style(worker): black格式化edit_plan_service.py

根因：像素扰动功能(bb98620)在ffmpeg滤镜链中使用了color_balance滤镜名，
但ffmpeg的实际滤镜名是colorbalance（无下划线），导致滤镜解析失败，
报错'No such filter: color_balance'，ffmpeg退出码8。

影响：任何启用了像素扰动的变体渲染（默认批量变体会随机启用）都会失败，
表现为进度卡在56%，error_message只显示ffmpeg版本banner。

修复：将ffmpeg滤镜名从color_balance=改为colorbalance=。
内部config key保持color_balance不变（避免兼容问题）。

apps/api/app/api/routes/ai_avatar_render.py

@@ -18,6 +18,7 @@ from app.dependencies import get_db_session
 from app.schemas.ai_avatar_render import (
     AiAvatarRenderJobResponse,
     CreateAiAvatarRenderRequest,
+    FinalizeRenderResponse,
     SmartCoverResponse,
 )
 from app.services.ai_avatar_cover_service import generate_smart_cover
@@ -257,3 +258,67 @@ def generate_render_smart_cover(
         cover_url[:120],
     )
     return SmartCoverResponse(cover_url=cover_url, status="completed")
+
+# ── POST /{job_id}/finalize — 封面选定后正式入库成片库 ────────────────────
+
+
+@router.post("/{job_id}/finalize", response_model=FinalizeRenderResponse)
+def finalize_render_job(
+    job_id: str,
+    current_user: AuthenticatedUser = Depends(get_current_user),
+    db: Session = Depends(get_db_session),
+):
+    """用户完成封面选择后，将视频正式保存到成片库.
+
+    - 必须等渲染任务 completed 后才可调用
+    - 如果已通过 smart-cover/custom-cover 设置了封面，会自动带上
+    - 返回成片库视频ID
+    - 幂等：已 finalize 的任务重复调用会返回 existing 记录
+    """
+    from app.services.ai_avatar_render_service import AiAvatarRenderError, AiAvatarRenderService
+
+    svc = AiAvatarRenderService(db)
+    job = svc.get_render_job(job_id, current_user.user.id)
+    if job is None:
+        raise HTTPException(status_code=404, detail="渲染任务不存在")
+    if job.status != "completed":
+        raise HTTPException(status_code=400, detail="请先完成视频生成")
+
+    # 幂等检查（通过 generation_task_id=job_id 识别，finalize_job 内部也做了一次，这里提前返回简化）
+    from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
+
+    existing = (
+        db.query(GeneratedVideoModel)
+        .filter(
+            GeneratedVideoModel.user_id == current_user.user.id,
+            GeneratedVideoModel.generation_task_id == job_id,
+        )
+        .first()
+    )
+    if existing is not None:
+        return FinalizeRenderResponse(
+            video_id=existing.id,
+            cover_url=existing.thumbnail_url or "",
+            status="already_finalized",
+        )
+
+    try:
+        video = svc.finalize_job(job_id, current_user.user.id)
+        return FinalizeRenderResponse(
+            video_id=video.id,
+            cover_url=video.thumbnail_url or job.output_cover_url or "",
+            status="success",
+        )
+    except AiAvatarRenderError as exc:
+        status_map = {
+            "RenderJobNotFound": 404,
+            "RenderNotCompleted": 400,
+            "OutputVideoMissing": 400,
+        }
+        raise HTTPException(
+            status_code=status_map.get(exc.code, 400),
+            detail=str(exc),
+        ) from exc
+    except Exception as exc:
+        logger.error("渲染任务finalize失败: job_id=%s err=%s", job_id, exc, exc_info=True)
+        raise HTTPException(status_code=500, detail=f"保存到成片库失败: {str(exc)}") from exc

apps/api/app/api/routes/generation_tasks.py

@@ -61,24 +61,38 @@ def _query_voice_durations(db: Session, voice_ids: list[str]) -> list[float]:
     """批量查询配音素材时长（秒），#1749 配音时长分配用。
 
     逐项 try/float 硬化：MagicMock/异常/缺失 → 0.0（无配音不分配，不阻断）。
+
+    #1855 P0修复：不再对 voice_ids 去重，保持与调用方传入顺序/长度一致，
+    允许同配音id多次出现时返回相同时长（支持"同配音N变体"的时长对齐）。
     """
-    ids = [v for v in dict.fromkeys(voice_ids or []) if v]
-    if not ids:
+    # 先去重查询（IN 查询性能优化），但最终按原始 voice_ids 顺序返回
+    raw_ids = list(voice_ids or [])
+    if not raw_ids:
         return []
+    # 去重且保序，用于 SQL IN 查询；空字符串/None 视为无效id → 0.0
+    unique_ids: list[str] = []
+    _seen: set[str] = set()
+    for v in raw_ids:
+        if v and v not in _seen:
+            _seen.add(v)
+            unique_ids.append(v)
+    if not unique_ids:
+        return [0.0 for _ in raw_ids]
     try:
         from packages.adapters.sqlalchemy_impl.models import AssetModel
 
-        rows = db.query(AssetModel.id, AssetModel.duration).filter(AssetModel.id.in_(ids)).all()
+        rows = db.query(AssetModel.id, AssetModel.duration).filter(AssetModel.id.in_(unique_ids)).all()
         dur_map: dict[str, float] = {}
         for row in rows:
             try:
                 dur_map[row[0]] = float(row[1] or 0.0)
             except (TypeError, ValueError):
                 dur_map[row[0]] = 0.0
-        return [dur_map.get(v, 0.0) for v in ids]
+        # 按原始 voice_ids 顺序返回，保持长度一致；空/None/未查到 → 0.0
+        return [dur_map.get(v, 0.0) if v else 0.0 for v in raw_ids]
     except Exception:
         logger.warning("[生成任务] 配音时长查询失败（按无配音处理，不阻断）", exc_info=True)
-        return [0.0 for _ in ids]
+        return [0.0 for _ in raw_ids]
 
 
 def _to_generation_task_response(task) -> GenerationTaskResponse:
@@ -552,7 +566,26 @@ def create_generation_task(
                 ) from clone_err
             variant_plan_ids.append(_plan0.id)
 
-            # 变体 1..N-1 独立选片
+            # #1855 P0：批次区间避让表，从变体0实际clips构建初始值
+            def _collect_segments(pid):
+                segs = {}
+                _sk, _pg = 0, 500
+                while True:
+                    _b = _plan_svc._clip_repo.list_by_plan(pid, skip=_sk, limit=_pg)
+                    if not _b:
+                        break
+                    for _c in _b:
+                        if _c.asset_id and float(_c.duration or 0) > 0:
+                            _st = float(_c.start_time or 0.0)
+                            segs.setdefault(_c.asset_id, []).append((_st, _st + float(_c.duration)))
+                    if len(_b) < _pg:
+                        break
+                    _sk += _pg
+                return segs
+
+            _batch_segments = _collect_segments(_plan0.id)
+
+            # 变体 1..N-1 独立选片（传入累积batch_segments做素材区间避让）
             for task_index in range(1, count):
                 variant = None
                 last_err: Exception | None = None
@@ -564,6 +597,7 @@ def create_generation_task(
                             created_by_user_id=user_id,
                             name_suffix=f"批量{task_index + 1}",
                             voice_duration=voice_durations[task_index] if task_index < len(voice_durations) else 0.0,
+                            batch_segments=_batch_segments,
                         )
                         break
                     except ValueError as ve:
@@ -595,7 +629,16 @@ def create_generation_task(
                     ) from last_err
                 variant_plan_ids.append(variant.id)
 
-        # ③ 配音时长分配（回传 plan / clone 变体0 均需幂等分配；reselect 已在选片时分配）
+                # #1855 P0：把新变体的clips区间追加到batch_segments，供下一变体避让
+                try:
+                    _new_segs = _collect_segments(variant.id)
+                    for _aid, _ivs in _new_segs.items():
+                        _batch_segments.setdefault(_aid, []).extend(_ivs)
+                except Exception:
+                    logger.exception("[生成任务] 变体%d 区间收集失败（不阻断）", task_index)
+
+        # ③ 配音时长分配（回传 plan / clone 变体0 均需幂等分配；reselect 已在选片时分配，
+        #    #1855：apply_voice_duration_to_plan 已内置幂等判断，重复调用安全）
         for _vi, _pid in enumerate(variant_plan_ids):
             _vd = voice_durations[_vi] if _vi < len(voice_durations) else 0.0
             if _vd > 0:

apps/api/app/api/routes/lipsync.py

@@ -1,11 +1,12 @@
-"""对口型 API 路由 — #1796 MediaKit 对口型, #1809 参数调整.
+"""对口型 API 路由 — #1796 MediaKit 对口型, #1809 参数调整, #1845 配音前置.
 
 接口：
-  POST   /api/v1/lipsync/jobs          提交对口型任务
+  POST   /api/v1/lipsync/jobs          提交对口型任务（支持 TTS/直传/预合成 三种模式）
   GET    /api/v1/lipsync/jobs          任务列表
   GET    /api/v1/lipsync/jobs/{id}     任务详情
   POST   /api/v1/lipsync/jobs/{id}/refresh  刷新任务状态
   POST   /api/v1/lipsync/jobs/{id}/cancel   取消任务
+  POST   /api/v1/lipsync/tts-preview   #1845 步骤1 TTS 预合成（同步 HTTP，~2-3s）
 """
 
 from __future__ import annotations
@@ -17,7 +18,12 @@ from app.dependencies import (
     get_db_session,
     get_voice_clone_profile_repository,
 )
-from app.schemas.lipsync import CreateLipsyncJobRequest, LipsyncJobResponse
+from app.schemas.lipsync import (
+    AiAvatarTtsPreviewRequest,
+    AiAvatarTtsPreviewResponse,
+    CreateLipsyncJobRequest,
+    LipsyncJobResponse,
+)
 from app.services.lipsync_service import LipsyncService
 from app.services.mediakit_client import MediaKitError
 from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException, Query
@@ -33,7 +39,6 @@ def _get_service(
     voice_clone_repo=Depends(get_voice_clone_profile_repository),
 ) -> LipsyncService:
     # voice_clone_repo 用于克隆音色 profile 解析
-    # TTS 合成已移至 Celery 异步任务，无需同步注入 cosyvoice_service
     return LipsyncService(
         db,
         voice_clone_repo=voice_clone_repo,
@@ -51,15 +56,20 @@ def create_lipsync_job(
 ):
     """提交对口型任务.
 
-    #1809/#1822: 前端传 {video_url, voice_id, script_text, speed?, emotion?}，
-    后端创建任务记录（状态 tts_processing），dispatch Celery 异步任务执行 TTS 合成 + MediaKit 提交；
-    也支持直接传 {video_url, audio_url}（同步提交 MediaKit）。
+    三种模式：
+    - TTS 直生（旧版/降级）：传 {video_url, voice_id, script_text, speed?, emotion?}，
+      后端 dispatch Celery 异步任务。
+    - 直接音频：传 {video_url, audio_url}，后端同步下载+算timings+提交MediaKit。
+    - 预合成音频（#1845 新主路径）：传 {video_url, audio_url, audio_duration, sentence_timings}，
+      后端同步ffprobe+写入timings+直接提交MediaKit（~2-3s）。
     """
     try:
         job = svc.create_job(
             user_id=current_user.user.id,
             video_url=body.video_url,
             audio_url=body.audio_url,
+            audio_duration=body.audio_duration,
+            sentence_timings=body.sentence_timings,
             voice_id=body.voice_id,
             script_text=body.script_text,
             speed=body.speed,
@@ -68,10 +78,8 @@ def create_lipsync_job(
             project_id=body.project_id,
         )
     except ValueError as exc:
-        # 参数无效（如 voice_id 格式不对、文本过长等）
         raise HTTPException(status_code=400, detail=str(exc)) from exc
     except MediaKitError as exc:
-        # 音色无权访问 → 403；参数无效 → 400；MediaKit 提交失败 → 502
         status_code = 502
         if exc.code in ("VoiceForbidden",):
             status_code = 403
@@ -86,7 +94,6 @@ def create_lipsync_job(
             },
         ) from exc
     except Exception as exc:
-        # 兜底：任何未预期的错误返回 400 而非 500
         logger.error("创建对口型任务异常: %s", exc, exc_info=True)
         raise HTTPException(
             status_code=400,
@@ -96,6 +103,52 @@ def create_lipsync_job(
     return job
 
 
+# ── POST /tts-preview — #1845 步骤1 TTS 预合成 ──────────────────────────
+
+
+@router.post("/tts-preview", response_model=AiAvatarTtsPreviewResponse)
+def preview_tts(
+    body: AiAvatarTtsPreviewRequest,
+    current_user: AuthenticatedUser = Depends(get_current_user),
+    svc: LipsyncService = Depends(_get_service),
+):
+    """步骤1「生成配音」同步 TTS 预合成.
+
+    同步执行 TTS 合成 → 下载音频 → ffprobe 时长 → 句子时间戳计算，
+    不创建 LipsyncJob、不转存 OSS，直接返回 CosyVoice 临时 URL（~24h 有效）。
+    耗时约 2-3 秒。
+    """
+    try:
+        result = svc.preview_tts(
+            user_id=current_user.user.id,
+            voice_id=body.voice_id,
+            script_text=body.script_text,
+            speed=body.speed,
+            emotion=body.emotion,
+        )
+    except MediaKitError as exc:
+        status_code = 400
+        if exc.code in ("VoiceForbidden",):
+            status_code = 403
+        elif exc.code in ("TTSNoAudio",):
+            status_code = 502
+        raise HTTPException(
+            status_code=status_code,
+            detail={
+                "code": exc.code,
+                "message": str(exc),
+            },
+        ) from exc
+    except Exception as exc:
+        logger.error("TTS 预合成异常: %s", exc, exc_info=True)
+        raise HTTPException(
+            status_code=400,
+            detail=f"TTS 合成失败: {exc}",
+        ) from exc
+
+    return result
+
+
 # ── GET /jobs — 任务列表 ─────────────────────────────────────────────────
 
 
@@ -134,11 +187,7 @@ def get_lipsync_job(
     current_user: AuthenticatedUser = Depends(get_current_user),
     svc: LipsyncService = Depends(_get_service),
 ):
-    """获取对口型任务详情.
-
-    非终态任务：先返回 DB 缓存，挂后台刷新（下次轮询拿到新状态），
-    避免 MediaKit 慢响应阻塞前端轮询。
-    """
+    """获取对口型任务详情."""
     job = svc.get_job(job_id, current_user.user.id)
     if job is None:
         raise HTTPException(status_code=404, detail="任务不存在")

apps/api/app/schemas/ai_avatar_render.py

@@ -116,3 +116,11 @@ class SmartCoverResponse(BaseModel):
     cover_url: str = Field("", description="封面图公网 URL（OSS，非临时）；失败为空")
     status: str = Field("completed", description="completed / fallback_failed")
     message: str = Field("", description="失败原因（如有）")
+
+
+class FinalizeRenderResponse(BaseModel):
+    """封面选好后点「完成」，正式入库成片库的响应."""
+
+    video_id: str = Field(..., description="成片库视频ID")
+    cover_url: str = Field("", description="封面URL")
+    status: str = Field("success", description="success/already_finalized")

apps/api/app/schemas/lipsync.py

@@ -1,9 +1,12 @@
-"""对口型 API Schema 定义 — #1796 / #1809 / #1822.
+"""对口型 API Schema 定义 — #1796 / #1809 / #1822 / #1845（配音前置）.
 
-支持两种输入模式（二选一）：
-1. TTS 直生模式（推荐）：传 voice_id + script_text（+ speed/emotion），
-   后端内部先调 CosyVoice 合成音频，再提交 MediaKit 对口型。
+支持三种输入模式：
+1. TTS 直生模式（兼容旧版前端）：传 voice_id + script_text（+ speed/emotion），
+   后端 Celery 异步做 TTS 合成 + MediaKit 提交。
 2. 直接音频模式：传 video_url + audio_url（音频已由调用方准备好）。
+3. 预合成音频模式（#1845 配音前置新主路径）：前端先调 POST /lipsync/tts-preview
+   拿到 audio_url + sentence_timings，再在 create_job 时传 audio_url + audio_duration
+   + sentence_timings，后端跳过 TTS 和时间戳计算，直接 ffprobe 校验后提交 MediaKit。
 """
 
 from __future__ import annotations
@@ -46,15 +49,19 @@ class LipsyncJobResponse(BaseModel):
 class CreateLipsyncJobRequest(BaseModel):
     """创建对口型任务请求.
 
