4.0:
+        return "pos"
+    if value <= 2.0:
+        return "warn"
+    return "neutral"
+
+
+def _base_turn_evaluation(turn_seq: int, stage: str) -> dict[str, Any]:
+    return {
+        "loop": "fast",
+        "turn_seq": turn_seq,
+        "stage": stage,
+        "techniques": [],
+        "client_state_read": [],
+        "appropriateness": "neutral",
+    }
+
+
+def _rebuild_turn_evaluations(
+    turn_refs: list[tuple[str, int, str]],
+    *,
+    feedback_rows: Iterable[Any],
+    technique_rows: Iterable[Any],
+    client_state_rows: Iterable[Any],
+    comment_rows: Iterable[Any],
+) -> dict[str, dict[str, Any]]:
+    """Rehydrate normalized DB rows back into the TurnRecord.evaluation shape."""
+    refs = {turn_id: (turn_seq, stage) for turn_id, turn_seq, stage in turn_refs}
+    evaluations: dict[str, dict[str, Any]] = {}
+    rationale_by_dimension: dict[str, dict[str, str]] = {}
+
+    def ensure(turn_id: str) -> dict[str, Any]:
+        if turn_id not in evaluations:
+            turn_seq, stage = refs[turn_id]
+            evaluations[turn_id] = _base_turn_evaluation(turn_seq, stage)
+        return evaluations[turn_id]
+
+    for row in feedback_rows:
+        turn_id = str(row["turn_id"])
+        if turn_id not in refs:
+            continue
+        ev = ensure(turn_id)
+        loop = _clean_text(row["loop"])
+        if loop in {"fast", "deep"}:
+            ev["loop"] = loop
+        dimension = _clean_text(row["dimension"]) or ""
+        rationale = _clean_text(row["rationale"])
+        if rationale:
+            rationale_by_dimension.setdefault(turn_id, {})[dimension] = rationale
+        if dimension == "appropriateness":
+            ev["appropriateness"] = _appropriateness_from_score(row["score"])
+            if rationale:
+                ev["appropriateness_note"] = rationale
+        elif dimension == "rapport_signal":
+            score = _safe_float(row["score"])
+            if score is not None:
+                ev["rapport_signal"] = max(-1.0, min(1.0, score))
+        elif dimension == "theory_mode" and rationale:
+            ev["theory_mode"] = rationale
+        elif dimension == "error" and rationale:
+            ev["error"] = rationale
+
+    for row in technique_rows:
+        turn_id = str(row["turn_id"])
+        if turn_id not in refs:
+            continue
+        code = _clean_text(row["code"])
+        if not code:
+            continue
+        item = {
+            "code": code,
+            "label_ko": _clean_text(row["label_ko"]) or code,
+            "category": _clean_text(row["category"]) or "",
+        }
+        rationale = rationale_by_dimension.get(turn_id, {}).get(f"technique:{code}")
+        if rationale:
+            item["rationale"] = rationale
+        ensure(turn_id)["techniques"].append(item)
+
+    for row in client_state_rows:
+        turn_id = str(row["turn_id"])
+        if turn_id not in refs:
+            continue
+        code = _clean_text(row["code"])
+        if not code:
+            continue
+        item = {
+            "code": code,
+            "label_ko": _clean_text(row["label_ko"]) or code,
+        }
+        rationale = rationale_by_dimension.get(turn_id, {}).get(f"client_state:{code}")
+        if rationale:
+            item["rationale"] = rationale
+        ensure(turn_id)["client_state_read"].append(item)
+
+    for row in comment_rows:
+        turn_id = str(row["turn_id"])
+        if turn_id not in refs:
+            continue
+        deviation = row["intent_deviation"]
+        if isinstance(deviation, dict):
+            ensure(turn_id)["intent_deviation"] = deviation
+
+    return evaluations
+
+
 async def _record_session_read_audit(
     conn: Any,
     principal: Principal,
@@ -113,7 +360,7 @@ def _state_from_row(row, card: PersonaCard) -> state_machine.SessionState:
     )
 