-    两种模式（二选一）：
-    - TTS 直生：voice_id + script_text 必填（+ 可选 speed/emotion）；audio_url 留空。
+    三种模式（三选一）：
+    - TTS 直生（旧版/降级）：voice_id + script_text 必填；audio_url 留空。
     - 直接音频：video_url + audio_url 必填。
+    - 预合成音频（#1845 新主路径）：audio_url 必填 + 可选 audio_duration/sentence_timings；
+      后端同步 ffprobe 校验时长、写入 timings，直接提交 MediaKit。
     """
 
     video_url: str = Field(..., description="人物视频 URL（MP4，≤30min，单人真人）")
 
-    # 模式 2：直接音频
+    # 模式 2/3：直接/预合成音频
     audio_url: str = Field("", description="驱动音频 URL（mp3/aac/wav/m4a/flac）；直生模式留空")
+    audio_duration: Optional[float] = Field(None, ge=0, description="预合成音频时长（秒），可选；后端会 ffprobe 校验")
+    sentence_timings: Optional[list] = Field(None, description="预合成接口返回的句子时间戳，可选；若传入则直接写入 job")
 
     # 模式 1：TTS 直生
     voice_id: str = Field("", description="音色 ID（预置音色或克隆音色 profile UUID）")
@@ -62,7 +69,9 @@ class CreateLipsyncJobRequest(BaseModel):
     speed: float = Field(1.0, ge=0.5, le=2.0, description="语速（0.5-2.0），默认 1.0")
     emotion: str = Field("", description="情绪（natural/excited/calm/friendly 或中文 自然/兴奋/沉稳/亲切）")
 
-    enable_video_loop: bool = Field(False, description="音频长于视频时是否循环画面")
+    enable_video_loop: bool = Field(
+        True, description="音频长于视频时是否循环画面（AI数字人默认开启，防止音频长于视频被截断）"
+    )
     project_id: str = Field("", description="项目 ID（可选）")
 
     @model_validator(mode="after")
@@ -73,15 +82,16 @@ class CreateLipsyncJobRequest(BaseModel):
         if not video.startswith(("http://", "https://")):
             raise ValueError("video_url 必须是 HTTP/HTTPS URL")
         lower = video.lower().split("?")[0]
-        if not lower.endswith(".mp4"):
-            raise ValueError("video_url 仅支持 MP4 格式")
+        allowed_video_exts = (".mp4", ".mov", ".m4v", ".webm", ".avi", ".mkv", ".3gp")
+        if not any(lower.endswith(ext) for ext in allowed_video_exts):
+            raise ValueError("video_url 格式不支持，仅支持: " + ", ".join(allowed_video_exts))
 
         has_audio = bool((self.audio_url or "").strip())
         has_tts = bool((self.voice_id or "").strip()) and bool((self.script_text or "").strip())
 
         if not has_audio and not has_tts:
             raise ValueError(
-                "必须提供驱动音频：要么传 audio_url（直接音频模式），"
+                "必须提供驱动音频：要么传 audio_url（直接/预合成音频模式），"
                 "要么同时传 voice_id + script_text（TTS 直生模式）"
             )
 
@@ -99,3 +109,23 @@ class CreateLipsyncJobRequest(BaseModel):
             self.audio_url = au
 
         return self
+
+
+# ── #1845 TTS 预合成接口 ────────────────────────────────────────────────
+
+
+class AiAvatarTtsPreviewRequest(BaseModel):
+    """步骤1「生成配音」预合成请求（同步 HTTP，~2-3s）."""
+
+    voice_id: str = Field(..., min_length=1, max_length=128, description="音色 ID")
+    script_text: str = Field(..., min_length=1, max_length=5000, description="要合成的文案")
+    speed: float = Field(1.0, ge=0.5, le=2.0, description="语速（0.5-2.0），默认 1.0")
+    emotion: str = Field("natural", max_length=32, description="情绪")
+
+
+class AiAvatarTtsPreviewResponse(BaseModel):
+    """TTS 预合成响应（临时 URL，24h 内有效，足够当前会话使用）."""
+
+    audio_url: str = Field(..., description="CosyVoice 临时音频 URL")
+    duration: float = Field(..., ge=0, description="音频总时长（秒），ffprobe 测得")
+    sentence_timings: list[dict] = Field(..., description="句子级精确时间戳")

apps/api/app/services/ai_avatar_cover_service.py

@@ -22,9 +22,9 @@ from urllib.parse import urlparse
 
 logger = logging.getLogger(__name__)
 
-# MediaKit 抽帧轮询参数：poll_interval=1s × max_poll=15 → 最长 15s，配合前端 120s 超时足够
-COVER_POLL_INTERVAL = 1.0
-COVER_MAX_POLL_ATTEMPTS = 15
+# MediaKit 抽帧轮询参数：poll_interval=2s × max_poll=30 → 最长 60s（与 mediakit_client 默认值/lipsync 轮询保持一致，防止合成视频下载+抽帧超时）
+COVER_POLL_INTERVAL = 2.0
+COVER_MAX_POLL_ATTEMPTS = 30
 
 # 帧图片下载超时（秒）
 FRAME_DOWNLOAD_TIMEOUT = 20

apps/api/app/services/ai_avatar_render_service.py

@@ -399,41 +399,9 @@ class AiAvatarRenderService:
             self.db.commit()
             logger.info("渲染任务完成: %s", job_id)
 
-            # 7. 自动保存成片记录到成片库
-            if job.output_video_url:
-                try:
-                    from packages.adapters.sqlalchemy_impl.generated_video_repository import (
-                        SQLAlchemyGeneratedVideoRepository,
-                    )
-                    from packages.domain.generated_video import GeneratedVideo
-
-                    clip_name = f"AI数字人_{job_id[:8]}"
-                    # AI数字人入口是独立页面，前端可能不传 project_id（无项目概念），
-                    # 兜底为 "ai_avatar" 避免 DB 非空约束/查询问题；generation_task_id 同样兜底用 render_job_id
-                    clip_project_id = (job.project_id or "").strip() or "ai_avatar"
-                    clip_generation_task_id = (job.lipsync_job_id or "").strip() or job_id
-                    clip = GeneratedVideo.create(
-                        project_id=clip_project_id,
-                        generation_task_id=clip_generation_task_id,
-                        name=clip_name,
-                        file_url=job.output_video_url,
-                        user_id=job.user_id,
-                        duration=job.output_duration or 0.0,
-                        thumbnail_url=job.output_cover_url or None,
-                        generation_params={
-                            "source": "ai_avatar_render",
-                            "render_job_id": job.id,
-                        },
-                    )
-                    video_repo = SQLAlchemyGeneratedVideoRepository(self.db)
-                    video_repo.create(clip)
-                    logger.info("成片记录已保存到成片库: clip_id=%s, render_job=%s", clip.id, job_id)
-                except Exception:
-                    logger.error(
-                        "自动保存成片记录失败（不影响渲染任务状态）: render_job=%s",
-                        job_id,
-                        exc_info=True,
-                    )
+            # 7. 渲染完成，停留在「待选封面」状态：不自动入库。
+            #    用户在前端选好封面、点「完成」后，由 /{job_id}/finalize 接口显式入库。
+            logger.info("渲染任务完成，等待用户选择封面后入库: job_id=%s", job_id)
 
         except AiAvatarRenderError as exc:
             job.status = "failed"
@@ -450,6 +418,87 @@ class AiAvatarRenderService:
             logger.exception("渲染任务异常 [%s]", job_id)
             raise
 
+    def _persist_to_library(self, job: AiAvatarRenderJob, cover_url: Optional[str] = None):
+        """将渲染结果写入成片库，返回 GeneratedVideo 领域对象.
+
+        Args:
+            job: 渲染任务（必须 status=completed 且 output_video_url 非空）
+            cover_url: 可选的封面 URL 覆盖（finalize 时传入即优先使用，否则取 job.output_cover_url）
+        """
+        from packages.adapters.sqlalchemy_impl.generated_video_repository import (
+            SQLAlchemyGeneratedVideoRepository,
+        )
+        from packages.domain.generated_video import GeneratedVideo
+
+        clip_name = f"AI数字人_{job.id[:8]}"
+        # AI数字人入口是独立页面，前端可能不传 project_id（无项目概念），
+        # 兜底为 "ai_avatar" 避免 DB 非空约束/查询问题；generation_task_id 用 render_job_id 便于反查。
+        clip_project_id = (job.project_id or "").strip() or "ai_avatar"
+        clip_generation_task_id = job.id
+        effective_cover = (cover_url or "").strip() if cover_url else (job.output_cover_url or "").strip()
+        clip = GeneratedVideo.create(
+            project_id=clip_project_id,
+            generation_task_id=clip_generation_task_id,
+            name=clip_name,
+            file_url=job.output_video_url,
+            user_id=job.user_id,
+            duration=job.output_duration or 0.0,
+            thumbnail_url=effective_cover or None,
+            generation_params={
+                "source": "ai_avatar_render",
+                "render_job_id": job.id,
+            },
+        )
+        video_repo = SQLAlchemyGeneratedVideoRepository(self.db)
+        video_repo.create(clip)
+        logger.info("[数字人渲染] 成片已入库: clip_id=%s render_job=%s", clip.id, job.id)
+        return clip
+
+    def finalize_job(self, job_id: str, user_id: str, cover_url: Optional[str] = None):
+        """用户在前端点「完成」后调用：将已 completed 的渲染任务正式入库到成片库.
+
+        - 必须 status=completed 才可调用
+        - cover_url 若传入则优先使用并回写 job.output_cover_url；否则使用 job.output_cover_url（smart-cover/custom-cover 已写入）
+        - 幂等：已入库则返回已存在的 GeneratedVideo
+        """
+        from packages.adapters.sqlalchemy_impl.generated_video_repository import (
+            SQLAlchemyGeneratedVideoRepository,
+        )
+        from packages.adapters.sqlalchemy_impl.models import GeneratedVideoModel
+
+        job = self.get_render_job(job_id, user_id)
+        if job is None:
+            raise AiAvatarRenderError("渲染任务不存在", code="RenderJobNotFound")
+        if job.status != "completed":
+            raise AiAvatarRenderError(f"渲染任务未完成（当前状态: {job.status}），无法入库", code="RenderNotCompleted")
+        if not (job.output_video_url or "").strip():
+            raise AiAvatarRenderError("渲染成片视频 URL 为空，无法入库", code="OutputVideoMissing")
+
+        # 幂等检查：已入库直接返回现有记录（通过 generation_task_id=job_id 识别，
+        # 因为入库时 generation_task_id 被设置为 render_job_id 自身）
+        existing = (
+            self.db.query(GeneratedVideoModel)
+            .filter(
+                GeneratedVideoModel.user_id == user_id,
+                GeneratedVideoModel.generation_task_id == job_id,
+            )
+            .first()
+        )
+        if existing is not None:
+            logger.info("[数字人渲染] finalize 幂等命中，返回已存在记录: clip_id=%s job_id=%s", existing.id, job_id)
+            return SQLAlchemyGeneratedVideoRepository(self.db).get(existing.id)
+
+        # 传入 cover_url 时回写到 job
+        if cover_url and cover_url.strip():
+            job.output_cover_url = cover_url.strip()
+            # 同步更新 cover_config，保持 smart-cover 路径一致
+            if isinstance(job.cover_config, dict):
+                job.cover_config = {**job.cover_config, "mode": "auto_frame", "url": cover_url.strip()}
+            job.updated_at = datetime.now(timezone.utc)
+            self.db.commit()
+
+        return self._persist_to_library(job, cover_url=cover_url)
+
     def _download_video(self, url: str) -> str:
         """下载视频到临时文件."""
         import httpx

apps/api/app/services/edit_plan_service.py

@@ -473,6 +473,7 @@ class EditPlanService:
         name_suffix: str = "变体",
         voice_duration: float = 0.0,
         rng=None,
+        batch_segments: dict[str, list[tuple[float, float]]] | None = None,
     ) -> EditPlan:
         """为批量变体生成独立 plan：完整重跑单视频选片流程（#1743）。
 
@@ -489,6 +490,8 @@ class EditPlanService:
             created_by_user_id: 新 plan 归属用户。
             name_suffix: plan 名后缀。
             rng: 可选随机数（测试注入种子）。
+            batch_segments: 可选，外部传入的批次内已使用素材区间（前序变体避让用）。
+                传入时作为初始避让对象；未传则保持原逻辑从源 plan clips 自建（向后兼容）。
 
         Raises:
             ValueError: 源 plan 不存在/无片段、素材池为空或时长全未知。
@@ -537,16 +540,26 @@ class EditPlanService:
             voice = float(voice_duration or 0.0)
         except (TypeError, ValueError):
             voice = 0.0
+        rhythm_template_for_reselect = None
+        if source.config:
+            rhythm_template_for_reselect = source.config.get("rhythm_template")
+
         if voice > 0 and source_clips_data:
             from packages.domain.voice_duration_planner import plan_clip_durations
 
             _effects: list[str | None] = [c.get("transition_effect") for c in source_clips_data]
             _tdurs: list[float] = [float(c.get("transition_duration") or 0.0) for c in source_clips_data]
+            # #1855 P0：先占位durations为空dict，真正查durations在后面pool_ids确定后执行；
+            # plan_clip_durations 的 asset_durations 参数在该函数中仅作最大段长钳制，
+            # 这里先不依赖它（durations 还没查），传 None 让planner用默认策略；
+            # 真正的asset_durations会在后面 clips_data 生成时传入 reselect_clips_for_variant
             target_durations = plan_clip_durations(
                 len(source_clips_data),
                 voice,
                 transition_effects=_effects,
                 transition_durations=_tdurs,
+                rhythm_template=rhythm_template_for_reselect,
+                asset_durations=None,
             )
             if target_durations:
                 for _c, _d in zip(source_clips_data, target_durations, strict=False):
@@ -582,12 +595,18 @@ class EditPlanService:
             created_by_user_id=created_by_user_id or (source.created_by_user_id or ""),
         )
 
-        # 批次内区间：以源 plan（变体 0）片段为初始避让对象
-        batch_segments: dict[str, list[tuple[float, float]]] = {}
-        for c in clips:
-            if c.asset_id and float(c.duration or 0) > 0:
-                st = float(c.start_time or 0.0)
-                batch_segments.setdefault(c.asset_id, []).append((st, st + float(c.duration)))
+        # 批次内区间：外部传入时使用外部传入（含前序变体已用区间）；
+        # 否则保持原逻辑从源 plan clips 自建（向后兼容）
+        if batch_segments is not None:
+            batch_segments_resolved: dict[str, list[tuple[float, float]]] = {
+                k: list(v) for k, v in batch_segments.items()
+            }
+        else:
+            batch_segments_resolved = {}
+            for c in clips:
+                if c.asset_id and float(c.duration or 0) > 0:
+                    st = float(c.start_time or 0.0)
+                    batch_segments_resolved.setdefault(c.asset_id, []).append((st, st + float(c.duration)))
 
         clips_data = reselect_clips_for_variant(
             source_clips_data,
@@ -595,7 +614,7 @@ class EditPlanService:
             asset_durations=durations,
             asset_scene_points=scene_points,
             historical_used_segments=historical,
-            batch_segments=batch_segments,
+            batch_segments=batch_segments_resolved,
             target_durations=target_durations,
             rng=rng,
         )
@@ -767,6 +786,17 @@ class EditPlanService:
         if plan is None:
             return None
 
+        # #1855 P0：幂等判断——如果已成功分配过且当前 total_duration 已接近 voice_duration，直接返回
+        try:
+            existing_mark = None
+            if plan.config:
+                existing_mark = plan.config.get("voice_duration_applied")
+            cur_total = float(plan.total_duration or 0.0)
+            if existing_mark is not None and abs(existing_mark - voice) < 1e-6 and abs(cur_total - voice) < 0.5:
+                return plan
+        except Exception:
+            pass
+
         clips: List[EditPlanClip] = []
         skip, page = 0, 500
         while True:
@@ -838,6 +868,10 @@ class EditPlanService:
         )
         try:
             plan.total_duration = net
+            # #1855 P0：写入幂等标记，避免二次调用时只重分配 duration 不重算 start_time
+            new_cfg = dict(plan.config or {})
+            new_cfg["voice_duration_applied"] = voice
+            plan.config = new_cfg
             db = self._clip_repo.session
             db.commit()
         except Exception:
@@ -878,6 +912,69 @@ class EditPlanService:
         rng = rng or _random.Random()
         plan_ids: list[str] = []
 
+        # #1855 P0：先确定片段数 clip_count（用于节奏模板生成长度匹配）
+        from packages.domain.bgm_pool import allocate_bgm_pool_for_variants
+        from packages.domain.variant_plan_selector import (
+            generate_pixel_perturbation,
+            generate_visual_perturbation,
+        )
+        from packages.domain.voice_duration_planner import RHYTHM_TEMPLATES, adapt_template_length
+
+        clip_count = 0
+        # 从源 plan 获取片段数（分页读，避免关系加载问题）
+        _sclips: list = []
+        _sk, _pg = 0, 500
+        while True:
+            _b = self._clip_repo.list_by_plan(source_plan_id, skip=_sk, limit=_pg)
+            if not _b:
+                break
+            _sclips.extend(_b)
+            if len(_b) < _pg:
+                break
+            _sk += _pg
+        clip_count = len(_sclips)
+
+        # 预先生成所有 N 个变体的节奏模板/BGM/扰动参数（时机提前到选片前写入config）
+        rhythm_templates_for_variants: list = []
+        for _idx in range(count):
+            if clip_count > 0:
+                variant_seed = rng.randint(0, 999999)
+                _tpl = adapt_template_length(RHYTHM_TEMPLATES[variant_seed % len(RHYTHM_TEMPLATES)], clip_count)
+                rhythm_templates_for_variants.append(_tpl)
+            else:
+                rhythm_templates_for_variants.append(None)
+
+        source_bgm_config: dict = {}
+        source_plan = self.get_plan(source_plan_id)
+        if source_plan and source_plan.config:
+            source_bgm_config = source_plan.config.get("bgm", {}) or {}
+        variant_seeds_for_bgm = [rng.randint(0, 999999) for _ in range(count)]
+        bgm_pool_assignments = allocate_bgm_pool_for_variants(source_bgm_config, variant_seeds_for_bgm)
+
+        def _build_variant_config_update(idx: int) -> dict:
+            """构建单个变体的 config 更新（节奏模板/BGM/视觉/像素扰动）。"""
+            upd: dict = {}
+            try:
+                perturbation = generate_visual_perturbation(rng)
+                if idx == 0:
+                    perturbation["hflip"] = False
+                upd["visual_perturbation"] = perturbation
+            except Exception:
+                logger.exception("变体 %d 视觉扰动生成失败（不阻断）", idx)
+            try:
+                pixel_pert = generate_pixel_perturbation(rng)
+                upd["pixel_perturbation"] = pixel_pert
+            except Exception:
+                logger.exception("变体 %d 像素扰动生成失败（不阻断）", idx)
+            rt = rhythm_templates_for_variants[idx] if idx < len(rhythm_templates_for_variants) else None
+            if rt is not None:
+                upd["rhythm_template"] = rt
+            if idx < len(bgm_pool_assignments):
+                existing_bgm = dict((source_plan.config or {}).get("bgm", {}) or {})
+                existing_bgm.update(bgm_pool_assignments[idx])
+                upd["bgm"] = existing_bgm
+            return upd
+
         # 变体 0：clone（片段结构同源 plan，起点重算），不污染源 plan
         plan0 = self.clone_plan_for_variant(
             source_plan_id,
@@ -890,6 +987,15 @@ class EditPlanService:
                 v0_voice = float(voice_durations[0] or 0.0)
             except (TypeError, ValueError):
                 v0_voice = 0.0
+
+        # #1855 P0：在配音分配前先写入变体0的节奏模板/扰动/BGM，确保 apply_voice_duration_to_plan 能读到 rhythm_template
+        try:
+            _cfg0 = _build_variant_config_update(0)
+            if _cfg0:
+                self.update_plan_config(plan0.id, _cfg0)
+        except Exception:
+            logger.exception("变体0 配置写入失败（不阻断）: plan=%s", plan0.id)
+
         if v0_voice > 0:
             try:
                 self.apply_voice_duration_to_plan(plan0.id, v0_voice)
@@ -897,7 +1003,27 @@ class EditPlanService:
                 logger.exception("变体0 配音分配失败（不阻断）: plan=%s", plan0.id)
         plan_ids.append(plan0.id)
 
-        # 变体 1..N-1：独立选片
+        # #1855 P0：批次内素材区间避让表——从变体0实际落库的clips构建初始值
+        def _collect_plan_segments(pid: str) -> dict[str, list[tuple[float, float]]]:
+            """分页读取 plan 所有 clips，构建 {asset_id: [(start, end), ...]} 区间表。"""
+            segs: dict[str, list[tuple[float, float]]] = {}
+            _sk2, _pg2 = 0, 500
+            while True:
+                _b2 = self._clip_repo.list_by_plan(pid, skip=_sk2, limit=_pg2)
+                if not _b2:
+                    break
+                for _c in _b2:
+                    if _c.asset_id and float(_c.duration or 0) > 0:
+                        _st = float(_c.start_time or 0.0)
+                        segs.setdefault(_c.asset_id, []).append((_st, _st + float(_c.duration)))
+                if len(_b2) < _pg2:
+                    break
+                _sk2 += _pg2
+            return segs
+
+        batch_segments_acc: dict[str, list[tuple[float, float]]] = _collect_plan_segments(plan0.id)
+
+        # 变体 1..N-1：独立选片（传入累积的 batch_segments 做区间避让）
         for i in range(1, count):
             voice = 0.0
             if voice_durations and i < len(voice_durations):
@@ -905,6 +1031,14 @@ class EditPlanService:
                     voice = float(voice_durations[i] or 0.0)
                 except (TypeError, ValueError):
                     voice = 0.0
+
+            # #1855 P0：在reselect前先为"变体i"准备配置更新——但reselect内部复制的是source.config，
+            # 所以每个变体独立的节奏模板需要在reselect后单独写入config；
+            # 但 plan_clip_durations 用的是 source.config.rhythm_template（即源plan的节奏模板），
+            # 为了让每个变体在选片阶段就使用自己的节奏模板分配段长，这里采用：
+            #   - reselect 仍使用源 plan 的 rhythm_template（保持片段骨架一致）
+            #   - 选片完成后立即写入该变体自己的 rhythm_template/扰动/BGM 到config，
+            #     后续不再二次 apply_voice_duration_to_plan（由幂等标记跳过）
             variant = self.reselect_plan_for_variant(
                 source_plan_id,
                 candidate_asset_ids,
@@ -912,73 +1046,26 @@ class EditPlanService:
                 name_suffix=f"变体{i + 1}",
                 voice_duration=voice,
                 rng=rng,
+                batch_segments=batch_segments_acc,
             )
+
+            # 选片完成后写入该变体的独立配置（节奏模板/扰动/BGM）
+            try:
+                _cfgi = _build_variant_config_update(i)
+                if _cfgi:
+                    self.update_plan_config(variant.id, _cfgi)
+            except Exception:
+                logger.exception("变体 %d 配置写入失败（不阻断）: plan=%s", i, variant.id)
+
             plan_ids.append(variant.id)
 
-        # #1764：为每个变体生成独立节奏模板（让批量视频片段时长分布不同）
-        from packages.domain.voice_duration_planner import RHYTHM_TEMPLATES, adapt_template_length
-
-        clip_count = 0
-        if voice_durations and len(voice_durations) > 0:
-            # 从源 plan 获取片段数
-            source_plan = self.get_plan(source_plan_id)
-            if source_plan and hasattr(source_plan, "clips"):
-                clip_count = len(list(source_plan.clips)) if source_plan.clips else 0
-
-        rhythm_templates_for_variants = []
-        if clip_count > 0:
-            for idx in range(len(plan_ids)):
-                # 每个变体用不同的 seed 选择节奏模板
-                variant_seed = rng.randint(0, 999999)
-                template = adapt_template_length(RHYTHM_TEMPLATES[variant_seed % len(RHYTHM_TEMPLATES)], clip_count)
-                rhythm_templates_for_variants.append(template)
-                logger.info("变体 %d 节奏模板: plan=%s template=%s", idx, plan_ids[idx], template)
-
-        # #1767：BGM 池差异化分配（让批量变体使用不同 BGM / 段落 / 音量）
-        from packages.domain.bgm_pool import allocate_bgm_pool_for_variants
-
-        source_bgm_config = {}
-        source_plan = self.get_plan(source_plan_id)
-        if source_plan and source_plan.config:
-            source_bgm_config = source_plan.config.get("bgm", {}) or {}
-
-        variant_seeds_for_bgm = [rng.randint(0, 999999) for _ in plan_ids]
-        bgm_pool_assignments = allocate_bgm_pool_for_variants(source_bgm_config, variant_seeds_for_bgm)
-
-        # 为每个变体生成独立视觉扰动参数（让批量视频画面本身更不同）
-        from packages.domain.variant_plan_selector import generate_visual_perturbation
-
-        for idx, pid in enumerate(plan_ids):
+            # #1855 P0：把当前新变体的 clips 区间追加到 batch_segments，供下一变体避让
             try:
-                perturbation = generate_visual_perturbation(rng)
-                # 变体 0 不做 hflip（保持预览 plan 原始画面方向）
-                if idx == 0:
-                    perturbation["hflip"] = False
-                config_update = {"visual_perturbation": perturbation}
-                # #1764：写入节奏模板
-                if idx < len(rhythm_templates_for_variants):
-                    config_update["rhythm_template"] = rhythm_templates_for_variants[idx]
-                # #1765：写入像素级扰动滤镜
-                from packages.domain.variant_plan_selector import generate_pixel_perturbation
-
-                pixel_pert = generate_pixel_perturbation(rng)
-                config_update["pixel_perturbation"] = pixel_pert
-                # #1767：写入 BGM 池分配（覆盖 bgm 配置中的 preset_id / audio_offset / volume_adjust_db）
-                if idx < len(bgm_pool_assignments):
-                    existing_bgm = dict((source_plan.config or {}).get("bgm", {}) or {})
-                    existing_bgm.update(bgm_pool_assignments[idx])
-                    config_update["bgm"] = existing_bgm
-                self.update_plan_config(pid, config_update)
-                logger.info(
-                    "变体 %d 视觉扰动+像素扰动+BGM池: plan=%s vis=%s pix=%s bgm=%s",
-                    idx,
-                    pid,
-                    perturbation,
-                    pixel_pert,
-                    bgm_pool_assignments[idx] if idx < len(bgm_pool_assignments) else None,
-                )
+                _new_segs = _collect_plan_segments(variant.id)
+                for _aid, _ivs in _new_segs.items():
+                    batch_segments_acc.setdefault(_aid, []).extend(_ivs)
             except Exception:
-                logger.exception("变体 %d 视觉扰动生成失败（不阻断）: plan=%s", idx, pid)
+                logger.exception("变体 %d 区间收集失败（不阻断）: plan=%s", i, variant.id)
 
         # 标记所有变体 plan 的 clips 为 ready（已分配素材+起点，语义上就是 ready）
         for pid in plan_ids:

apps/api/app/services/lipsync_service.py

@@ -1,8 +1,12 @@
-"""对口型 Service — #1796 MediaKit 对口型业务逻辑, #1809 参数调整.
+"""对口型 Service — #1796 MediaKit 对口型业务逻辑, #1809 参数调整, #1845 配音前置.
 
 职责：
 - 创建/查询对口型任务
-- 双输入模式：TTS 直生（voice_id + script_text，内部先合成音频转存 OSS）或直接音频（audio_url）
+- 三输入模式：
+  1. TTS 直生（voice_id + script_text）→ 走 Celery 异步（降级路径）
+  2. 直接音频（audio_url，前端未传 timings）→ 同步下载 + 算 timings + 提交 MediaKit
+  3. 预合成音频（audio_url + sentence_timings，#1845 新主路径）→ 同步 ffprobe 校验时长 +
+     写入前端传来的 timings → 直接提交 MediaKit（~2-3s）
 - 调用 MediaKit 客户端提交异步任务
 - 轮询更新任务状态（中间状态同步 DB，成片转存自家 OSS）
 - 用户隔离（每个用户只能操作自己的任务）
@@ -26,19 +30,22 @@ from app.services.mediakit_client import (
     get_mediakit_client,
 )
 
-# Celery 异步任务：TTS 合成 + MediaKit 提交（#lipsync-speed-optimization）
+# Celery 异步任务：TTS 合成 + MediaKit 提交（降级路径）
 from app.tasks.lipsync_tts import tts_synthesize_and_submit
 from sqlalchemy.orm import Session
 
 from packages.adapters.sqlalchemy_impl.models import LipsyncJobModel
 from packages.application.cosyvoice_service import CosyVoiceError, normalize_emotion
+from packages.domain.sentence_timings import (
+    compute_sentence_timings,
+    probe_audio_duration,
+)
 from packages.shared.storage import get_shared_storage_service
 from packages.shared.url_security import ALLOWED_AUDIO_MIME_TYPES, safe_download_bytes
 
 logger = logging.getLogger(__name__)
 
 # 传给 MediaKit GPU worker / 回给前端播放的 OSS 预签名有效期：7 天。
-# MediaKit 排队 + 拉取可能延迟，私有桶裸 URL 或 1 小时短预签名都会 403，故统一重签长有效期。
 MEDIAKIT_URL_TTL_SECONDS = 7 * 24 * 3600
 
 
@@ -142,6 +149,100 @@ class LipsyncService:
             logger.warning("TTS 音频转存 OSS 失败，回退临时 URL: job_id=%s err=%s", job_id, exc)
             return temp_url
 
+    def _submit_audio_direct(
+        self,
+        *,
+        job: LipsyncJobModel,
+        supplied_timings: Optional[list] = None,
+        supplied_duration: Optional[float] = None,
+    ) -> None:
+        """音频直传模式（包含 #1845 预合成路径）：同步下载 → ffprobe → timings → 提交 MediaKit.
+
+        直接在 HTTP 请求内完成，不走 Celery。job.status 成功后置为 submitted。
+        失败时把 job 标成 failed 并 commit，然后抛 MediaKitError。
+
+        Args:
+            job: 已 commit 的 LipsyncJobModel（audio_url / video_url 已写入）
+            supplied_timings: 前端传来的预合成 timings（可选，可信时直接用）
+            supplied_duration: 前端传来的预合成时长（可选，用于优先避免重复探测）
+        """
+        # 1. 下载音频
+        audio_data: bytes | None = None
+        try:
+            audio_data = safe_download_bytes(
+                job.audio_url,
+                purpose="lipsync_direct_audio",
+                allowed_mime_types=ALLOWED_AUDIO_MIME_TYPES,
+                timeout=60.0,
+            )
+            logger.info(
+                "[lipsync] 直传音频下载完成: job_id=%s size=%d",
+                job.id,
+                len(audio_data) if audio_data else 0,
+            )
+        except Exception as exc:
+            logger.warning("[lipsync] 直传音频下载失败，跳过 timings 计算: job_id=%s err=%s", job.id, exc)
+
+        # 2. ffprobe 探测时长（优先用前端传入的预合成时长，但以 ffprobe 为准做兜底校验）
+        audio_duration = 0.0
+        if audio_data:
+            audio_duration = probe_audio_duration(audio_data)
+        if audio_duration <= 0 and supplied_duration and supplied_duration > 0:
+            audio_duration = supplied_duration
+            logger.info(
+                "[lipsync] ffprobe 失败，使用前端传入的预合成时长: job_id=%s duration=%.2f", job.id, audio_duration
+            )
+
+        # 3. 句子时间戳：优先用前端预合成传入的 timings（后端预合成接口已经算过，可信）；
+        #    否则若音频下载成功则重算；否则不设置（不阻塞主流程）
+        timings: Optional[list] = None
+        if supplied_timings:
+            timings = supplied_timings
+            logger.info("[lipsync] 使用前端预合成句子时间戳: job_id=%s sentences=%d", job.id, len(timings))
+        elif audio_data and audio_duration > 0 and job.script_text:
+            try:
+                timings = compute_sentence_timings(audio_data, job.script_text, audio_duration)
+                logger.info(
+                    "[lipsync] 后端重算句子时间戳: job_id=%s sentences=%d duration=%.2f",
+                    job.id,
+                    len(timings) if timings else 0,
+                    audio_duration,
+                )
+            except Exception as exc:
+                logger.warning("[lipsync] 句子时间戳计算失败（不阻塞）: job_id=%s err=%s", job.id, exc)
+
+        if timings:
+            job.sentence_timings = timings
+
+        # 4. 签名 URL 并提交 MediaKit
+        video_url = self._sign_media_url(job.video_url)
+        signed_audio_url = self._sign_media_url(job.audio_url)
+        job.audio_url = signed_audio_url
+
+        try:
+            result = self.client.submit_lipsync(
+                video_url=video_url,
+                audio_url=signed_audio_url,
+                enable_video_loop=job.enable_video_loop,
+                client_token=job.id,
+            )
+            job.mediakit_task_id = result["task_id"]
+            job.status = "submitted"
+            job.submitted_at = datetime.now(timezone.utc)
+            self.db.commit()
+            logger.info(
+                "[lipsync] 直传音频已提交 MediaKit: job_id=%s task_id=%s",
+                job.id,
+                result["task_id"],
+            )
+        except MediaKitError as exc:
+            job.status = "failed"
+            job.error_message = str(exc)
+            job.error_code = exc.code
+            logger.error("[lipsync] 直传音频提交 MediaKit 失败: job_id=%s err=%s", job.id, exc)
+            self.db.commit()
+            raise
+
     # ── 创建任务 ──────────────────────────────────────────────────────────
 
     def create_job(
@@ -150,27 +251,35 @@ class LipsyncService:
         user_id: str,
         video_url: str,
         audio_url: str = "",
+        audio_duration: Optional[float] = None,
+        sentence_timings: Optional[list] = None,
         voice_id: str = "",
         script_text: str = "",
         speed: float = 1.0,
         emotion: str = "",
-        enable_video_loop: bool = False,
+        enable_video_loop: bool = True,
         project_id: str = "",
     ) -> LipsyncJobModel:
         """创建对口型任务.
 
-        两种输入模式：
+        三种输入模式：
         - TTS 直生：voice_id + script_text（audio_url 留空）
-          → 先创建 DB 记录（状态 tts_processing），再 dispatch Celery 异步任务
-            执行 TTS 合成 + MediaKit 提交。API 响应 <1s。
-        - 直接音频：提供 audio_url
-          → 同步提交 MediaKit，状态直接设为 submitted。
+          → 创建 DB 记录（状态 tts_processing），dispatch Celery 异步任务（降级路径）。
+          API 响应 <1s。
+        - 直接音频：audio_url 非空 + 无 sentence_timings
+          → 同步下载音频 + 重算 timings + 提交 MediaKit（几秒完成）。
+        - 预合成音频（#1845 新主路径）：audio_url 非空 + 传 sentence_timings
+          → 同步 ffprobe 校验时长 + 写入 timings + 提交 MediaKit（~2-3s）。
 
         Raises:
-            MediaKitError: 参数校验失败或 MediaKit 提交失败（仅直接音频模式）
+            MediaKitError: 参数校验失败或 MediaKit 提交失败
         """
         # 0. 输入校验
-        if not audio_url:
+        is_pre_synth = bool(audio_url) and bool(sentence_timings)
+        bool(audio_url) and not is_pre_synth
+        is_tts_mode = not bool(audio_url)
+
+        if is_tts_mode:
             if not (voice_id and script_text):
                 raise MediaKitError(
                     "必须提供 audio_url 或 voice_id+script_text",
@@ -178,10 +287,13 @@ class LipsyncService:
                 )
             # TTS 模式：在 HTTP 请求中同步校验音色归属，快速失败
             self._resolve_voice_id(voice_id, user_id)
+        elif is_pre_synth:
+            # 预合成模式：script_text 可空（因为 timings 已自带句子文本），但仍建议传
+            if not isinstance(sentence_timings, list) or len(sentence_timings) == 0:
+                raise MediaKitError("预合成模式 sentence_timings 不能为空", code="InvalidInput")
 
         # 1. 创建数据库记录
         job_id = str(uuid.uuid4())
-        is_tts_mode = not bool(audio_url)
         job = LipsyncJobModel(
             id=job_id,
             user_id=user_id,
@@ -192,20 +304,19 @@ class LipsyncService:
             voice_id=voice_id or "",
             script_text=script_text or "",
             speed=speed,
-            emotion=normalize_emotion(emotion),
+            emotion=normalize_emotion(emotion) if is_tts_mode else (emotion or ""),
+            # 音频直传（含预合成）直接进入 pending（后续同步改为 submitted）；TTS 模式进入 tts_processing
             status="tts_processing" if is_tts_mode else "pending",
         )
         self.db.add(job)
         self.db.flush()
 
-        # ⚠️ 必须先 commit 再发 Celery 任务，避免事务竞态：
-        # worker 是独立进程+独立DB连接，任务被消费（<4ms）时若本事务还未提交，
-        # worker 查询 job 会返回 None → 静默 return 不重试，job 永远卡在 tts_processing。
+        # ⚠️ 必须先 commit 再发 Celery 任务 / 后续同步操作，避免事务竞态
         self.db.commit()
         self.db.refresh(job)
 
         if is_tts_mode:
-            # 2a. TTS 模式：dispatch Celery 异步任务处理 TTS 合成 + MediaKit 提交
+            # 2a. TTS 模式：dispatch Celery 异步任务处理 TTS 合成 + MediaKit 提交（降级路径）
             try:
                 tts_synthesize_and_submit.apply_async(
                     args=(
@@ -218,8 +329,6 @@ class LipsyncService:
                     )
                 )
             except Exception as exc:
-                # 投递失败时立即把 job 标成 failed 并写入 error_message，
-                # 前端轮询时能直接看到失败原因，不会无限卡在 tts_processing。
                 logger.exception(
                     "Celery 任务提交失败，TTS 任务已创建但未触发执行: job_id=%s err=%s",
                     job_id,
@@ -229,35 +338,102 @@ class LipsyncService:
                 job.error_message = f"Celery 任务投递失败: {exc}"
                 job.error_code = "AsyncDispatchFailed"
                 job.updated_at = datetime.now(timezone.utc)
-                self.db.commit()  # 投递失败也要落库失败状态
-        else:
-            # 2b. 直接音频模式：同步签名并提交 MediaKit
-            video_url = self._sign_media_url(video_url)
-            if audio_url:
-                audio_url = self._sign_media_url(audio_url)
-                job.audio_url = audio_url
-
-            try:
-                result = self.client.submit_lipsync(
-                    video_url=video_url,
-                    audio_url=audio_url,
-                    enable_video_loop=enable_video_loop,
-                    client_token=job_id,
-                )
-                job.mediakit_task_id = result["task_id"]
-                job.status = "submitted"
-                job.submitted_at = datetime.now(timezone.utc)
-                self.db.commit()  # submitted 状态落库
-            except MediaKitError as exc:
-                job.status = "failed"
-                job.error_message = str(exc)
-                job.error_code = exc.code
-                logger.error("提交对口型任务失败: %s", exc)
                 self.db.commit()
-                raise
+        else:
+            # 2b/2c. 直接音频 / 预合成音频：同步路径
+            self._submit_audio_direct(
+                job=job,
+                supplied_timings=sentence_timings,
+                supplied_duration=audio_duration,
+            )
+            self.db.refresh(job)
 