 
-def _turn_from_row(row) -> TurnRecord:
+def _turn_from_row(row, evaluation: dict[str, Any] | None = None) -> TurnRecord:
     created_at = _ts(row["created_at"]) or time.time()
     return TurnRecord(
         turn_seq=int(row["seq"]),
@@ -121,6 +368,7 @@ def _turn_from_row(row) -> TurnRecord:
         stage=row["stage"],
         text=row["text"] or row["text_masked"] or "",
         text_masked=row["text_masked"] or row["text"] or "",
+        turn_id=str(_row_value(row, "id")) if _row_value(row, "id") is not None else None,
         created_at=created_at,
         llm_provider=_row_value(row, "llm_provider"),
         model=_row_value(row, "model"),
@@ -131,10 +379,179 @@ def _turn_from_row(row) -> TurnRecord:
         silence_ms=_row_value(row, "silence_ms"),
         speech_rate=_row_value(row, "speech_rate"),
         barge_in=_row_value(row, "barge_in"),
+        evaluation=evaluation,
         visible_to=tuple(_row_value(row, "visible_to") or DEFAULT_TURN_VISIBLE_TO),
     )
 
 
+async def _persist_turn_evaluation(conn: Any, turn_id: str, evaluation: dict[str, Any] | None) -> None:
+    if not isinstance(evaluation, dict):
+        return
+    await conn.execute("SELECT set_config('app.ai_context', '1', true)")
+    await conn.execute("SELECT set_config('app.current_ai_view', 'evaluator', true)")
+    await conn.execute("SELECT set_config('app.current_sens_max', '2', true)")
+
+    for row in _evaluation_feedback_rows(evaluation):
+        await conn.execute(
+            """
+            INSERT INTO app.feedback_scores (
+              turn_id, dimension, score, rationale, top1_score, loop
+            )
+            VALUES ($1::uuid, $2, $3, $4, $5, $6)
+            ON CONFLICT (turn_id, dimension) DO UPDATE SET
+              score = EXCLUDED.score,
+              rationale = EXCLUDED.rationale,
+              top1_score = EXCLUDED.top1_score,
+              loop = EXCLUDED.loop
+            """,
+            turn_id,
+            row["dimension"],
+            row["score"],
+            row["rationale"],
+            row["top1_score"],
+            row["loop"],
+        )
+
+    for row in _evaluation_technique_rows(evaluation):
+        label_id = await conn.fetchval(
+            """
+            INSERT INTO app.technique_label_def (code, display_name, category)
+            VALUES ($1, $2, $3)
+            ON CONFLICT (code, version) DO UPDATE SET
+              display_name = EXCLUDED.display_name,
+              category = EXCLUDED.category,
+              is_active = TRUE
+            RETURNING label_id
+            """,
+            row["code"],
+            row["label_ko"],
+            row["category"],
+        )
+        await conn.execute(
+            """
+            INSERT INTO app.turn_technique (turn_id, label_id)
+            VALUES ($1::uuid, $2)
+            ON CONFLICT DO NOTHING
+            """,
+            turn_id,
+            label_id,
+        )
+
+    for row in _evaluation_client_state_rows(evaluation):
+        label_id = await conn.fetchval(
+            """
+            INSERT INTO app.client_state_def (code, display_name)
+            VALUES ($1, $2)
+            ON CONFLICT (code, version) DO UPDATE SET
+              display_name = EXCLUDED.display_name,
+              is_active = TRUE
+            RETURNING label_id
+            """,
+            row["code"],
+            row["label_ko"],
+        )
+        await conn.execute(
+            """
+            INSERT INTO app.turn_client_state (turn_id, label_id)
+            VALUES ($1::uuid, $2)
+            ON CONFLICT DO NOTHING
+            """,
+            turn_id,
+            label_id,
+        )
+
+    for row in _evaluation_comment_rows(evaluation):
+        await conn.execute(
+            """
+            INSERT INTO app.supervisor_comment (
+              turn_id, kind, text, intent_deviation
+            )
+            VALUES ($1::uuid, $2, $3, $4::jsonb)
+            """,
+            turn_id,
+            row["kind"],
+            row["text"],
+            row["intent_deviation"],
+        )
+
+
+async def _load_turn_evaluations(
+    conn: Any,
+    turn_refs: list[tuple[str, int, str]],
+) -> dict[str, dict[str, Any]]:
+    turn_ids = [turn_id for turn_id, _, _ in turn_refs]
+    if not turn_ids:
+        return {}
+    feedback_rows = await conn.fetch(
+        """
+        SELECT turn_id::text AS turn_id, dimension, score, rationale, top1_score, loop
+        FROM app.feedback_scores
+        WHERE turn_id = ANY($1::uuid[])
+        ORDER BY created_at, dimension
+        """,
+        turn_ids,
+    )
+    technique_rows = await conn.fetch(
+        """
+        SELECT
+          tt.turn_id::text AS turn_id,
+          d.code,
+          d.display_name AS label_ko,
+          d.category
+        FROM app.turn_technique tt
+        JOIN app.technique_label_def d ON d.label_id = tt.label_id
+        WHERE tt.turn_id = ANY($1::uuid[])
+        ORDER BY tt.turn_id, d.code
+        """,
+        turn_ids,
+    )
+    client_state_rows = await conn.fetch(
+        """
+        SELECT
+          ts.turn_id::text AS turn_id,
+          d.code,
+          d.display_name AS label_ko
+        FROM app.turn_client_state ts
+        JOIN app.client_state_def d ON d.label_id = ts.label_id
+        WHERE ts.turn_id = ANY($1::uuid[])
+        ORDER BY ts.turn_id, d.code
+        """,
+        turn_ids,
+    )
+    comment_rows = await conn.fetch(
+        """
+        SELECT turn_id::text AS turn_id, intent_deviation
+        FROM app.supervisor_comment
+        WHERE turn_id = ANY($1::uuid[])
+          AND intent_deviation IS NOT NULL
+        ORDER BY created_at
+        """,
+        turn_ids,
+    )
+    return _rebuild_turn_evaluations(
+        turn_refs,
+        feedback_rows=feedback_rows,
+        technique_rows=technique_rows,
+        client_state_rows=client_state_rows,
+        comment_rows=comment_rows,
+    )
+
+
+async def _hydrate_session_turn_evaluations(sess: InProcSession) -> None:
+    turn_refs = [
+        (turn.turn_id, turn.turn_seq, turn.stage)
+        for turn in sess.turns
+        if turn.turn_id is not None
+    ]
+    if not turn_refs:
+        return
+    async with acquire(ai_view="evaluator") as conn:
+        evaluations = await _load_turn_evaluations(conn, turn_refs)
+    for turn in sess.turns:
+        if turn.turn_id and turn.turn_id in evaluations:
+            turn.evaluation = evaluations[turn.turn_id]
+
+
 async def ensure_review_tables() -> None:
     """Create runtime review/evaluation storage when the DB role allows it."""
     try:
@@ -441,6 +858,7 @@ async def load_session(
     principal: Principal,
     *,
     allow_ended: bool = False,
+    include_turn_evaluation: bool = False,
 ) -> InProcSession | None:
     try:
         get_pool()
@@ -496,7 +914,7 @@ async def load_session(
             )
             turn_rows = await conn.fetch(
                 """
-                SELECT seq, speaker, stage, text, text_masked, created_at,
+                SELECT id, seq, speaker, stage, text, text_masked, created_at,
                        llm_provider, model, tokens_in, tokens_out, cost_usd,
                        audio_ref, silence_ms, speech_rate, barge_in, visible_to
                 FROM app.turns
@@ -518,7 +936,9 @@ async def load_session(
                         "learner_id": sess.learner_id,
                     },
                 )
-            return sess
+        if sess is not None and include_turn_evaluation:
+            await _hydrate_session_turn_evaluations(sess)
+        return sess
     except Exception:
         require_runtime_fallback_allowed("session load")
         return None
@@ -546,7 +966,7 @@ async def append_turn(
                 )
                 or 1
             )
-            await conn.execute(
+            inserted_turn_id = await conn.fetchval(
                 """
                 INSERT INTO app.turns (
                   session_id, seq, speaker, stage, text, text_masked, actor_kind,
@@ -559,6 +979,7 @@ async def append_turn(
                   $13, $14, $15, $16, $17::text[]
                 )
                 ON CONFLICT (session_id, seq) DO NOTHING
+                RETURNING id
                 """,
                 session_id,
                 seq,
@@ -578,6 +999,11 @@ async def append_turn(
                 turn.barge_in,
                 list(turn.visible_to or DEFAULT_TURN_VISIBLE_TO),
             )
+            if inserted_turn_id is None:
+                return False
+            turn.turn_id = str(inserted_turn_id)
+            # 원시 평가 row는 학습자 축어록이 아니라 evaluator 전용 데이터다.
+            await _persist_turn_evaluation(conn, turn.turn_id, turn.evaluation)
         return True
     except Exception:
         require_runtime_fallback_allowed("session turn append")
@@ -703,7 +1129,7 @@ async def list_sessions(principal: Principal) -> tuple[list[InProcSession], bool
                 )
                 turn_rows = await conn.fetch(
                     """
-                    SELECT seq, speaker, stage, text, text_masked, created_at,
+                    SELECT id, seq, speaker, stage, text, text_masked, created_at,
                            llm_provider, model, tokens_in, tokens_out, cost_usd,
                            audio_ref, silence_ms, speech_rate, barge_in, visible_to
                     FROM app.turns