         return job
 
+    # ── TTS 预合成（#1845 步骤1「生成配音」同步接口使用） ──────────────────
+
+    def preview_tts(
+        self,
+        *,
+        user_id: str,
+        voice_id: str,
+        script_text: str,
+        speed: float = 1.0,
+        emotion: str = "natural",
+    ) -> dict:
+        """同步做 TTS 合成 + 下载 + ffprobe + 句子时间戳计算.
+
+        不创建 LipsyncJob、不转存 OSS，直接返回 CosyVoice 临时 URL（~24h 有效期）。
+        耗时约 2-3 秒，由前端在步骤1点「生成配音」时同步等待。
+
+        Returns:
+            {"audio_url": str, "duration": float, "sentence_timings": list[dict]}
+
+        Raises:
+            MediaKitError: TTS 合成失败 / 下载失败 / ffprobe 失败
+        """
+        # 1. 音色解析（校验克隆音色归属）
+        actual_voice_id = self._resolve_voice_id(voice_id, user_id)
+        cosyvoice = self._get_cosyvoice()
+
+        # 2. TTS 合成（同步，~2-3s）
+        try:
+            result = cosyvoice.submit_synthesize_task(
+                text=script_text,
+                voice_id=actual_voice_id,
+                speed=speed,
+                emotion=normalize_emotion(emotion),
+            )
+        except CosyVoiceError as exc:
+            raise MediaKitError(f"TTS 合成失败: {exc}", code="TTSSynthesisFailed") from exc
+        except ValueError as exc:
+            raise MediaKitError(f"TTS 参数错误: {exc}", code="TTSInvalidParam") from exc
+
+        temp_url = result.get("audio_url", "")
+        if not temp_url:
+            raise MediaKitError("TTS 未返回音频 URL", code="TTSNoAudio")
+
+        # 3. 下载音频到内存（用于 ffprobe + 静音检测）
+        try:
+            audio_data = safe_download_bytes(
+                temp_url,
+                purpose="tts_preview_audio",
+                allowed_mime_types=ALLOWED_AUDIO_MIME_TYPES,
+                timeout=60.0,
+            )
+        except Exception as exc:
+            logger.warning("[tts-preview] TTS 音频下载失败，仍返回 audio_url: user_id=%s err=%s", user_id, exc)
+            return {
+                "audio_url": temp_url,
+                "duration": 0.0,
+                "sentence_timings": [],
+            }
+
+        # 4. ffprobe 时长
+        duration = probe_audio_duration(audio_data)
+        if duration <= 0:
+            logger.warning("[tts-preview] ffprobe 未返回有效时长，timings 留空: user_id=%s", user_id)
+            return {
+                "audio_url": temp_url,
+                "duration": 0.0,
+                "sentence_timings": [],
+            }
+
+        # 5. 句子时间戳
+        timings = compute_sentence_timings(audio_data, script_text, duration)
+
+        logger.info(
+            "[tts-preview] TTS 预合成完成: user_id=%s duration=%.2f sentences=%d",
+            user_id,
+            duration,
+            len(timings),
+        )
+        return {
+            "audio_url": temp_url,
+            "duration": round(duration, 2),
+            "sentence_timings": timings,
+        }
+
     # ── 查询任务 ──────────────────────────────────────────────────────────
 
     def get_job(self, job_id: str, user_id: str) -> Optional[LipsyncJobModel]:
@@ -291,11 +467,7 @@ class LipsyncService:
     # ── 更新任务状态（轮询） ──────────────────────────────────────────────
 
     def refresh_job_status(self, job_id: str, user_id: str) -> Optional[LipsyncJobModel]:
-        """从 MediaKit 拉取最新状态并更新本地记录.
-
-        Returns:
-            更新后的 Job，或 None（任务不存在/不属于该用户）
-        """
+        """从 MediaKit 拉取最新状态并更新本地记录."""
         job = self.get_job(job_id, user_id)
         if job is None:
             return None
@@ -321,13 +493,12 @@ class LipsyncService:
             result = status_data.get("result", {})
             job.status = STATUS_COMPLETED
             temp_url = result.get("video_url", "")
-            # 先以临时 URL 立即返回前端（前端可立即播放），再异步 Celery 任务转存自家 OSS（步骤⑦）
             job.output_video_url = temp_url
             job.output_duration = result.get("duration", 0.0)
             job.completed_at = datetime.now(timezone.utc)
             job.updated_at = datetime.now(timezone.utc)
             self.db.commit()
-            # 异步转存到自家 OSS（注意：必须在 commit 之后 dispatch，避免 commit 失败任务已发出）
+            # 异步转存自家 OSS
             try:
                 from app.tasks.lipsync_tts import persist_output_video_task
 
@@ -347,7 +518,6 @@ class LipsyncService:
             job.error_code = error.get("code", "TaskFailed")
             job.completed_at = datetime.now(timezone.utc)
         else:
-            # 中间状态（running/processing/queued 等）同步到 DB，避免前端永远卡在 submitted
             if isinstance(mk_status, str) and mk_status:
                 job.status = mk_status
         job.updated_at = datetime.now(timezone.utc)
@@ -356,10 +526,7 @@ class LipsyncService:
         return job
 
     def _persist_output_video(self, temp_url: str, job_id: str, user_id: str) -> str:
-        """将 MediaKit 输出的临时视频 URL 转存到自家 OSS.
-
-        失败时回退返回原始临时 URL，不影响任务完成。
-        """
+        """将 MediaKit 输出的临时视频 URL 转存到自家 OSS. 失败时回退返回原始临时 URL."""
         if not temp_url:
             return ""
         try:
@@ -379,27 +546,21 @@ class LipsyncService:
             return temp_url
 
     def _sign_media_url(self, url: str) -> str:
-        """对自家 OSS 私有桶 URL 重签长有效期预签名，供 MediaKit 拉取 / 前端播放。
-
-        - 裸 public_url（upload_file 返回，不带签名）→ 私有桶匿名访问 403，重签。
-        - 已带签名但即将过期的 URL（如前端 1h 预签名）→ 抽 storage_key 后重签。
-        - 外部 URL（CosyVoice/MediaKit 临时链接，非本桶 host）→ 原样透传。
-        - 任何异常都降级原样返回，不阻断主流程。
-        """
+        """对自家 OSS 私有桶 URL 重签长有效期预签名."""
         if not url:
             return url
         try:
             storage = get_shared_storage_service()
             public_base = getattr(storage, "public_url", "")
             if not isinstance(public_base, str) or not public_base:
-                return url  # 无法判定归属，保守透传
+                return url
             own_host = urlparse(public_base).netloc.lower()
             host = urlparse(url).netloc.lower()
             if not own_host or host != own_host:
-                return url  # 非自家 OSS（外部临时链接），不处理
+                return url  # 外部临时链接原样透传
             signed = storage.get_download_url(url, expires_seconds=MEDIAKIT_URL_TTL_SECONDS)
             return signed or url
-        except Exception as exc:  # noqa: BLE001 - 签名失败不阻断，降级原 URL
+        except Exception as exc:
             logger.warning("对口型 URL 重签失败，原样返回: url_prefix=%s err=%s", url[:80], exc)
             return url
 

apps/api/app/services/mediakit_client.py

@@ -75,7 +75,7 @@ class MediaKitClient:
         *,
         video_url: str,
         audio_url: str,
-        enable_video_loop: bool = False,
+        enable_video_loop: bool = True,
         callback_url: Optional[str] = None,
         callback_args: Optional[str] = None,
         client_token: Optional[str] = None,
@@ -103,8 +103,7 @@ class MediaKitClient:
             "video_url": video_url,
             "audio_url": audio_url,
         }
-        if enable_video_loop:
-            payload["enable_video_loop"] = True
+        payload["enable_video_loop"] = bool(enable_video_loop)
         if callback_url:
             payload["callback_url"] = callback_url
         if callback_args:

apps/api/app/tasks/lipsync_tts.py

@@ -12,6 +12,10 @@
 
 注意：使用 @shared_task 而非绑定到某个 celery_app 实例，
 确保任务能被 Worker 侧 celery_app 正确注册，同时 API 侧 send_task/apply_async 仍可正常调用。
+
+#1845：句子时间戳计算已提取至 packages/domain/sentence_timings.py，本模块保留
+_ 开头别名兼容历史导入，但 _compute_sentence_timings/_split_script_into_sentences/
+_estimate_sentence_timings_by_chars 等内部函数已复用共享实现，避免重复代码。
 """
 
 import io
@@ -21,6 +25,12 @@ from urllib.parse import urlparse
 
 from celery import shared_task
 
+# 复用共享的句子时间戳工具（#1845 配音前置）
+from packages.domain.sentence_timings import compute_sentence_timings as _compute_sentence_timings
+from packages.domain.sentence_timings import (
+    probe_audio_duration,
+)
+
 logger = logging.getLogger(__name__)
 
 # MediaKit 预签名 URL 有效期（7天，秒），与 LipsyncService._sign_media_url 保持一致
@@ -54,164 +64,6 @@ def _sign_media_url(url: str) -> str:
         return url
 
 
-def _split_script_into_sentences(script_text: str) -> list[str]:
-    """按句号/问号/感叹号/分号/逗号/换行分句（与前端 SENTENCE_SPLIT_RE 一致）.
-
-    中文短视频文案习惯用「，」断小句（如"卖花的叫花无缺，卖姜的叫姜子牙"），
-    必须把逗号也纳入分隔符，否则多句文案会被识别成一整句，导致 B-roll 时间戳错位。
-    """
-    import re
-
-    text = (script_text or "").strip()
-    if not text:
-        return []
-    parts = re.split(r"[。！？!?？！;；,，\n\r]+", text)
-    return [p.strip() for p in parts if p.strip()]
-
-
-def _compute_sentence_timings(audio_data: bytes, script_text: str, total_duration: float) -> list[dict]:
-    """基于 TTS 音频的静音检测，精确计算每句文案的起止时间.
-
-    使用 ffmpeg silencedetect 检测静音段，将静音点与句子边界对齐。
-    比字数比例估算准确得多。
-
-    Args:
-        audio_data: TTS 音频二进制数据（MP3）
-        script_text: 文案全文
-        total_duration: 音频总时长（秒）
-
-    Returns:
-        list[{"index": int, "text": str, "start_time": float, "end_time": float}]
-    """
-    import re
-    import subprocess
-    import tempfile
-
-    sentences = _split_script_into_sentences(script_text)
-    if not sentences:
-        return []
-
-    # 写入临时音频文件
-    with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
-        tmp.write(audio_data)
-        tmp_path = tmp.name
-
-    try:
-        # 用 ffmpeg silencedetect 检测静音段
-        result = subprocess.run(
-            [
-                "ffmpeg",
-                "-i",
-                tmp_path,
-                "-af",
-                "silencedetect=noise=-25dB:d=0.3",
-                "-f",
-                "null",
-                "-",
-            ],
-            capture_output=True,
-            text=True,
-            timeout=30,
-        )
-        stderr = result.stderr or ""
-
-        # 解析静音结束时间点（silence_end: X.XXX）
-        silence_ends = []
-        for match in re.finditer(r"silence_end:\s*([\d.]+)", stderr):
-            t = float(match.group(1))
-            if 0 < t < total_duration:
-                silence_ends.append(t)
-
-        # 如果没有检测到足够的静音点，降级为字数比例估算
-        if len(silence_ends) < len(sentences) - 1:
-            logger.warning(
-                "[sentence_timings] 静音点不足(%d < %d)，降级为字数比例估算",
-                len(silence_ends),
-                len(sentences) - 1,
-            )
-            return _estimate_sentence_timings_by_chars(sentences, total_duration)
-
-        # 贪心匹配：N-1 个句子边界对应 N-1 个静音点
-        # 按时间均匀分布期望值，选择最近的静音点
-        n_boundaries = len(sentences) - 1
-        boundaries = []
-        used_indices = set()
-
-        for i in range(n_boundaries):
-            # 期望的边界位置（按句子数量均匀分布）
-            expected_pos = (i + 1) / len(sentences) * total_duration
-            # 找最近的未使用静音点
-            best_idx = None
-            best_dist = float("inf")
-            for j, t in enumerate(silence_ends):
-                if j in used_indices:
-                    continue
-                dist = abs(t - expected_pos)
-                if dist < best_dist:
-                    best_dist = dist
-                    best_idx = j
-            if best_idx is not None:
-                used_indices.add(best_idx)
-                boundaries.append(silence_ends[best_idx])
-
-        boundaries.sort()
-
-        # 构建 sentence_timings
-        timings = []
-        prev_end = 0.0
-        for i, sent in enumerate(sentences):
-            start = prev_end
-            end = boundaries[i] if i < len(boundaries) else total_duration
-            timings.append(
-                {
-                    "index": i,
-                    "text": sent,
-                    "start_time": round(start, 2),
-                    "end_time": round(end, 2),
-                }
-            )
-            prev_end = end
-
-        return timings
-
-    except Exception as exc:
-        logger.warning("[sentence_timings] 静音检测异常，降级为字数比例估算: %s", exc)
-        return _estimate_sentence_timings_by_chars(sentences, total_duration)
-    finally:
-        import os
-
-        try:
-            os.unlink(tmp_path)
-        except Exception:
-            pass
-
-
-def _estimate_sentence_timings_by_chars(sentences: list[str], total_duration: float) -> list[dict]:
-    """降级方案：按字数比例估算句子时间（与原前端逻辑一致）."""
-    if not sentences or total_duration <= 0:
-        return []
-    total_chars = sum(len(s.replace(r"\s", "")) for s in sentences)
-    if total_chars == 0:
-        return []
-
-    timings = []
-    acc = 0
-    for i, sent in enumerate(sentences):
-        chars = len(sent.replace(r"\s", ""))
-        start = (acc / total_chars) * total_duration
-        end = ((acc + chars) / total_chars) * total_duration
-        timings.append(
-            {
-                "index": i,
-                "text": sent,
-                "start_time": round(start, 2),
-                "end_time": round(end, 2),
-            }
-        )
-        acc += chars
-    return timings
-
-
 @shared_task(
     bind=True,
     name="lipsync_tts.synthesize_and_submit",
@@ -234,7 +86,8 @@ def tts_synthesize_and_submit(
 ):
     """异步执行 TTS 合成 + OSS 转存 + MediaKit 提交.
 
-    在 Celery worker 中运行，不阻塞 HTTP 请求。
+    在 Celery worker 中运行，不阻塞 HTTP 请求。保留作为降级路径
+    （预合成失败 / 旧版前端未传 audio_url 时走此路径）。
     """
     from app.services.mediakit_client import MediaKitError, get_mediakit_client
     from sqlalchemy.orm import Session as DBSession
@@ -267,9 +120,6 @@ def tts_synthesize_and_submit(
 
         if job is None:
             # 事务竞态防御：API 在 commit 前投递了任务，worker 消费时事务尚未提交。
-            # Celery 内置 autoretry_for 不支持"业务条件重试"，这里手动 retry 3 次，
-            # 间隔递增（1s/3s/7s），让 API 事务有时间提交。
-            # max_retries 由 self.request(retries) 维护；默认 self.max_retries=3 由装饰器 soft_time_limit 下方指定。
             retries = getattr(self.request, "retries", 0)
             max_retries = 3
             if retries < max_retries:
@@ -340,7 +190,6 @@ def tts_synthesize_and_submit(
 
         # 2. 下载 TTS 音频到内存（用于 2.5 静音检测；不转存自家 OSS，直接使用 CosyVoice 临时 URL）
         audio_data: bytes | None = None
-        _st_tmp_path: str | None = None
         try:
             audio_data = safe_download_bytes(
                 temp_url,
@@ -349,7 +198,7 @@ def tts_synthesize_and_submit(
                     "audio/mpeg",
                     "audio/mp3",
                     "audio/wav",
-                    "audio/x-wav",  # CosyVoice 部分接口返回 audio/x-wav，与 audio/wav 等价（RIFF/WAVE）
+                    "audio/x-wav",  # CosyVoice 部分接口返回 audio/x-wav
                     "audio/mp4",
                     "audio/x-m4a",
                 },
@@ -361,7 +210,6 @@ def tts_synthesize_and_submit(
                 len(audio_data) if audio_data else 0,
             )
         except Exception as exc:
-            # 下载失败：audio_data 保持 None，2.5 静音检测会跳过；后续仍用 temp_url 提交 MediaKit
             logger.warning(
                 "[lipsync_tts] TTS 音频下载失败，跳过静音检测，直接使用临时 URL 提交: job_id=%s err=%s",
                 job_id,
@@ -373,45 +221,16 @@ def tts_synthesize_and_submit(
 
         db.commit()
 
-        # 2.5 计算精确句子时间戳（基于 TTS 音频静音检测）
-        # 直接复用步骤 2 已下载到内存的 audio_data，避免重新下载
-        import os as _os
-
+        # 2.5 计算精确句子时间戳（基于 TTS 音频静音检测）—— 复用共享工具
         try:
-            import subprocess as _sp
-            import tempfile as _tmpf
-
             if not audio_data:
                 logger.warning("[lipsync_tts] 无音频数据，跳过句子时间戳计算: job_id=%s", job_id)
             else:
-                # 写入临时文件供 ffprobe/ffmpeg 使用
-                with _tmpf.NamedTemporaryFile(suffix=".mp3", delete=False) as _atmp:
-                    _atmp.write(audio_data)
-                    _st_tmp_path = _atmp.name
-
-                # ffprobe 获取音频时长
-                _probe_result = _sp.run(
-                    [
-                        "ffprobe",
-                        "-v",
-                        "error",
-                        "-show_entries",
-                        "format=duration",
-                        "-of",
-                        "default=noprint_wrappers=1:nokey=1",
-                        _st_tmp_path,
-                    ],
-                    capture_output=True,
-                    text=True,
-                    timeout=10,
-                )
-                _audio_duration = float(_probe_result.stdout.strip()) if _probe_result.stdout.strip() else 0.0
+                _audio_duration = probe_audio_duration(audio_data)
                 logger.info(
-                    "[lipsync_tts] 音频时长探测: job_id=%s duration=%.2f probe_stdout=%s probe_stderr=%s",
+                    "[lipsync_tts] 音频时长探测: job_id=%s duration=%.2f",
                     job_id,
                     _audio_duration,
-                    _probe_result.stdout.strip()[:50],
-                    _probe_result.stderr.strip()[:100] if _probe_result.stderr else "",
                 )
 
                 if _audio_duration > 0:
@@ -428,22 +247,14 @@ def tts_synthesize_and_submit(
                         logger.warning("[lipsync_tts] 句子时间戳计算返回空结果: job_id=%s", job_id)
                 else:
                     logger.warning(
-                        "[lipsync_tts] ffprobe 未获取到有效时长，跳过句子时间戳: job_id=%s stdout=%s stderr=%s",
+                        "[lipsync_tts] ffprobe 未获取到有效时长，跳过句子时间戳: job_id=%s",
                         job_id,
-                        _probe_result.stdout.strip()[:100],
-                        _probe_result.stderr.strip()[:200] if _probe_result.stderr else "",
                     )
                 db.commit()
         except Exception as _st_err:
             logger.warning(
                 "[lipsync_tts] 句子时间戳计算失败（不影响主流程）: job_id=%s err=%s", job_id, _st_err, exc_info=True
             )
-        finally:
-            if _st_tmp_path:
-                try:
-                    _os.unlink(_st_tmp_path)
-                except Exception:
-                    pass
 
         # 3. 签名 URL 并提交到 MediaKit（复用模块内 _sign_media_url，避免对 LipsyncService 的耦合）
         audio_url = _sign_media_url(job.audio_url)
@@ -495,12 +306,7 @@ def tts_synthesize_and_submit(
     default_retry_delay=30,
 )
 def persist_output_video_task(job_id: str, user_id: str, temp_url: str):
-    """异步转存对口型输出视频到自家 OSS（步骤⑦ — 将同步阻塞挪到后台，加速前端响应）.
-
-    - MediaKit 返回 completed 后先以 temp_url 回前端（前端可立即播放临时 URL）
-    - Celery 后台下载 temp_url 并转存 OSS，成功后更新 job.output_video_url 为永久 URL
-    - 失败则保留 temp_url，不阻断主流程
-    """
+    """异步转存对口型输出视频到自家 OSS（步骤⑦ — 将同步阻塞挪到后台，加速前端响应）."""
 