apps/api/app/store.py

@@ -31,6 +31,7 @@ class TurnRecord:
     stage: str
     text: str                    # 원문(개발용; 실제 저장은 마스킹본)
     text_masked: str
+    turn_id: str | None = None    # DB app.turns.id. 런타임 폴백 턴은 None.
     created_at: float = field(default_factory=time.time)
     llm_provider: str | None = None
     model: str | None = None

apps/api/app/test_auth_providers.py

@@ -29,13 +29,13 @@ def patched_settings(**values: Any):
             setattr(settings, key, value)
 
 
-def _request() -> Request:
+def _request(headers: list[tuple[bytes, bytes]] | None = None) -> Request:
     return Request(
         {
             "type": "http",
             "method": "GET",
             "path": "/auth/login",
-            "headers": [(b"host", b"localhost:8000")],
+            "headers": headers or [(b"host", b"localhost:8000")],
         }
     )
 
@@ -111,6 +111,64 @@ class AuthProviderScaffoldTest(unittest.IsolatedAsyncioTestCase):
         self.assertTrue(providers["saml"].enabled)
         self.assertEqual(providers["saml"].login_path, "/auth/login?provider=saml")
 
+    async def test_auth_config_allows_dev_login_from_configured_tailnet_origin(self) -> None:
+        request = _request(
+            [
+                (b"host", b"127.0.0.1:8000"),
+                (b"origin", b"https://alpaca-home.taile93291.ts.net"),
+            ]
+        )
+
+        with patched_settings(
+            environment="dev",
+            auth_dev_login_enabled=True,
+            auth_dev_login_extra_origins=["https://alpaca-home.taile93291.ts.net"],
+        ):
+            config = await auth_routes.auth_config(request)
+
+        self.assertTrue(config.dev_login_enabled)
+
+    async def test_auth_config_allows_dev_login_from_configured_tailnet_forwarded_host(self) -> None:
+        request = _request(
+            [
+                (b"host", b"127.0.0.1:8000"),
+                (b"x-forwarded-host", b"alpaca-home.taile93291.ts.net"),
+            ]
+        )
+
+        with patched_settings(
+            environment="dev",
+            auth_dev_login_enabled=True,
+            auth_dev_login_extra_origins=["https://alpaca-home.taile93291.ts.net"],
+        ):
+            config = await auth_routes.auth_config(request)
+
+        self.assertTrue(config.dev_login_enabled)
+
+    async def test_auth_config_allows_dev_login_config_probe_without_origin_when_extra_origin_is_set(self) -> None:
+        request = _request([(b"host", b"127.0.0.1:8000")])
+
+        with patched_settings(
+            environment="dev",
+            auth_dev_login_enabled=True,
+            auth_dev_login_extra_origins=["https://alpaca-home.taile93291.ts.net"],
+        ):
+            config = await auth_routes.auth_config(request)
+
+        self.assertTrue(config.dev_login_enabled)
+
+    async def test_frontend_origin_map_routes_vnet_api_callbacks_to_vnet_frontend(self) -> None:
+        request = _request([(b"host", b"api-vnet.18ka.net")])
+
+        with patched_settings(
+            frontend_base_url="https://vignette.chanpaca.net",
+            frontend_origin_map={"api-vnet.18ka.net": "https://vnet.18ka.net"},
+            cors_origins=["https://vignette.chanpaca.net", "https://vnet.18ka.net"],
+        ):
+            origin = auth_routes._frontend_origin_for_request(request)
+
+        self.assertEqual(origin, "https://vnet.18ka.net")
+
     async def test_saml_login_builds_redirect_authn_request_and_relay_state(self) -> None:
         with patched_settings(
             auth_saml_enabled=True,