     try:
         from worker_app.db import SessionLocal  # type: ignore
@@ -532,7 +338,6 @@ def persist_output_video_task(job_id: str, user_id: str, temp_url: str):
             storage = get_shared_storage_service()
             storage_key = f"lipsync-outputs/{user_id}/{job_id}.mp4"
             permanent_url = storage.upload_file(io.BytesIO(data), storage_key, content_type="video/mp4")
-            # 对自家 OSS URL 重签 7 天有效期预签名，供前端播放
             final_url = _sign_media_url(permanent_url) if permanent_url else temp_url
             job.output_video_url = final_url
             job.updated_at = datetime.now(timezone.utc)

apps/web/src/pages/ai-avatar/AiAvatar.css

@@ -1265,3 +1265,20 @@
   width: auto;
   min-width: 300px;
 }
+
+/* 渲染完成后的封面确认区 */
+.aa-finalize-section {
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  padding: 16px 0 8px;
+}
+
+.aa-finalize-cover {
+  width: 100%;
+  max-width: 240px;
+  aspect-ratio: 9/16;
+  border-radius: 12px;
+  overflow: hidden;
+  position: relative;
+}

apps/web/src/pages/ai-avatar/AiAvatarPage.tsx

@@ -1,7 +1,7 @@
 /**
- * AI数字人 — 主页面（v3 两步骤版）
- * 步骤1：出镜视频 / 配音库 / 文案
- * 步骤2：对口型预览（含插入画面）/ 标题配置 / 封面&生成
+ * AI数字人 — 主页面（v3 两步骤版 + #1845 配音前置）
+ * 步骤1：出镜视频 / 配音库 / 文案 → 点击「🎵 生成配音」做 TTS 预合成（同步，~2-3s）
+ * 步骤2：对口型预览（音频已就绪、B-roll 句子时间戳立即可用）/ 标题配置 / 封面&生成
  */
 import React, { useState, useCallback, useEffect, useRef } from "react"
 import { message } from "antd"
@@ -16,17 +16,20 @@ import PanelTitleConfig from "./components/PanelTitleConfig"
 import PanelCoverAndGenerate from "./components/PanelCoverAndGenerate"
 import { ModalAssetPicker } from "./components/ModalAssetPicker"
 import ModalBRollEditor from "./components/ModalBRollEditor"
+import ModalCoverSelect from "./components/ModalCoverSelect"
 import {
   getScripts,
   getAssetById,
   createLipsyncJob,
   getLipsyncJob,
+  previewTts,
   submitRender,
   getRenderJob,
   generateRenderSmartCover,
+  finalizeRenderJob,
 } from "./api/aiAvatar"
 import { getOrCreateDefaultProject } from "@/api/projects"
-import type { RenderJob } from "./types"
+import type { RenderJob, SentenceTiming } from "./types"
 import {
   normalizeEmotion,
   buildTitleConfigPayload,
@@ -50,6 +53,12 @@ const AiAvatarPage: React.FC = () => {
     cover: false,
   })
 
+  /* ── #1845 TTS 预合成弹窗 ── */
+  const [showTtsModal, setShowTtsModal] = useState(false)
+  const [ttsProgress, setTtsProgress] = useState(0)
+  const [ttsErrorMessage, setTtsErrorMessage] = useState("")
+  const ttsProgressTimerRef = useRef<ReturnType<typeof setInterval> | null>(null)
+
   /* ── 对口型生成弹窗 ── */
   const [showLipsyncModal, setShowLipsyncModal] = useState(false)
   const [lipsyncStatus, setLipsyncStatus] = useState<"generating" | "completed" | "failed">(
@@ -63,8 +72,12 @@ const AiAvatarPage: React.FC = () => {
   )
   const [renderProgress, setRenderProgress] = useState(0)
   const [renderErrorMessage, setRenderErrorMessage] = useState("")
-  /* ── 当前渲染任务对象（轮询更新；用于封面区判断渲染是否完成） ── */
+  /* ── 当前渲染任务对象 ── */
   const [currentRenderJob, setCurrentRenderJob] = useState<RenderJob | null>(null)
+  /* ── 封面选择弹窗 ── */
+  const [showCoverModal, setShowCoverModal] = useState(false)
+  const [selectedCoverUrl, setSelectedCoverUrl] = useState("")
+  const [finalizeLoading, setFinalizeLoading] = useState(false)
 
   /* ── 对口型轮询 ── */
   const lipsyncTimerRef = useRef<ReturnType<typeof setInterval> | null>(null)
@@ -75,8 +88,29 @@ const AiAvatarPage: React.FC = () => {
     setCollapsed((prev) => ({ ...prev, [key]: !prev[key] }))
   }, [])
 
-  /* ── 步骤切换 ── */
-  const handleNextStep = useCallback(() => {
+  /* ── #1845 文案/音色/语速变更时重置 TTS 预合成状态，避免音频与文案不一致 ── */
+  useEffect(() => {
+    if (state.ttsPreview.status !== "idle") {
+      state.resetTtsPreview()
+    }
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [state.scriptText, state.selectedVoice?.voice_id, state.speed, state.emotion])
+
+  const _clearTtsProgressTimer = useCallback(() => {
+    if (ttsProgressTimerRef.current) {
+      clearInterval(ttsProgressTimerRef.current)
+      ttsProgressTimerRef.current = null
+    }
+  }, [])
+
+  useEffect(() => {
+    return () => {
+      _clearTtsProgressTimer()
+    }
+  }, [_clearTtsProgressTimer])
+
+  /* ── #1845 步骤1：点击「🎵 生成配音」→ 同步 TTS 预合成 ── */
+  const handleGenerateTts = useCallback(async () => {
     const missing: string[] = []
     if (!state.selectedVideo) missing.push("出镜视频")
     if (!state.selectedVoice) missing.push("配音")
@@ -85,45 +119,120 @@ const AiAvatarPage: React.FC = () => {
       message.warning(`请先完成${missing.join("、")}`)
       return
     }
-    setCurrentStep(2)
-  }, [state.selectedVideo, state.selectedVoice, state.scriptText])
 
+    // 打开弹窗 & 启动模拟进度条
+    setShowTtsModal(true)
+    setTtsProgress(0)
+    setTtsErrorMessage("")
+    state.setTtsPreview({
+      audioUrl: null,
+      duration: 0,
+      sentenceTimings: [],
+      status: "generating",
+      error: null,
+    })
+
+    // 模拟进度：每 300ms +10%，到 90% 停住，真完成后瞬间到 100%
+    _clearTtsProgressTimer()
+    let fake = 0
+    ttsProgressTimerRef.current = setInterval(() => {
+      fake = Math.min(fake + 10, 90)
+      setTtsProgress(fake)
+      if (fake >= 90) {
+        _clearTtsProgressTimer()
+      }
+    }, 300)
+
+    try {
+      const res = await previewTts({
+        voice_id: state.selectedVoice!.voice_id,
+        script_text: state.scriptText,
+        speed: state.speed,
+        emotion: normalizeEmotion(state.emotion),
+      })
+      _clearTtsProgressTimer()
+      setTtsProgress(100)
+      state.setTtsPreview({
+        audioUrl: res.audio_url,
+        duration: res.duration,
+        sentenceTimings: res.sentence_timings as SentenceTiming[],
+        status: "done",
+        error: null,
+      })
+      message.success("配音合成完成")
+    } catch (err) {
+      _clearTtsProgressTimer()
+      const errMsg =
+        (err as { response?: { data?: { message?: string; detail?: unknown } } })?.response?.data
+          ?.message || (err instanceof Error ? err.message : "配音合成失败，请重试")
+      setTtsErrorMessage(typeof errMsg === "string" ? errMsg : "配音合成失败，请重试")
+      state.setTtsPreview({
+        audioUrl: null,
+        duration: 0,
+        sentenceTimings: [],
+        status: "failed",
+        error: typeof errMsg === "string" ? errMsg : "配音合成失败",
+      })
+    }
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [state.selectedVideo, state.selectedVoice, state.scriptText, state.speed, state.emotion])
+
+  const handleRetryTts = useCallback(() => {
+    handleGenerateTts()
+  }, [handleGenerateTts])
+
+  const handleTtsNext = useCallback(() => {
+    setShowTtsModal(false)
+    setTtsProgress(0)
+    setCurrentStep(2)
+  }, [])
+
+  const handleCancelTts = useCallback(() => {
+    _clearTtsProgressTimer()
+    setShowTtsModal(false)
+    setTtsProgress(0)
+    setTtsErrorMessage("")
+    // 若用户在生成中途关闭，把状态重置回 idle，允许重新点击
+    if (state.ttsPreview.status === "generating") {
+      state.resetTtsPreview()
+    }
+  }, [_clearTtsProgressTimer, state])
+
+  /* ── 上一步（返回步骤1，不会丢失 TTS 预合成结果） ── */
   const handlePrevStep = useCallback(() => {
     setCurrentStep(1)
   }, [])
 
   /* ── 对口型 ── */
   const handleGenerateLipsync = useCallback(async () => {
-    // ② 缺项明确提示（#1809）：不再静默 return
     const video = state.selectedVideo
-    const voice = state.selectedVoice
     const text = state.scriptText.trim()
     const missing: string[] = []
     if (!video) missing.push("出镜视频")
-    if (!voice) missing.push("音色")
     if (!text) missing.push("文案")
-    if (missing.length > 0 || !video || !voice) {
+    if (missing.length > 0 || !video) {
       message.warning(`请先选择${missing.join("、")}`)
       return
     }
+
+    // #1845：预合成模式下必须要有 audioUrl（理论上到了步骤2肯定有，兜底防御）
+    const isPreSynth = state.ttsPreview.status === "done" && !!state.ttsPreview.audioUrl
+    if (!isPreSynth && !state.selectedVoice) {
+      message.warning("请先选择音色或完成配音合成")
+      return
+    }
+
     try {
-      // 显示生成弹窗
       setShowLipsyncModal(true)
       setLipsyncStatus("generating")
       setLipsyncErrorMessage("")
 
-      // ① 先按素材 id 拿 file_url（#1809 补充：对齐后端新参数 video_url）
       console.log("[对口型] 开始生成:", {
         videoId: video.id,
-        voiceId: voice.voice_id,
-        voiceType: voice.type,
+        mode: isPreSynth ? "pre-synth" : "tts-direct",
         textLen: state.scriptText.length,
       })
       const asset = await getAssetById(video.id)
-      console.log("[对口型] getAssetById 响应:", {
-        id: asset?.id,
-        file_url: asset?.file_url?.substring(0, 100),
-      })
       const videoUrl = asset?.file_url
       if (!videoUrl) {
         console.error("[对口型] file_url 为空，asset:", asset)
@@ -131,19 +240,35 @@ const AiAvatarPage: React.FC = () => {
         message.error("获取出镜视频播放地址失败，请重新选择素材")
         return
       }
-      // ② 模式A TTS直生：video_url + voice_id + script_text，语速/情绪英文枚举透传（#1822）
-      const payload = {
-        voice_id: voice.voice_id,
-        script_text: state.scriptText,
-        video_url: videoUrl,
-        speed: state.speed, // 语速 0.5~2.0
-        emotion: normalizeEmotion(state.emotion), // natural/excited/calm/friendly
+
+      type LipsyncPayload = Parameters<typeof createLipsyncJob>[0]
+      let payload: LipsyncPayload
+      if (isPreSynth) {
+        // 预合成模式：传 audio_url + audio_duration + sentence_timings（后端直接提交 MediaKit，~2-3s）
+        payload = {
+          video_url: videoUrl,
+          audio_url: state.ttsPreview.audioUrl!,
+          audio_duration: state.ttsPreview.duration,
+          sentence_timings: state.ttsPreview.sentenceTimings,
+          enable_video_loop: true,
+        }
+      } else {
+        // 降级：TTS 直生（旧路径，前端未预合成时）
+        payload = {
+          voice_id: state.selectedVoice!.voice_id,
+          script_text: state.scriptText,
+          video_url: videoUrl,
+          speed: state.speed,
+          emotion: normalizeEmotion(state.emotion),
+        }
       }
       console.log("[对口型] createLipsyncJob 请求:", payload)
       const job = await createLipsyncJob(payload)
       console.log("[对口型] createLipsyncJob 响应:", { id: job.id, status: job.status })
       state.setLipsyncJob(job)
-      // 开始轮询
+
+      // 如果是预合成模式，后端会同步把状态置为 submitted（甚至可能已返回 running），
+      // 但仍需轮询等 completed
       if (lipsyncTimerRef.current) clearInterval(lipsyncTimerRef.current)
       lipsyncTimerRef.current = setInterval(async () => {
         try {
@@ -180,7 +305,14 @@ const AiAvatarPage: React.FC = () => {
       message.error(err instanceof Error ? err.message : "对口型任务提交失败，请重试")
     }
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [state.selectedVideo, state.selectedVoice, state.scriptText, state.speed, state.emotion])
+  }, [
+    state.selectedVideo,
+    state.selectedVoice,
+    state.scriptText,
+    state.speed,
+    state.emotion,
+    state.ttsPreview,
+  ])
 
   // 取消对口型生成
   const handleCancelLipsync = useCallback(() => {
@@ -201,6 +333,14 @@ const AiAvatarPage: React.FC = () => {
     }
   }, [])
 
+  /* ── B-roll 弹窗可用的句子时间戳：优先 lipsyncJob.sentence_timings，否则用 ttsPreview.sentenceTimings ── */
+  const bRollSentenceTimings: SentenceTiming[] | undefined =
+    (state.lipsyncJob?.sentence_timings as SentenceTiming[] | undefined) ??
+    (state.ttsPreview.status === "done" ? state.ttsPreview.sentenceTimings : undefined)
+
+  /* ── B-roll 可用的总时长：优先 lipsyncJob.output_duration，否则用 ttsPreview.duration ── */
+  const bRollDuration = state.lipsyncJob?.output_duration || state.ttsPreview.duration || 0
+
   /* ── 生成视频（含实时进度轮询） ── */
   const handleGenerate = useCallback(async () => {
     if (!state.lipsyncJob || state.lipsyncJob.status !== "completed") {
@@ -209,10 +349,9 @@ const AiAvatarPage: React.FC = () => {
     }
     state.setIsGenerating(true)
     try {
-      // 确保有 project_id（AI数字人入口独立，不在项目内，自动取默认项目；#1860 P0 bugfix）
       const defaultProject = await getOrCreateDefaultProject()
 
-      // 用 Canvas 预渲染标题为 PNG dataURL（所见即所得，后端用 overlay 直接叠加）
+      // 用 Canvas 预渲染标题为 PNG dataURL
       let titleImageDataUrl: string | null = null
       if (state.titleConfig.title?.trim()) {
         try {
@@ -242,7 +381,6 @@ const AiAvatarPage: React.FC = () => {
           pip_scale: seg.pip_scale,
         })) as never,
         title_config: buildTitleConfigPayload(state.titleConfig, titleImageDataUrl),
-        // 封面不阻塞渲染：用户未选定封面时传空 dict，后端不生成封面；渲染完成后再单独抽帧
         cover_config:
           state.coverConfig.smart_cover_url ||
           (state.coverConfig.upload_url && !state.coverConfig.upload_url.startsWith("blob:"))
@@ -250,12 +388,14 @@ const AiAvatarPage: React.FC = () => {
             : {},
       })
 
-      // 打开渲染进度弹窗，启动轮询
       setShowRenderModal(true)
       setRenderStatus("generating")
       setRenderProgress(job.progress ?? 0)
       setRenderErrorMessage("")
       setCurrentRenderJob(job as RenderJob)
+      // 每次新渲染重置封面状态
+      setSelectedCoverUrl("")
+      setShowCoverModal(false)
 
       if (renderTimerRef.current) clearInterval(renderTimerRef.current)
       renderTimerRef.current = setInterval(async () => {
@@ -267,8 +407,8 @@ const AiAvatarPage: React.FC = () => {
             if (renderTimerRef.current) clearInterval(renderTimerRef.current)
             renderTimerRef.current = null
             setRenderStatus("completed")
-            // 渲染完成后：如果后端已返回封面（用户预上传/预设）则同步到前端；
-            // 否则不自动设置封面，由用户在封面区点击"智能获取封面"主动抽帧（步骤③④）
+            // 不再自动入库/自动跳转：渲染完成后停留在主页面，等用户选封面、点「完成」才入库
+            // 如果后端在透传时已经带了封面（旧逻辑兜底），同步本地状态
             if (updated.output_cover_url) {
               state.setCoverConfig((prev) => ({
                 ...prev,
@@ -276,8 +416,9 @@ const AiAvatarPage: React.FC = () => {
                 smart_cover_url: updated.output_cover_url,
                 thumbnail_url: updated.output_cover_url,
               }))
+              setSelectedCoverUrl(updated.output_cover_url)
             }
-            message.success("视频已生成并保存到成片库")
+            message.success("视频生成完成，请选择封面")
           } else if (updated.status === "failed") {
             if (renderTimerRef.current) clearInterval(renderTimerRef.current)
             renderTimerRef.current = null
@@ -309,7 +450,7 @@ const AiAvatarPage: React.FC = () => {
     setRenderErrorMessage("")
   }, [])
 
-  /* ── 智能封面：从最终渲染成片抽帧（POST /renders/{id}/smart-cover，步骤③④） ── */
+  /* ── 智能封面 ── */
   const handleGenerateRenderSmartCover = useCallback(
     async (renderId: string): Promise<{ cover_url: string; message?: string }> => {
       try {
@@ -334,11 +475,72 @@ const AiAvatarPage: React.FC = () => {
         return { cover_url: "", message: errMsg }
       }
     },
-    // state.setCoverConfig 是 zustand action 引用稳定，eslint 不需要检查
     // eslint-disable-next-line react-hooks/exhaustive-deps
     [],
   )
 
+  /* ── 封面弹窗回调 ── */
+  const handleCoverSelected = useCallback((coverUrl: string) => {
+    setSelectedCoverUrl(coverUrl || "")
+  }, [])
+
+  const handleOpenCoverModal = useCallback(() => {
+    if (currentRenderJob?.status !== "completed") {
+      message.warning("请先完成视频生成")
+      return
+    }
+    setShowCoverModal(true)
+  }, [currentRenderJob])
+
+  const handleCloseCoverModal = useCallback(() => {
+    setShowCoverModal(false)
+  }, [])
+
+  /* ── 自定义上传封面（本地预览，不单独上传；点完成时一起入库） ── */
+  const handleUploadCover = useCallback(
+    (file: File) => {
+      const url = URL.createObjectURL(file)
+      state.setCoverConfig((prev) => ({
+        ...prev,
+        mode: "upload",
+        upload_url: url,
+        thumbnail_url: url,
+      }))
+      setSelectedCoverUrl(url)
+    },
+    [state],
+  )
+
+  /* ── 点「完成」：调用 finalize 入库成片库，成功后跳转到成片库 ── */
+  const handleFinalize = useCallback(async () => {
+    if (!currentRenderJob?.id) {
+      message.error("渲染任务不存在")
+      return
+    }
+    if (currentRenderJob.status !== "completed") {
+      message.warning("请先完成视频生成")
+      return
+    }
+    setFinalizeLoading(true)
+    try {
+      const res = await finalizeRenderJob(currentRenderJob.id)
+      if (res.data?.status === "success" || res.data?.status === "already_finalized") {
+        message.success("已保存到成片库")
+        navigate("/app/products")
+      } else {
+        message.error("保存失败，请重试")
+      }
+    } catch (err) {
+      console.error("finalize 失败:", err)
+      const errMsg =
+        (err as { response?: { data?: { detail?: unknown } } })?.response?.data?.detail ||
+        (err instanceof Error ? err.message : "保存到成片库失败")
+      message.error(typeof errMsg === "string" ? errMsg : "保存到成片库失败")
+    } finally {
+      setFinalizeLoading(false)
+    }
+  }, [currentRenderJob, navigate])
+
   /* ── 配置汇总 ── */
   const coverStatus: "not_ready" | "pending" | "selected" = (() => {
     if (
@@ -431,19 +633,33 @@ const AiAvatarPage: React.FC = () => {
                   onOpenScriptModal={() => state.setShowScriptModal(true)}
                 />
                 <div className="aa-step-btn-row">
-                  <button type="button" className="aa-btn aa-btn--primary" onClick={handleNextStep}>
-                    下一步 →
+                  <button
+                    type="button"
+                    className="aa-btn aa-btn--primary"
+                    onClick={handleGenerateTts}
+                    disabled={state.ttsPreview.status === "generating"}
+                  >
+                    {state.ttsPreview.status === "done" ? "🎵 重新生成配音" : "🎵 生成配音"}
                   </button>
+                  {state.ttsPreview.status === "done" && (
+                    <button
+                      type="button"
+                      className="aa-btn aa-btn--primary"
+                      onClick={() => setCurrentStep(2)}
+                      style={{ marginLeft: 12 }}
+                    >
+                      下一步 →
+                    </button>
+                  )}
                 </div>
               </div>
             </div>
           </>
         )}
 
-        {/* ════ 步骤 2：对口型预览（含插入画面）/ 标题配置 / 封面&生成 ════ */}
+        {/* ════ 步骤 2：对口型预览 / 标题配置 / 封面&生成 ════ */}
         {currentStep === 2 && (
           <>
-            {/* 面板：对口型预览 + 插入画面 */}
             <div className={`aa-panel aa-panel--s2-wide${collapsed.lipsync ? " collapsed" : ""}`}>
               <div className="aa-panel__header" onClick={() => togglePanel("lipsync")}>
                 <span className="aa-panel__title">对口型预览</span>
@@ -481,27 +697,98 @@ const AiAvatarPage: React.FC = () => {
               </div>
             </div>
 
-            {/* 面板5：封面 & 生成 */}
+            {/* 面板5：封面 & 生成
+                - 渲染未完成：显示分辨率/配置摘要/「开始生成」按钮（PanelCoverAndGenerate setup 变体，无封面区）
+                - 渲染完成：显示封面预览 + 「🎬 选择封面」/「✅ 完成」按钮，封面选择在弹窗中完成 */}
             <div className={`aa-panel aa-panel--s2${collapsed.cover ? " collapsed" : ""}`}>
               <div className="aa-panel__header" onClick={() => togglePanel("cover")}>
-                <span className="aa-panel__title">封面 & 生成</span>
+                <span className="aa-panel__title">
+                  {currentRenderJob?.status === "completed" ? "视频已生成" : "封面 & 生成"}
+                </span>
                 <span className="aa-panel__toggle">▼</span>
               </div>
               <div className="aa-panel__body">
-                <PanelCoverAndGenerate
-                  coverConfig={state.coverConfig}
-                  onCoverConfigChange={(partial) =>
-                    state.setCoverConfig((prev) => ({ ...prev, ...partial }))
-                  }
-                  titleConfig={state.titleConfig}
-                  renderJob={currentRenderJob}
-                  onGenerateRenderSmartCover={handleGenerateRenderSmartCover}
-                  resolution={state.resolution}
-                  onResolutionChange={state.setResolution}
-                  isGenerating={state.isGenerating}
-                  onGenerate={handleGenerate}
-                  summary={summary}
-                />
+                {currentRenderJob?.status !== "completed" ? (
+                  <PanelCoverAndGenerate
+                    variant="setup"
+                    coverConfig={state.coverConfig}
+                    onCoverConfigChange={(partial) =>
+                      state.setCoverConfig((prev) => ({ ...prev, ...partial }))
+                    }
+                    renderJob={currentRenderJob}
+                    onGenerateRenderSmartCover={handleGenerateRenderSmartCover}
+                    resolution={state.resolution}
+                    onResolutionChange={state.setResolution}
+                    isGenerating={state.isGenerating}
+                    onGenerate={handleGenerate}
+                    summary={summary}
+                  />
+                ) : (
+                  <div className="aa-finalize-section">
+                    <div
+                      style={{ marginBottom: 8, fontSize: 13, color: "#1a1a2e", fontWeight: 500 }}
+                    >
+                      ✅ 视频生成完成，请选择封面后点「完成」入库
+                    </div>
+                    <div className="aa-finalize-cover">
+                      {selectedCoverUrl ? (
+                        <img
+                          src={selectedCoverUrl}
+                          alt="封面预览"
+                          draggable={false}
+                          style={{
+                            width: "100%",
+                            height: "100%",
+                            objectFit: "cover",
+                            borderRadius: 8,
+                          }}
+                        />
+                      ) : (
+                        <div
+                          style={{
+                            width: "100%",
+                            height: "100%",
+                            border: "2px dashed #d9d9d9",
+                            borderRadius: 8,
+                            display: "flex",
+                            alignItems: "center",
+                            justifyContent: "center",
+                            color: "#8c8ca1",
+                            fontSize: 12,
+                            flexDirection: "column",
+                            gap: 4,
+                          }}
+                        >
+                          <span style={{ fontSize: 24 }}>🎬</span>
+                          <span>尚未选择封面</span>
+                        </div>
+                      )}
+                    </div>
+                    <div
+                      style={{
+                        display: "flex",
+                        gap: 10,
+                        marginTop: 12,
+                      }}
+                    >
+                      <button
+                        type="button"
+                        className="aa-btn aa-btn--ghost"
+                        onClick={handleOpenCoverModal}
+                      >
+                        🎬 选择封面
+                      </button>
+                      <button
+                        type="button"
+                        className="aa-btn aa-btn--primary"
+                        onClick={handleFinalize}
+                        disabled={finalizeLoading}
+                      >
+                        {finalizeLoading ? "⏳ 保存中..." : "✅ 完成"}
+                      </button>
+                    </div>
+                  </div>
+                )}
               </div>
             </div>
           </>
@@ -527,20 +814,144 @@ const AiAvatarPage: React.FC = () => {
         />
       )}
 
-      {/* B-roll 编辑器弹窗 */}
+      {/* B-roll 编辑器弹窗 — #1845：timings 在对口型完成前就可用（来自 TTS 预合成） */}
       {state.showBRollModal && (
         <ModalBRollEditor
           open={state.showBRollModal}
           onClose={() => state.setShowBRollModal(false)}
           existingSegments={state.bRollSegments}
           scriptText={state.lipsyncJob?.script_text || state.scriptText}
-          outputDuration={state.lipsyncJob?.output_duration ?? 0}
-          sentenceTimings={state.lipsyncJob?.sentence_timings}
+          outputDuration={bRollDuration}
+          sentenceTimings={bRollSentenceTimings}
           onConfirm={state.addBRollSegment}
           onRemove={state.removeBRollSegment}
         />
       )}
 
+      {/* #1845 TTS 预合成弹窗 */}
+      {showTtsModal && (
+        <div className="aa-modal-overlay">
+          <div className="aa-modal" onClick={(e) => e.stopPropagation()}>
+            <div className="aa-modal__header">
+              <span className="aa-modal__title">配音合成中</span>
+              {state.ttsPreview.status !== "generating" && (
+                <button className="aa-modal__close" onClick={handleCancelTts}>
+                  ✕
+                </button>
+              )}
+            </div>
+            <div
+              className="aa-modal__body"
+              style={{
+                display: "flex",
+                flexDirection: "column",
+                alignItems: "center",
+                padding: "40px 20px",
+              }}
+            >
+              {state.ttsPreview.status === "generating" && (
+                <>
+                  <div className="aa-lipsync-spinner" />
+                  <div style={{ marginTop: 20, fontSize: 15, color: "#1a1a2e" }}>
+                    正在合成配音，请稍候…
+                  </div>
+                  <div
+                    style={{
+                      marginTop: 20,
+                      fontSize: 32,
+                      fontWeight: 700,
+                      color: "#1890ff",
+                    }}
+                  >
+                    {ttsProgress}%
+                  </div>
+                  <div
+                    style={{
+                      marginTop: 12,
+                      width: "80%",
+                      height: 8,
+                      backgroundColor: "#f0f0f0",
+                      borderRadius: 4,
+                      overflow: "hidden",
+                    }}
+                  >
+                    <div
+                      style={{
+                        width: `${ttsProgress}%`,
+                        height: "100%",
+                        backgroundColor: "#1890ff",
+                        borderRadius: 4,
+                        transition: "width 0.3s ease",
+                      }}
+                    />
+                  </div>
+                  <div style={{ marginTop: 12, fontSize: 13, color: "#8c8ca1" }}>
+                    请勿关闭页面，完成后将自动提示
+                  </div>
+                </>
+              )}
+              {state.ttsPreview.status === "done" && (
+                <>
+                  <div style={{ fontSize: 48 }}>✅</div>
+                  <div style={{ marginTop: 16, fontSize: 15, color: "#1a1a2e" }}>
+                    配音合成完成，点击下一步继续
+                  </div>
+                  <div style={{ marginTop: 8, fontSize: 13, color: "#8c8ca1" }}>
+                    音频时长 {state.ttsPreview.duration.toFixed(1)}s，共{" "}
+                    {state.ttsPreview.sentenceTimings.length} 句
+                  </div>
+                </>
+              )}
+              {state.ttsPreview.status === "failed" && (
+                <>
+                  <div style={{ fontSize: 48 }}>❌</div>
+                  <div style={{ marginTop: 16, fontSize: 15, color: "#1a1a2e" }}>配音合成失败</div>
+                  {ttsErrorMessage && (
+                    <div
+                      style={{
+                        marginTop: 8,
+                        fontSize: 13,
+                        color: "#ff4d4f",
+                        textAlign: "center",
+                        padding: "0 20px",
+                      }}
+                    >
+                      {ttsErrorMessage}
+                    </div>
+                  )}
+                </>
+              )}
+            </div>
+            <div className="aa-modal__footer">
+              {state.ttsPreview.status === "generating" && (
+                <button className="aa-btn aa-btn--danger" onClick={handleCancelTts}>
+                  取消
+                </button>
+              )}
+              {state.ttsPreview.status === "done" && (
+                <button className="aa-btn aa-btn--primary" onClick={handleTtsNext}>
+                  下一步 →
+                </button>
+              )}
+              {state.ttsPreview.status === "failed" && (
+                <>
+                  <button className="aa-btn" onClick={handleCancelTts}>
+                    关闭
+                  </button>
+                  <button
+                    className="aa-btn aa-btn--primary"
+                    onClick={handleRetryTts}
+                    style={{ marginLeft: 12 }}
+                  >
+                    重试
+                  </button>
+                </>
+              )}
+            </div>
+          </div>
+        </div>
+      )}
+
       {/* 对口型生成弹窗 */}
       {showLipsyncModal && (
         <div className="aa-modal-overlay">
@@ -673,17 +1084,21 @@ const AiAvatarPage: React.FC = () => {
                 <>
                   <div style={{ fontSize: 48 }}>✅</div>
                   <div style={{ marginTop: 16, fontSize: 15, color: "#1a1a2e" }}>
-                    视频已保存到成片库
+                    视频生成完成，请选择封面
+                  </div>
+                  <div
+                    style={{ marginTop: 8, fontSize: 13, color: "#8c8ca1", textAlign: "center" }}
+                  >
+                    关闭此窗口后可在下方选择封面并保存到成片库
                   </div>
                   <button
-                    className="aa-btn"
+                    className="aa-btn aa-btn--primary"
                     style={{ marginTop: 16 }}
                     onClick={() => {
                       setShowRenderModal(false)
-                      navigate("/app/products")
                     }}
                   >
-                    📁 查看成片
+                    🎬 选择封面
                   </button>
                 </>
               )}
@@ -714,6 +1129,18 @@ const AiAvatarPage: React.FC = () => {
           </div>
         </div>
       )}
+
+      {/* 封面选择弹窗 */}
+      <ModalCoverSelect
+        open={showCoverModal}
+        onClose={handleCloseCoverModal}
+        renderJob={currentRenderJob}
+        coverConfig={state.coverConfig}
+        onCoverConfigChange={(partial) => state.setCoverConfig((prev) => ({ ...prev, ...partial }))}
+        onGenerateRenderSmartCover={handleGenerateRenderSmartCover}
+        onUploadCover={handleUploadCover}
+        onCoverSelected={handleCoverSelected}
+      />
     </div>
   )
 }

apps/web/src/pages/ai-avatar/api/aiAvatar.ts

@@ -2,7 +2,7 @@
  * AI数字人 — API 封装（#1822 契约对齐）
  */
 import apiClient from "@/api/client"
-import type { Script, LipsyncJob, RenderJob, BRollSegment } from "../types"
+import type { Script, LipsyncJob, RenderJob, BRollSegment, SentenceTiming } from "../types"
 
 /* ── 文案库 ── */
 export const getScripts = async (): Promise<Script[]> => {
@@ -34,17 +34,28 @@ export const getAssetById = async (id: string): Promise<{ file_url?: string; id:
   return response.data
 }
 
-/* ── 对口型（模式A：TTS 直生，后端内部合成音频；不要先调 TTS 拿 audio_url） ── */
+/* ── 对口型（支持三种模式） ──
+ * 1. TTS 直生（降级/旧版）：传 voice_id + script_text（+speed/emotion），后端 Celery 异步合成
+ * 2. 直接音频：传 video_url + audio_url，后端同步下载+算timings+提交MediaKit
+ * 3. 预合成音频（#1845 新主路径）：先调 previewTts 拿 audio_url+sentence_timings，
+ *    再把 audio_url + audio_duration + sentence_timings 一起传过来，后端直接提交 MediaKit
+ */
 export const createLipsyncJob = async (data: {
-  /** 人物视频 URL（MP4）；由素材 id 经 getAssetById 拿 file_url，禁止传 video_asset_id */
+  /** 人物视频 URL（MP4）；由素材 id 经 getAssetById 拿 file_url */
   video_url: string
-  /** 音色 ID（预置音色 或 克隆音色 profile UUID，后端会解析） */
-  voice_id: string
-  /** 要合成的文案（手动输入或文案库内容） */
-  script_text: string
-  /** 语速 0.5~2.0，默认 1.0 */
+  /** 预合成/直接音频模式：音频 URL（#1845 步骤1 预合成的 CosyVoice 临时 URL，或外部音频 URL） */
+  audio_url?: string
+  /** 预合成音频时长（秒），由 previewTts 返回 */
+  audio_duration?: number
+  /** 预合成接口返回的句子时间戳（精确），后端直接写入 job */
+  sentence_timings?: SentenceTiming[]
+  /** 音色 ID（TTS 直生模式用） */
+  voice_id?: string
+  /** 要合成的文案（TTS 直生模式用） */
+  script_text?: string
+  /** 语速 0.5~2.0，默认 1.0（TTS 直生模式用） */
   speed?: number
-  /** 情绪英文枚举：natural/excited/calm/friendly */
+  /** 情绪英文枚举：natural/excited/calm/friendly（TTS 直生模式用） */
   emotion?: string
   enable_video_loop?: boolean
   project_id?: string
@@ -53,6 +64,25 @@ export const createLipsyncJob = async (data: {
   return response.data
 }
 
+/* ── #1845 TTS 预合成（步骤1「生成配音」同步接口，~2-3s） ── */
+export const previewTts = async (data: {
+  voice_id: string
+  script_text: string
+  speed?: number
+  emotion?: string
+}): Promise<{
+  audio_url: string
+  duration: number
+  sentence_timings: SentenceTiming[]
+}> => {
+  const response = await apiClient.post<{
+    audio_url: string
+    duration: number
+    sentence_timings: SentenceTiming[]
+  }>("/lipsync/tts-preview", data, { timeout: 30000 })
+  return response.data
+}
+
 export const getLipsyncJob = async (id: string): Promise<LipsyncJob> => {
   const response = await apiClient.get<LipsyncJob>(`/lipsync/jobs/${id}`, { timeout: 60000 })
   return response.data
@@ -93,3 +123,9 @@ export const generateRenderSmartCover = async (
   )
   return response.data
 }
+
+/* ── 封面选定后点「完成」正式入库（POST /ai-avatar/render/{job_id}/finalize） ── */
+export const finalizeRenderJob = (renderId: string) =>
+  apiClient.post<{ video_id: string; cover_url: string; status: string }>(
+    `/ai-avatar/render/${renderId}/finalize`,
+  )

apps/web/src/pages/ai-avatar/components/ModalCoverSelect.tsx

@@ -0,0 +1,59 @@
+/**
+ * AI数字人 — 封面选择弹窗
+ * 渲染完成后由主页面唤起，内部用 PanelCoverAndGenerate（select-cover 变体）提供
+ * 智能抽帧 + 自定义上传 + 预览 + 确定按钮。
+ */
+import React from "react"
+import type { AiAvatarCoverConfig, RenderJob } from "../types"
+import PanelCoverAndGenerate from "./PanelCoverAndGenerate"
+
+interface ModalCoverSelectProps {
+  open: boolean
+  onClose: () => void
+  renderJob: RenderJob | null
+  coverConfig: AiAvatarCoverConfig
+  onCoverConfigChange: (partial: Partial<AiAvatarCoverConfig>) => void
+  onGenerateRenderSmartCover: (renderId: string) => Promise<{ cover_url: string; message?: string }>
+  onUploadCover?: (file: File) => void
+  onCoverSelected: (coverUrl: string) => void
+}
+
+const ModalCoverSelect: React.FC<ModalCoverSelectProps> = ({
+  open,
+  onClose,
+  renderJob,
+  coverConfig,
+  onCoverConfigChange,
+  onGenerateRenderSmartCover,
+  onUploadCover,
+  onCoverSelected,
+}) => {
+  if (!open) return null
+
+  return (
+    <div className="aa-modal-overlay" onClick={onClose}>
+      <div className="aa-modal" onClick={(e) => e.stopPropagation()} style={{ maxWidth: 480 }}>
+        <div className="aa-modal__header">
+          <span className="aa-modal__title">选择封面</span>
+          <button type="button" className="aa-modal__close" onClick={onClose} aria-label="关闭">
+            ×
+          </button>
+        </div>
+        <div className="aa-modal__body" style={{ padding: 20 }}>
+          <PanelCoverAndGenerate
+            variant="select-cover"
+            coverConfig={coverConfig}
+            onCoverConfigChange={onCoverConfigChange}
+            renderJob={renderJob}
+            onGenerateRenderSmartCover={onGenerateRenderSmartCover}
+            onUploadCover={onUploadCover}
+            onClose={onClose}
+            onCoverSelected={onCoverSelected}
+          />
+        </div>
+      </div>
+    </div>
+  )
+}
+
+export default ModalCoverSelect