apps/api/app/test_evaluation_persistence.py

@@ -0,0 +1,247 @@
+"""Regression tests for turn-evaluation persistence mapping."""
+
+from __future__ import annotations
+
+from pathlib import Path
+import unittest
+from unittest.mock import patch
+
+from .deps import Principal, Role
+from . import session_persistence
+from .routes import sessions
+from .services import persona as persona_service
+from .services import state_machine
+from .store import InProcSession
+
+
+class FakeEvaluationConn:
+    def __init__(self) -> None:
+        self.executed: list[tuple[str, tuple[object, ...]]] = []
+        self.fetchvals: list[tuple[str, tuple[object, ...]]] = []
+
+    async def execute(self, query: str, *args: object) -> str:
+        self.executed.append((query, args))
+        return "INSERT 0 1"
+
+    async def fetchval(self, query: str, *args: object) -> int:
+        self.fetchvals.append((query, args))
+        if "app.technique_label_def" in query:
+            return 101
+        if "app.client_state_def" in query:
+            return 202
+        raise AssertionError(f"unexpected fetchval query: {query}")
+
+
+class EvaluationPersistenceMappingTest(unittest.TestCase):
+    def test_feedback_rows_preserve_review_scalar_contract(self) -> None:
+        evaluation = {
+            "loop": "fast",
+            "turn_seq": 2,
+            "stage": "탐색",
+            "appropriateness": "pos",
+            "appropriateness_note": "정서를 먼저 반영했다.",
+            "rapport_signal": 0.75,
+            "theory_mode": "humanistic",
+            "techniques": [
+                {
+                    "code": "empathy",
+                    "label_ko": "공감",
+                    "category": "relational",
+                    "rationale": "감정을 명시적으로 반영했다.",
+                }
+            ],
+            "client_state_read": [
+                {
+                    "code": "affect_contact",
+                    "label_ko": "정서 접촉/표현",
+                    "rationale": "내담자가 감정을 언급했다.",
+                }
+            ],
+        }
+
+        rows = {
+            row["dimension"]: row
+            for row in session_persistence._evaluation_feedback_rows(evaluation)
+        }
+
+        self.assertEqual(rows["appropriateness"]["score"], 5.0)
+        self.assertEqual(rows["appropriateness"]["rationale"], "정서를 먼저 반영했다.")
+        self.assertEqual(rows["rapport_signal"]["score"], 0.75)
+        self.assertEqual(rows["theory_mode"]["rationale"], "humanistic")
+        self.assertEqual(rows["technique:empathy"]["rationale"], "감정을 명시적으로 반영했다.")
+        self.assertEqual(rows["client_state:affect_contact"]["rationale"], "내담자가 감정을 언급했다.")
+
+    def test_rebuild_turn_evaluation_restores_review_shape(self) -> None:
+        rebuilt = session_persistence._rebuild_turn_evaluations(
+            [("11111111-1111-1111-1111-111111111111", 2, "탐색")],
+            feedback_rows=[
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "dimension": "appropriateness",
+                    "score": 1.0,
+                    "rationale": "조언이 너무 빨랐다.",
+                    "top1_score": None,
+                    "loop": "fast",
+                },
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "dimension": "technique:empathy",
+                    "score": None,
+                    "rationale": "정서 반영이 포함됐다.",
+                    "top1_score": None,
+                    "loop": "fast",
+                },
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "dimension": "rapport_signal",
+                    "score": -0.4,
+                    "rationale": None,
+                    "top1_score": None,
+                    "loop": "fast",
+                },
+            ],
+            technique_rows=[
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "code": "empathy",
+                    "label_ko": "공감",
+                    "category": "relational",
+                }
+            ],
+            client_state_rows=[
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "code": "defensive",
+                    "label_ko": "방어",
+                }
+            ],
+            comment_rows=[
+                {
+                    "turn_id": "11111111-1111-1111-1111-111111111111",
+                    "intent_deviation": {
+                        "dimension": "pacing",
+                        "expected": "감정 탐색",
+                        "actual": "해결 조언",
+                        "severity": "moderate",
+                    },
+                }
+            ],
+        )
+
+        ev = rebuilt["11111111-1111-1111-1111-111111111111"]
+        self.assertEqual(ev["turn_seq"], 2)
+        self.assertEqual(ev["stage"], "탐색")
+        self.assertEqual(ev["appropriateness"], "warn")
+        self.assertEqual(ev["appropriateness_note"], "조언이 너무 빨랐다.")
+        self.assertEqual(ev["rapport_signal"], -0.4)
+        self.assertEqual(ev["techniques"][0]["rationale"], "정서 반영이 포함됐다.")
+        self.assertEqual(ev["client_state_read"][0]["label_ko"], "방어")
+        self.assertEqual(ev["intent_deviation"]["dimension"], "pacing")
+
+    def test_evaluation_rls_blocks_raw_learner_writes(self) -> None:
+        root = Path(__file__).resolve().parents[3]
+        sql = (root / "infra/db/init/04_audit_eval_rls.sql").read_text(encoding="utf-8")
+
+        self.assertIn("ALTER TABLE app.feedback_scores ENABLE ROW LEVEL SECURITY", sql)
+        self.assertIn("ALTER TABLE app.turn_technique ENABLE ROW LEVEL SECURITY", sql)
+        self.assertIn("ALTER TABLE app.turn_client_state ENABLE ROW LEVEL SECURITY", sql)
+        self.assertIn("ALTER TABLE app.supervisor_comment ENABLE ROW LEVEL SECURITY", sql)
+        self.assertIn("ALTER TABLE app.alternative_utterance ENABLE ROW LEVEL SECURITY", sql)
+        feedback_insert = sql.split("CREATE POLICY p_feedback_insert", 1)[1].split(");", 1)[0]
+        self.assertNotIn("learner_id = app.current_uid()", feedback_insert)
+
+    def test_append_turn_requires_inserted_turn_id(self) -> None:
+        source = Path(session_persistence.__file__).read_text(encoding="utf-8")
+
+        self.assertIn("RETURNING id", source)
+        self.assertIn("if inserted_turn_id is None:", source)
+
+
+class EvaluationPersistenceIOTest(unittest.IsolatedAsyncioTestCase):
+    async def test_persist_turn_evaluation_uses_evaluator_context_and_real_fast_tables(self) -> None:
+        conn = FakeEvaluationConn()
+        evaluation = {
+            "loop": "fast",
+            "turn_seq": 3,
+            "stage": "탐색",
+            "appropriateness": "warn",
+            "appropriateness_note": "해결 제안이 빨랐다.",
+            "techniques": [
+                {
+                    "code": "empathy",
+                    "label_ko": "공감",
+                    "category": "relational",
+                    "rationale": "정서 반영.",
+                }
+            ],
+            "client_state_read": [
+                {
+                    "code": "defensive",
+                    "label_ko": "방어",
+                    "rationale": "짧은 회피 반응.",
+                }
+            ],
+            "intent_deviation": {
+                "dimension": "pacing",
+                "expected": "탐색",
+                "actual": "조언",
+                "severity": "minor",
+            },
+        }
+
+        await session_persistence._persist_turn_evaluation(
+            conn,
+            "11111111-1111-1111-1111-111111111111",
+            evaluation,
+        )
+
+        executed_sql = "\n".join(query for query, _ in conn.executed)
+        self.assertIn("set_config('app.ai_context', '1', true)", executed_sql)
+        self.assertIn("set_config('app.current_ai_view', 'evaluator', true)", executed_sql)
+        self.assertIn("INSERT INTO app.feedback_scores", executed_sql)
+        self.assertIn("INSERT INTO app.turn_technique", executed_sql)
+        self.assertIn("INSERT INTO app.turn_client_state", executed_sql)
+        self.assertIn("INSERT INTO app.supervisor_comment", executed_sql)
+        self.assertNotIn("app.alternative_utterance", executed_sql)
+
+    async def test_route_loader_only_hydrates_when_requested(self) -> None:
+        principal = Principal(
+            user_id="00000000-0000-0000-0000-000000000101",
+            role=Role.LEARNER,
+            cohort_ids=[],
+            email="eval-map@hs.ac.kr",
+            display_name="Eval Map",
+        )
+        card = persona_service.P1
+        sess = InProcSession(
+            session_id="eval-map-session",
+            case_id="eval-map-case",
+            learner_id=principal.user_id,
+            persona_code=card.code,
+            theory_mode="humanistic",
+            persona=card,
+            state=state_machine.SessionState(
+                resistance=card.base_resistance(),
+                ideation_stage=card.ideation_baseline(),
+            ),
+        )
+        calls: list[bool] = []
+
+        async def fake_load_session(*args, **kwargs):
+            calls.append(bool(kwargs.get("include_turn_evaluation")))
+            return sess
+
+        with patch.object(sessions.session_persistence, "load_session", fake_load_session):
+            await sessions._load_session_or_404(sess.session_id, principal)
+            await sessions._load_session_or_404(
+                sess.session_id,
+                principal,
+                allow_ended=True,
+                include_turn_evaluation=True,
+            )
+
+        self.assertEqual(calls, [False, True])
+
+
+if __name__ == "__main__":
+    unittest.main()