apps/web/src/pages/ai-avatar/components/PanelCoverAndGenerate.tsx

@@ -1,32 +1,37 @@
 /**
- * AI数字人 — 面板5：分辨率/配置摘要/生成按钮/封面
- * v3 调整（步骤③④）：
- * - 布局顺序：分辨率 → 配置摘要卡片 → 🔘「开始生成视频」按钮 → （渲染完成后）封面区域
- * - 渲染未完成时封面区域显示占位态，按钮 disabled
- * - 「智能获取封面」从最终成片抽帧（调用 POST /renders/{id}/smart-cover），不再依赖 lipsync 状态
- * - 修复点 2 次 bug：内部维护 smartCoverLoading，不依赖外层异步 state 更新
+ * AI数字人 — 面板5 / 封面选择弹窗内容：
+ * - variant="setup"（默认）：分辨率 / 配置摘要 / 「开始生成视频」按钮，用于主页面步骤2配置阶段；
+ *   渲染完成后仍内嵌封面预览与按钮，方便不打开弹窗直接操作。
+ * - variant="select-cover"：只渲染封面选择区（智能获取封面 + 自定义上传 + 预览），
+ *   用于 ModalCoverSelect 弹窗中；传 onClose 时底部显示「确定」按钮。
  *
- * 注意：v3 已删除"画面插入模式"，本面板不包含该选项。
+ * 封面一律从最终成片（已叠加标题/B-roll）抽帧，本面板不再叠加标题。
  */
-import React, { useMemo, useRef, useState } from "react"
-import type { AiAvatarCoverConfig, AiAvatarTitleConfig, RenderJob } from "../types"
+import React, { useRef, useState } from "react"
+import type { AiAvatarCoverConfig, RenderJob } from "../types"
+
+type PanelVariant = "setup" | "select-cover"
 
 interface PanelCoverAndGenerateProps {
+  variant?: PanelVariant
   coverConfig: AiAvatarCoverConfig
-  titleConfig: AiAvatarTitleConfig
   onCoverConfigChange: (partial: Partial<AiAvatarCoverConfig>) => void
-  resolution: string
-  onResolutionChange: (r: string) => void
-  isGenerating: boolean
-  onGenerate: () => void
+  resolution?: string
+  onResolutionChange?: (r: string) => void
+  isGenerating?: boolean
+  onGenerate?: () => void
   /** 当前渲染任务（渲染完成后才有 output_video_url，才能抽封面） */
   renderJob: RenderJob | null
   /** 从最终成片智能抽帧（参数 renderId），返回 { cover_url } */
   onGenerateRenderSmartCover: (renderId: string) => Promise<{ cover_url: string; message?: string }>
   /** 自定义上传封面（选择本地文件后由父组件处理实际上传） */
   onUploadCover?: (file: File) => void
-  /** 配置汇总信息 */
-  summary: {
+  /** 弹窗关闭回调（传入则表示在弹窗中使用，底部显示「确定」按钮） */
+  onClose?: () => void
+  /** 封面选好（智能抽帧/自定义上传成功）后通知父组件，参数为封面 URL */
+  onCoverSelected?: (coverUrl: string) => void
+  /** 配置汇总信息（仅 variant="setup" 使用） */
+  summary?: {
     videoName: string | null
     voiceName: string | null
     scriptLength: number
@@ -52,27 +57,19 @@ const LIPSYNC_STATUS_LABEL: Record<string, { text: string; cls: string }> = {
   failed: { text: "失败", cls: "aa-status-badge--failed" },
 }
 
-/** 字体名 → CSS font-family 映射（与后端 drawtext 对齐） */
-const FONT_FAMILY_MAP: Record<string, string> = {
-  思源黑体: "'Noto Sans SC', 'Source Han Sans SC', 'PingFang SC', 'Microsoft YaHei', sans-serif",
-  思源宋体: "'Noto Serif SC', 'Source Han Serif SC', 'SimSun', serif",
-  楷体: "KaiTi, 'STKaiti', serif",
-  黑体: "'Heiti SC', 'SimHei', 'Microsoft YaHei', sans-serif",
-}
-
-const getFontFamily = (font: string): string => FONT_FAMILY_MAP[font] || FONT_FAMILY_MAP["思源黑体"]
-
 const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
+  variant = "setup",
   coverConfig,
-  titleConfig,
   onCoverConfigChange,
-  resolution,
+  resolution = "720p",
   onResolutionChange,
-  isGenerating,
+  isGenerating = false,
   onGenerate,
   renderJob,
   onGenerateRenderSmartCover,
   onUploadCover,
+  onClose,
+  onCoverSelected,
   summary,
 }) => {
   const uploadInputRef = useRef<HTMLInputElement>(null)
@@ -84,16 +81,31 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     uploadInputRef.current?.click()
   }
 
+  const _applyCoverUrl = (url: string, mode: "upload" | "auto_frame") => {
+    const partial: Partial<AiAvatarCoverConfig> = {
+      mode,
+      thumbnail_url: url,
+    }
+    if (mode === "auto_frame") {
+      partial.smart_cover_url = url
+    } else {
+      partial.upload_url = url
+    }
+    onCoverConfigChange(partial)
+    onCoverSelected?.(url)
+  }
+
   const handleFileChange = (e: React.ChangeEvent<HTMLInputElement>) => {
     const file = e.target.files?.[0]
     if (!file) return
     if (onUploadCover) {
       onUploadCover(file)
-    } else {
-      // 本地预览兜底（实际上传由父级处理；blob URL 仅作本地展示）
-      const url = URL.createObjectURL(file)
-      onCoverConfigChange({ mode: "upload", upload_url: url, thumbnail_url: url })
+      e.target.value = ""
+      return
     }
+    // 本地预览兜底（实际上传由父级处理；blob URL 仅作本地展示）
+    const url = URL.createObjectURL(file)
+    _applyCoverUrl(url, "upload")
     e.target.value = ""
   }
 
@@ -104,11 +116,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     try {
       const res = await onGenerateRenderSmartCover(renderJob.id)
       if (res.cover_url) {
-        onCoverConfigChange({
-          mode: "auto_frame",
-          smart_cover_url: res.cover_url,
-          thumbnail_url: res.cover_url,
-        })
+        _applyCoverUrl(res.cover_url, "auto_frame")
       } else {
         // 失败由父组件 message 提示，这里不重复弹窗
         console.warn("[智能封面] 返回空 cover_url:", res.message)
@@ -120,8 +128,8 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     }
   }
 
-  const lipsync = summary.lipsyncStatus ? LIPSYNC_STATUS_LABEL[summary.lipsyncStatus] : null
-  const canGenerate = summary.lipsyncStatus === "completed" && !isGenerating
+  const lipsync = summary?.lipsyncStatus ? LIPSYNC_STATUS_LABEL[summary.lipsyncStatus] : null
+  const canGenerate = summary?.lipsyncStatus === "completed" && !isGenerating
   // 渲染已完成 → 封面区可用
   const isRenderCompleted = renderJob?.status === "completed"
   const canSmartCover = isRenderCompleted && !smartCoverLoading
@@ -131,63 +139,12 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     coverConfig.smart_cover_url || coverConfig.thumbnail_url || coverConfig.upload_url
   const hasCoverImage = Boolean(coverUrl)
 
-  /** 是否显示标题叠加层：有图、有文字、非加载中 */
-  const showTitleOverlay =
-    hasCoverImage && !smartCoverLoading && titleConfig.title.trim().length > 0
-
-  /** 计算标题叠加层的 inline 样式 */
-  const titleOverlayStyle = useMemo<React.CSSProperties>(() => {
-    const style: React.CSSProperties = {
-      position: "absolute",
-      left: "50%",
-      width: "90%",
-      transform: "translateX(-50%)",
-      textAlign: "center",
-      boxSizing: "border-box",
-      padding: "0 4px",
-      wordBreak: "break-word",
-      whiteSpace: "pre-wrap",
-      color: titleConfig.color || "#ffffff",
-      fontSize: `${(titleConfig.size || 48) * 0.35}px`,
-      fontFamily: getFontFamily(titleConfig.font),
-      fontWeight: titleConfig.bold ? "bold" : "normal",
-      fontStyle: titleConfig.italic ? "italic" : "normal",
-      lineHeight: 1.3,
-      pointerEvents: "none",
-    }
-
-    const pos = titleConfig.position || "bottom"
-    if (pos === "top") {
-      style.top = "40px"
-    } else if (pos === "center") {
-      style.top = "50%"
-      style.transform = "translate(-50%, -50%)"
-    } else if (pos === "custom" && titleConfig.pos_x != null && titleConfig.pos_y != null) {
-      style.left = `${titleConfig.pos_x}%`
-      style.top = `${titleConfig.pos_y}%`
-      style.transform = "translate(-50%, -50%)"
-    } else {
-      style.bottom = "40px"
-    }
-
-    if (titleConfig.stroke) {
-      const strokeWidth = Math.max(1, Math.round(titleConfig.size / 18))
-      ;(style as React.CSSProperties)["WebkitTextStroke"] = `${strokeWidth}px rgba(0,0,0,0.75)`
-      style.textShadow = "none"
-    } else if (titleConfig.shadow) {
-      style.textShadow = "0 2px 8px rgba(0,0,0,0.7), 0 0 2px rgba(0,0,0,0.5)"
-    } else {
-      style.textShadow = "none"
-    }
-
-    return style
-  }, [titleConfig])
-
   /** 封面区占位文字 */
   const coverPlaceholder = isRenderCompleted ? "暂无封面" : "视频生成后可选择封面"
 
-  /** 封面摘要状态文本 */
+  /** 配置摘要中的封面状态标签 */
   const coverSummaryNode = (() => {
+    if (!summary) return null
     if (summary.coverStatus === "selected") {
       return <span className="aa-config-summary__value">已选择</span>
     }
@@ -197,6 +154,69 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
     return <span className="aa-config-summary__empty">生成视频后可选</span>
   })()
 
+  // ── 封面选择区（两种 variant 共用） ─────────────────────────────────
+  const coverSection = (
+    <div className="aa-cover-section" style={{ marginTop: variant === "select-cover" ? 0 : 16 }}>
+      <div className="aa-label" style={{ marginBottom: 8 }}>
+        {variant === "select-cover" ? "选择封面" : "封面"}
+      </div>
+      {/* 封面预览（竖屏 9:16）——成片帧已经通过 Canvas PNG overlay 带有标题，直接展示原图即可 */}
+      <div className="aa-cover-preview" style={{ opacity: isRenderCompleted ? 1 : 0.5 }}>
+        {hasCoverImage ? (
+          <img src={coverUrl!} alt="封面预览" draggable={false} />
+        ) : (
+          <span className="aa-cover-preview__placeholder">{coverPlaceholder}</span>
+        )}
+        {smartCoverLoading && <div className="aa-cover-preview__loading">⏳ 智能选帧中…</div>}
+      </div>
+
+      <div className="aa-cover-actions">
+        <button
+          type="button"
+          className={`aa-btn aa-btn--ghost${coverConfig.mode === "auto_frame" ? " active" : ""}`}
+          onClick={handleSmartCover}
+          disabled={!canSmartCover}
+          title={isRenderCompleted ? "从成片智能选帧" : "请先生成视频"}
+        >
+          {smartCoverLoading ? "⏳ 智能选帧中…" : "🎬 智能获取封面"}
+        </button>
+        <button
+          type="button"
+          className={`aa-btn aa-btn--ghost${coverConfig.mode === "upload" ? " active" : ""}`}
+          onClick={handleUploadClick}
+          disabled={!isRenderCompleted || smartCoverLoading}
+          title={isRenderCompleted ? "自定义上传封面" : "请先生成视频"}
+        >
+          📷 自定义上传
+        </button>
+        <input
+          ref={uploadInputRef}
+          type="file"
+          accept="image/*"
+          style={{ display: "none" }}
+          onChange={handleFileChange}
+        />
+      </div>
+    </div>
+  )
+
+  // ── select-cover 变体：只渲染封面区 + 弹窗确定按钮 ──
+  if (variant === "select-cover") {
+    return (
+      <div className="aa-cover-generate">
+        {coverSection}
+        {onClose && (
+          <div style={{ marginTop: 16, display: "flex", justifyContent: "flex-end" }}>
+            <button type="button" className="aa-btn aa-btn--primary" onClick={onClose}>
+              确定
+            </button>
+          </div>
+        )}
+      </div>
+    )
+  }
+
+  // ── setup 变体：分辨率 / 配置摘要 / 生成按钮（渲染完成后内嵌封面区） ──
   return (
     <div className="aa-cover-generate">
       {/* 分辨率选择 */}
@@ -205,7 +225,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
         <select
           className="aa-select"
           value={resolution}
-          onChange={(e) => onResolutionChange(e.target.value)}
+          onChange={(e) => onResolutionChange?.(e.target.value)}
           disabled={isGenerating}
         >
           {RESOLUTION_OPTIONS.map((opt) => (
@@ -221,7 +241,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
         <div className="aa-config-summary">
           <div className="aa-config-summary__row">
             <span>出镜视频</span>
-            {summary.videoName ? (
+            {summary?.videoName ? (
               <span className="aa-config-summary__value">{summary.videoName}</span>
             ) : (
               <span className="aa-config-summary__empty">未配置</span>
@@ -229,7 +249,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
           </div>
           <div className="aa-config-summary__row">
             <span>音色</span>
-            {summary.voiceName ? (
+            {summary?.voiceName ? (
               <span className="aa-config-summary__value">{summary.voiceName}</span>
             ) : (
               <span className="aa-config-summary__empty">未配置</span>
@@ -237,7 +257,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
           </div>
           <div className="aa-config-summary__row">
             <span>文案</span>
-            {summary.scriptLength > 0 ? (
+            {summary && summary.scriptLength > 0 ? (
               <span className="aa-config-summary__value">{summary.scriptLength} 字</span>
             ) : (
               <span className="aa-config-summary__empty">未配置</span>
@@ -254,12 +274,12 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
           <div className="aa-config-summary__row">
             <span>B-roll 画面插入</span>
             <span className="aa-config-summary__value">
-              {summary.brollCount > 0 ? `${summary.brollCount} 段` : "无"}
+              {summary && summary.brollCount > 0 ? `${summary.brollCount} 段` : "无"}
             </span>
           </div>
           <div className="aa-config-summary__row">
             <span>标题</span>
-            {summary.hasTitle ? (
+            {summary?.hasTitle ? (
               <span className="aa-config-summary__value">已设置</span>
             ) : (
               <span className="aa-config-summary__empty">未配置</span>
@@ -280,7 +300,7 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
         >
           {isGenerating ? "⏳ 生成中..." : "🚀 开始生成视频"}
         </button>
-        {summary.lipsyncStatus !== "completed" && !isGenerating && (
+        {summary?.lipsyncStatus !== "completed" && !isGenerating && (
           <div style={{ marginTop: 8, fontSize: 11, color: "#8c8ca1", textAlign: "center" }}>
             请先完成对口型生成
           </div>
@@ -291,55 +311,6 @@ const PanelCoverAndGenerate: React.FC<PanelCoverAndGenerateProps> = ({
           </div>
         )}
       </div>
-
-      {/* 封面区域（视频生成后才激活；步骤③④要求：按钮在封面上方，完成后再显示封面区） */}
-      <div className="aa-cover-section" style={{ marginTop: 16 }}>
-        <div className="aa-label" style={{ marginBottom: 8 }}>
-          封面
-        </div>
-        {/* 封面预览（竖屏 9:16） */}
-        <div className="aa-cover-preview" style={{ opacity: isRenderCompleted ? 1 : 0.5 }}>
-          {hasCoverImage ? (
-            <img src={coverUrl!} alt="封面预览" draggable={false} />
-          ) : (
-            <span className="aa-cover-preview__placeholder">{coverPlaceholder}</span>
-          )}
-          {smartCoverLoading && <div className="aa-cover-preview__loading">⏳ 智能选帧中…</div>}
-          {showTitleOverlay && (
-            <div style={titleOverlayStyle} aria-hidden="true">
-              {titleConfig.title}
-            </div>
-          )}
-        </div>
-
-        <div className="aa-cover-actions">
-          <button
-            type="button"
-            className={`aa-btn aa-btn--ghost${coverConfig.mode === "auto_frame" ? " active" : ""}`}
-            onClick={handleSmartCover}
-            disabled={!canSmartCover}
-            title={isRenderCompleted ? "从成片智能选帧" : "请先生成视频"}
-          >
-            {smartCoverLoading ? "⏳ 智能选帧中…" : "🎬 智能获取封面"}
-          </button>
-          <button
-            type="button"
-            className={`aa-btn aa-btn--ghost${coverConfig.mode === "upload" ? " active" : ""}`}
-            onClick={handleUploadClick}
-            disabled={!isRenderCompleted || smartCoverLoading}
-            title={isRenderCompleted ? "自定义上传封面" : "请先生成视频"}
-          >
-            📷 自定义上传
-          </button>
-          <input
-            ref={uploadInputRef}
-            type="file"
-            accept="image/*"
-            style={{ display: "none" }}
-            onChange={handleFileChange}
-          />
-        </div>
-      </div>
     </div>
   )
 }

apps/web/src/pages/ai-avatar/components/PanelLipsyncPreview.tsx

@@ -1,9 +1,9 @@
 /**
  * AI数字人 — 对口型预览面板（步骤2用）
  * B-roll 画面插入 + 对口型视频预览 + 生成/重新生成按钮
- * v3.1: 预览容器按 1/2 缩放、标题实时叠加预览
+ * v3.1: 标题字号按预览容器实际宽度动态计算 previewScale（基准 720p），与成片一致
  */
-import React, { useRef } from "react"
+import React, { useCallback, useEffect, useRef, useState } from "react"
 import type { LipsyncJob, BRollSegment, AiAvatarTitleConfig } from "../types"
 
 interface PanelLipsyncPreviewProps {
@@ -29,6 +29,16 @@ function formatTime(seconds: number): string {
   return `${m}:${s.toString().padStart(2, "0")}`
 }
 
+/** 字体名 → CSS font-family 映射（与 titleCanvas 字体链对齐） */
+const FONT_FAMILY_MAP: Record<string, string> = {
+  思源黑体:
+    "'Noto Sans CJK SC', 'Source Han Sans CN', 'PingFang SC', 'Microsoft YaHei', sans-serif",
+  思源宋体: "'Noto Serif SC', 'Source Han Serif SC', 'SimSun', serif",
+  楷体: "KaiTi, 'STKaiti', serif",
+  黑体: "'Heiti SC', 'SimHei', 'Microsoft YaHei', sans-serif",
+}
+const getFontFamily = (font: string): string => FONT_FAMILY_MAP[font] || FONT_FAMILY_MAP["思源黑体"]
+
 export function PanelLipsyncPreview({
   lipsyncJob,
   onGenerateLipsync,
@@ -41,6 +51,8 @@ export function PanelLipsyncPreview({
   const titleDragRef = useRef<HTMLDivElement>(null)
   const draggingTitleRef = useRef(false)
   const previewContainerRef = useRef<HTMLDivElement>(null)
+  // 预览容器实际宽度（通过 ResizeObserver 监听），用于动态计算 previewScale
+  const [containerWidth, setContainerWidth] = useState(0)
   const isGenerating = lipsyncJob?.status === "pending" || lipsyncJob?.status === "processing"
   const isDone = lipsyncJob?.status === "completed"
   const isFailed = lipsyncJob?.status === "failed"
@@ -52,35 +64,84 @@ export function PanelLipsyncPreview({
         ? "排队中…"
         : "对口型生成中…"
 
-  /** 标题叠加样式 */
-  const titleOverlayStyle: React.CSSProperties | null = titleConfig?.title
-    ? {
-        position: "absolute",
-        color: titleConfig.color || "#ffffff",
-        fontFamily: titleConfig.font || "思源黑体",
-        fontSize: `${(titleConfig.size || 48) * 0.35}px`,
-        fontWeight: titleConfig.bold ? 700 : 400,
-        fontStyle: titleConfig.italic ? "italic" : "normal",
-        textAlign: "center",
-        width: "90%",
-        padding: "4px 8px",
-        textShadow: titleConfig.shadow ? "0 2px 4px rgba(0,0,0,0.8)" : undefined,
-        WebkitTextStroke: titleConfig.stroke ? "1.5px #000" : undefined,
-        ...(titleConfig.position === "custom" &&
-        titleConfig.pos_x != null &&
-        titleConfig.pos_y != null
-          ? {
-              left: `${titleConfig.pos_x}%`,
-              top: `${titleConfig.pos_y}%`,
-              transform: "translateX(-50%) translateY(-50%)",
-            }
-          : titleConfig.position === "top"
-            ? { left: "50%", top: 8, transform: "translateX(-50%)" }
-            : titleConfig.position === "bottom"
-              ? { left: "50%", bottom: 8, transform: "translateX(-50%)" }
-              : { left: "50%", top: "50%", transform: "translateX(-50%) translateY(-50%)" }),
-      }
-    : null
+  // 监听预览容器尺寸变化，动态测量宽度以计算 previewScale（基准 720p）
+  useEffect(() => {
+    const el = previewContainerRef.current
+    if (!el) return
+    const update = () => setContainerWidth(el.clientWidth || 0)
+    update()
+    if (typeof ResizeObserver !== "undefined") {
+      const ro = new ResizeObserver(update)
+      ro.observe(el)
+      return () => ro.disconnect()
+    }
+    window.addEventListener("resize", update)
+    return () => window.removeEventListener("resize", update)
+  }, [])
+
+  // 预览缩放比：预览宽度 / 720（基准宽度）
+  const previewScale = containerWidth > 0 ? containerWidth / 720 : 0.35
+  const ps = useCallback((v: number) => Math.round(v * previewScale * 100) / 100, [previewScale])
+
+  /** 标题叠加样式（字号/padding/描边/阴影均按 previewScale 缩放，保持与成片视觉一致） */
+  const titleOverlayStyle: React.CSSProperties | null =
+    titleConfig?.title && containerWidth > 0
+      ? (() => {
+          const baseSize = titleConfig.size || 48
+          const fontSize = ps(baseSize)
+          // 描边宽度基准 ≈ size * 0.06，最小 1.5px @720p
+          const strokeW = Math.max(ps(1.5), +(baseSize * 0.06 * previewScale).toFixed(2))
+          // 阴影按比例缩放
+          const shadowBlur = ps(4)
+          const shadowOffsetY = ps(2)
+          // padding / top 边距按比例（基准 8px 对应预览小窗，成片基准 16px，这里 8px 对应约 0.33 缩放）
+          const padV = ps(16) * 0.5 // ≈ 8px in ~240px container
+          const padH = ps(24) * 0.5
+
+          const style: React.CSSProperties = {
+            position: "absolute",
+            color: titleConfig.color || "#ffffff",
+            fontFamily: getFontFamily(titleConfig.font || "思源黑体"),
+            fontSize: `${fontSize}px`,
+            fontWeight: titleConfig.bold ? 700 : 400,
+            fontStyle: titleConfig.italic ? "italic" : "normal",
+            textAlign: "center",
+            width: "90%",
+            lineHeight: 1.2,
+            padding: `${ps(4)}px ${padH}px`,
+            textShadow: titleConfig.shadow
+              ? `0 ${shadowOffsetY}px ${shadowBlur}px rgba(0,0,0,0.8), 0 0 ${ps(2)}px rgba(0,0,0,0.5)`
+              : undefined,
+            WebkitTextStroke: titleConfig.stroke ? `${strokeW}px #000` : undefined,
+            boxSizing: "border-box",
+            wordBreak: "break-word",
+            whiteSpace: "pre-wrap",
+          }
+
+          if (
+            titleConfig.position === "custom" &&
+            titleConfig.pos_x != null &&
+            titleConfig.pos_y != null
+          ) {
+            style.left = `${titleConfig.pos_x}%`
+            style.top = `${titleConfig.pos_y}%`
+            style.transform = "translateX(-50%) translateY(-50%)"
+          } else if (titleConfig.position === "top") {
+            style.left = "50%"
+            style.top = padV
+            style.transform = "translateX(-50%)"
+          } else if (titleConfig.position === "bottom") {
+            style.left = "50%"
+            style.bottom = padV
+            style.transform = "translateX(-50%)"
+          } else {
+            style.left = "50%"
+            style.top = "50%"
+            style.transform = "translateX(-50%) translateY(-50%)"
+          }
+          return style
+        })()
+      : null
 
   const handleTitlePointerDown = (e: React.PointerEvent<HTMLDivElement>) => {
     if (!onTitlePositionChange || !previewContainerRef.current) return
@@ -183,7 +244,7 @@ export function PanelLipsyncPreview({
         )}
       </div>
 
-      {/* ── 对口型预览（v3.1: 缩放1/2 + 标题叠加） ─ */}
+      {/* ── 对口型预览（标题字号按 previewScale 动态缩放） ─ */}
       <div className="aa-lipsync-section">
         <div className="aa-lipsync-section__title">对口型预览</div>
 

apps/web/src/pages/ai-avatar/hooks/useAiAvatar.ts

@@ -1,5 +1,5 @@
 /**
- * AI数字人 — 页面全局状态管理 hook（v3）
+ * AI数字人 — 页面全局状态管理 hook（v3 + #1845 配音前置）
  */
 import { useState, useCallback } from "react"
 import type { AssetItem } from "@/api/assets"
@@ -13,10 +13,19 @@ import {
   type BRollSegment,
   type AiAvatarTitleConfig,
   type AiAvatarCoverConfig,
+  type TtsPreviewResult,
   DEFAULT_TITLE_CONFIG,
   DEFAULT_COVER_CONFIG,
 } from "../types"
 
+const DEFAULT_TTS_PREVIEW: TtsPreviewResult = {
+  audioUrl: null,
+  duration: 0,
+  sentenceTimings: [],
+  status: "idle",
+  error: null,
+}
+
 export function useAiAvatar() {
   /* ── 面板1：出镜视频 ── */
   const [selectedVideo, setSelectedVideo] = useState<AssetItem | null>(null)
@@ -36,6 +45,9 @@ export function useAiAvatar() {
   const [showScriptModal, setShowScriptModal] = useState(false)
   const [showBRollModal, setShowBRollModal] = useState(false)
 
+  /* ── #1845 TTS 预合成（步骤1「生成配音」） ── */
+  const [ttsPreview, setTtsPreview] = useState<TtsPreviewResult>(DEFAULT_TTS_PREVIEW)
+
   /* ── 面板3.5：B-roll ── */
   const [bRollSegments, setBRollSegments] = useState<BRollSegment[]>([])
 
@@ -81,6 +93,7 @@ export function useAiAvatar() {
     setScript(null)
     setScriptText("")
     setLipsyncJob(null)
+    setTtsPreview(DEFAULT_TTS_PREVIEW)
     setBRollSegments([])
     setTitleConfig(DEFAULT_TITLE_CONFIG)
     setCoverConfig(DEFAULT_COVER_CONFIG)
@@ -118,6 +131,10 @@ export function useAiAvatar() {
     showBRollModal,
     setShowBRollModal,
     selectScript,
+    // #1845 TTS 预合成
+    ttsPreview,
+    setTtsPreview,
+    resetTtsPreview: useCallback(() => setTtsPreview(DEFAULT_TTS_PREVIEW), []),
     // B-roll
     bRollSegments,
     addBRollSegment,

apps/web/src/pages/ai-avatar/types.ts

@@ -28,6 +28,17 @@ export const VOICE_LANGUAGE_OPTIONS: { value: VoiceLanguage; label: string }[] =
 /* ── 对口型任务状态 ── */
 export type LipsyncStatus = "idle" | "pending" | "processing" | "completed" | "failed"
 
+/* ── TTS 预合成（#1845 配音前置：步骤1「生成配音」状态） ── */
+export type TtsPreviewStatus = "idle" | "generating" | "done" | "failed"
+
+export interface TtsPreviewResult {
+  audioUrl: string | null
+  duration: number
+  sentenceTimings: SentenceTiming[]
+  status: TtsPreviewStatus
+  error: string | null
+}
+
 /* ── 文案 ── */
 export interface Script {
   id: string

apps/web/src/pages/ai-avatar/utils/titleCanvas.ts

@@ -4,6 +4,10 @@
  * 把标题按前端预览的 HTML/CSS 效果画到透明背景 PNG 上（与视频同分辨率），
  * 以 dataURL 形式传给后端，后端用 FFmpeg overlay 直接叠加图层，
  * 彻底解决前端 HTML/CSS 预览 ≠ FFmpeg drawtext 成片的 WYSIWYG 问题。
+ *
+ * 约定：titleConfig.size 的语义是"720p 基准宽度下的字号（px）"，
+ *       按 videoWidth / 720 得到 scale，所有长度类参数乘以 scale，
+ *       保证 1080p / 4K 成片里标题视觉大小与预览一致。
  */
 import type { AiAvatarTitleConfig } from "../types"
 
@@ -35,13 +39,18 @@ export function renderTitleToPngDataUrl(opts: RenderTitlePngOptions): string | n
     .filter((l) => l.length > 0)
   if (lines.length === 0) return null
 
+  // 分辨率缩放系数：基准 720p，所有长度类参数乘以 scale
+  const scale = videoWidth / 720
+  const r = (v: number) => Math.round(v * scale)
+
   const canvas = document.createElement("canvas")
   canvas.width = videoWidth
   canvas.height = videoHeight
   const ctx = canvas.getContext("2d")
   if (!ctx) return null
 
-  const size = Math.max(12, Math.round(titleConfig.size || 48))
+  const baseSize = Math.max(12, Math.round(titleConfig.size || 48))
+  const size = r(baseSize)
   const bold = !!titleConfig.bold
   const italic = !!titleConfig.italic
   const color = titleConfig.color || "#ffffff"
@@ -60,19 +69,16 @@ export function renderTitleToPngDataUrl(opts: RenderTitlePngOptions): string | n
   ctx.textAlign = "center"
   ctx.textBaseline = "middle"
 
-  // 阴影（shadow=true 时开启）
+  // 阴影（shadow=true 时开启）——按 scale 缩放
   if (shadow) {
     ctx.shadowColor = "rgba(0,0,0,0.8)"
-    ctx.shadowBlur = 4
+    ctx.shadowBlur = r(4)
     ctx.shadowOffsetX = 0
-    ctx.shadowOffsetY = 2
+    ctx.shadowOffsetY = r(2)
   }
 
-  // 位置计算：与 PanelLipsyncPreview 的 CSS 对齐
-  // 预览用 top/bottom 8px padding + transform translateX(-50%) 居中；
-  // 这里画到整尺寸 canvas，padding 按比例放大到全分辨率（预览缩放 0.35x 时 8px ≈ 23px 全尺寸，
-  // 为更贴近原 CSS 16px 安全边距，用 16px 作为内边距）。
-  const PAD = 16
+  // 位置计算：与 PanelLipsyncPreview 的 CSS 对齐（按 scale 缩放 PAD）
+  const PAD = r(16)
   let centerX = videoWidth / 2
   const position = titleConfig.position || "bottom"
   const lineGap = size * 1.2
@@ -97,10 +103,9 @@ export function renderTitleToPngDataUrl(opts: RenderTitlePngOptions): string | n
     firstLineY = videoHeight - totalTextH - PAD + size / 2
   }
 
-  // 描边参数（stroke=true 或 bold 默认细描边模拟粗体时都画；
-  // 注意：浏览器原生 bold 已经是粗体 glyph，Canvas 这里对 stroke=true 才加黑描边，
-  // 与预览 CSS 的 WebkitTextStroke 保持一致，不对 bold 自动加描边避免双粗）。
+  // 描边参数：描边 lineWidth 按 scale 缩放（基准 size * 0.06，最小 2px @720p）
   const doStroke = stroke
+  const strokeWidth = Math.max(r(2), Math.round(size * 0.06))
   // 逐行绘制
   lines.forEach((line, idx) => {
     const y = firstLineY + idx * lineGap
@@ -110,14 +115,14 @@ export function renderTitleToPngDataUrl(opts: RenderTitlePngOptions): string | n
       // 描边不要带阴影（避免黑色描边发虚）
       ctx.shadowColor = "rgba(0,0,0,0)"
       ctx.shadowBlur = 0
-      ctx.lineWidth = Math.max(2, size * 0.06)
+      ctx.lineWidth = strokeWidth
       ctx.strokeStyle = "#000000"
       ctx.lineJoin = "round"
       ctx.strokeText(line, centerX, y)
       // 恢复阴影
       if (shadow) {
         ctx.shadowColor = "rgba(0,0,0,0.8)"
-        ctx.shadowBlur = 4
+        ctx.shadowBlur = r(4)
       } else {
         ctx.shadowColor = prevShadowColor
         ctx.shadowBlur = prevShadowBlur

apps/worker/video_processing/render_adapter.py

@@ -221,17 +221,18 @@ class RenderAdapter:
 
         except subprocess.CalledProcessError as exc:
             stderr_text = (exc.stderr or "").strip()
+            stderr_tail = stderr_text[-5000:] if len(stderr_text) > 5000 else stderr_text
             logger.error(
                 "[render-adapter] ffmpeg渲染失败: plan_id=%s job_id=%s exit_code=%d\nstderr:\n%s",
                 plan_id,
                 job_id,
                 exc.returncode,
-                stderr_text[-2000:] if len(stderr_text) > 2000 else stderr_text,
+                stderr_tail,
             )
             return RenderAdapterResult(
                 success=False,
-                error_message=f"FFmpeg渲染失败(exit={exc.returncode}): {stderr_text[:200]}",
-                error_detail=stderr_text[-2000:] if len(stderr_text) > 2000 else stderr_text,
+                error_message=f"FFmpeg渲染失败(exit={exc.returncode}): {stderr_text[-500:]}",
+                error_detail=stderr_tail,
             )
         except Exception as exc:
             logger.exception(
@@ -721,16 +722,17 @@ class RenderAdapter:
 
         except subprocess.CalledProcessError as exc:
             stderr_text = (exc.stderr or "").strip()
+            stderr_tail = stderr_text[-5000:] if len(stderr_text) > 5000 else stderr_text
             logger.error(
                 "[render-adapter] 内存模式渲染失败: plan_id=%s exit_code=%d\nstderr:\n%s",
                 actual_plan_id,
                 exc.returncode,
-                stderr_text[-2000:] if len(stderr_text) > 2000 else stderr_text,
+                stderr_tail,
             )
             return RenderAdapterResult(
                 success=False,
-                error_message=f"FFmpeg渲染失败(exit={exc.returncode}): {stderr_text[:200]}",
-                error_detail=stderr_text[-2000:] if len(stderr_text) > 2000 else stderr_text,
+                error_message=f"FFmpeg渲染失败(exit={exc.returncode}): {stderr_text[-500:]}",
+                error_detail=stderr_tail,
             )
         except Exception as exc:
             logger.exception(

apps/worker/video_processing/unified_render_service.py

@@ -2301,13 +2301,13 @@ class UnifiedRenderService:
                     filters.append(f"eq=contrast={contrast:.3f}")
 
             elif filt == "color_balance":
-                # RGB 通道偏移：color_balance=rs=...:gs=...:bs=...
+                # RGB 通道偏移：colorbalance=rs=...:gs=...:bs=...
                 r = pixel_pert.get("color_r", 0)
                 g = pixel_pert.get("color_g", 0)
                 b = pixel_pert.get("color_b", 0)
                 if r != 0 or g != 0 or b != 0:
                     # color_balance 参数范围 -1.0 ~ 1.0，这里用 /100 转换
-                    filters.append(f"color_balance=rs={r/100:.3f}:gs={g/100:.3f}:bs={b/100:.3f}")
+                    filters.append(f"colorbalance=rs={r/100:.3f}:gs={g/100:.3f}:bs={b/100:.3f}")
 
     @staticmethod
     def _clip_volume(clip: ResolvedClip) -> float:

apps/worker/worker_app/tasks/generation.py

@@ -705,7 +705,8 @@ def _render_from_edit_plan(
         )
 
         if not result.success:
-            raise RuntimeError(f"渲染失败: {result.error_message}")
+            detail_suffix = f"\n[detail] {result.error_detail}" if result.error_detail else ""
+            raise RuntimeError(f"渲染失败: {result.error_message}{detail_suffix}")
 
         render_elapsed = time.monotonic() - render_start
         logger.info(

packages/domain/sentence_timings.py

@@ -0,0 +1,205 @@
+"""共享的句子时间戳计算工具 — 供 Celery TTS 任务和 /lipsync/tts-preview 同步接口复用.
+
+- `_split_script_into_sentences`: 按标点分句（中英文逗号/句号/问号/感叹号/分号/换行）
+- `_estimate_sentence_timings_by_chars`: 按字数比例估算（静音检测失败时降级）
+- `_probe_audio_duration`: ffprobe 读取音频时长
+- `compute_sentence_timings`: 基于 ffmpeg silencedetect 精确计算每句起止时间
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import re
+import subprocess
+import tempfile
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+def split_script_into_sentences(script_text: str) -> list[str]:
+    """按句号/问号/感叹号/分号/逗号/换行分句（与前端 SENTENCE_SPLIT_RE 一致）.
+
+    中文短视频文案习惯用「，」断小句（如"卖花的叫花无缺，卖姜的叫姜子牙"），
+    必须把逗号也纳入分隔符，否则多句文案会被识别成一整句，导致 B-roll 时间戳错位。
+    """
+    text = (