apps/api/app/test_runtime_policy.py

@@ -222,11 +222,13 @@ class RuntimeFallbackPolicyTest(unittest.IsolatedAsyncioTestCase):
                 auth_dev_login_enabled=True,
                 auto_seed_personas=True,
                 allow_seed_persona_fallback=True,
+                voice_poc_sample_tts_enabled=True,
             )
 
         self.assertIn("AUTH_DEV_LOGIN_ENABLED", str(caught.exception))
         self.assertIn("AUTO_SEED_PERSONAS", str(caught.exception))
         self.assertIn("ALLOW_SEED_PERSONA_FALLBACK", str(caught.exception))
+        self.assertIn("VIGNETTE_VOICE_POC_SAMPLE_TTS", str(caught.exception))
 
     def test_non_dev_rejects_missing_public_runtime_config(self) -> None:
         with self.assertRaises(ValueError) as caught:
@@ -264,6 +266,43 @@ class RuntimeFallbackPolicyTest(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(cfg.environment, "staging")
         self.assertEqual(cfg.cors_origins, ["https://vignette.chanpaca.net"])
 
+    def test_non_dev_accepts_vnet_frontend_origin_map(self) -> None:
+        cfg = Settings(
+            environment="prod",
+            auth_dev_login_enabled=False,
+            auto_seed_personas=False,
+            allow_seed_persona_fallback=False,
+            session_secret="prod-secret-change-me",
+            oauth_google_client_id="google-client-id",
+            oauth_google_client_secret="google-client-secret",
+            frontend_base_url="https://vignette.chanpaca.net",
+            frontend_origin_map={
+                "api-vignette.chanpaca.net": "https://vignette.chanpaca.net",
+                "api-vnet.18ka.net": "https://vnet.18ka.net",
+            },
+            cors_origins=["https://vignette.chanpaca.net", "https://vnet.18ka.net"],
+        )
+
+        self.assertEqual(cfg.frontend_origin_map["api-vnet.18ka.net"], "https://vnet.18ka.net")
+        self.assertIn("https://vnet.18ka.net", cfg.cors_origins)
+
+    def test_non_dev_rejects_local_frontend_origin_map(self) -> None:
+        with self.assertRaises(ValueError) as caught:
+            Settings(
+                environment="prod",
+                auth_dev_login_enabled=False,
+                auto_seed_personas=False,
+                allow_seed_persona_fallback=False,
+                session_secret="prod-secret-change-me",
+                oauth_google_client_id="google-client-id",
+                oauth_google_client_secret="google-client-secret",
+                frontend_base_url="https://vignette.chanpaca.net",
+                frontend_origin_map={"api.example.test": "http://localhost:9999"},
+                cors_origins=["https://vignette.chanpaca.net"],
+            )
+
+        self.assertIn("FRONTEND_ORIGIN_MAP", str(caught.exception))
+
     def test_non_dev_accepts_explicit_local_vite_cors_ports(self) -> None:
         cfg = Settings(
             environment="prod",

apps/api/app/test_voice_service.py

@@ -3,6 +3,8 @@
 from __future__ import annotations
 
 import unittest
+from pathlib import Path
+from tempfile import TemporaryDirectory
 
 from .services.voice import (
     DEFAULT_OPENAI_VOICE,
@@ -173,6 +175,49 @@ class VoiceServiceStreamTest(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(payload["instructions"], "Keep the tone grounded.")
         self.assertEqual([chunk.audio for chunk in chunks], [b"\x80\x80", b"\xff\x00"])
 
+    async def test_dev_p1_sample_tts_streams_local_mp3_without_openai_key(self) -> None:
+        with TemporaryDirectory() as tmp:
+            sample_dir = Path(tmp)
+            default_audio = b"default-mp3"
+            anxious_audio = (b"anxious-mp3-" * 500)
+            (sample_dir / "p1_seoyeon_01_depressed_slow.mp3").write_bytes(default_audio)
+            (sample_dir / "p1_seoyeon_03_anxious_guarded.mp3").write_bytes(anxious_audio)
+            service = VoiceService(
+                api_key="",
+                poc_sample_tts_enabled=True,
+                environment="dev",
+                poc_sample_tts_dir=sample_dir,
+            )
+            voice = VoicePreset(preset="soft-young-fem", openai_voice="coral")
+
+            chunks = [
+                chunk
+                async for chunk in service.synthesize_stream(
+                    "엄마한테 말하지 않는 거죠? 진짜 불안해요.",
+                    voice,
+                )
+            ]
+
+            self.assertFalse(service.is_available())
+            self.assertTrue(service.poc_sample_tts_available())
+            self.assertEqual(service.tts_provider(), "p1-sample-poc")
+            self.assertEqual(b"".join(chunk.audio for chunk in chunks), anxious_audio)
+
+    async def test_p1_sample_tts_is_disabled_outside_dev(self) -> None:
+        with TemporaryDirectory() as tmp:
+            sample_dir = Path(tmp)
+            (sample_dir / "p1_seoyeon_01_depressed_slow.mp3").write_bytes(b"default-mp3")
+            service = VoiceService(
+                api_key="",
+                poc_sample_tts_enabled=True,
+                environment="prod",
+                poc_sample_tts_dir=sample_dir,
+            )
+
+            self.assertFalse(service.is_available())
+            self.assertFalse(service.poc_sample_tts_available())
+            self.assertEqual(service.tts_provider(), "disabled-non-dev")
+
 
 class EndOfTurnDecisionTest(unittest.TestCase):
     def test_end_of_turn_requires_silence_threshold(self) -> None: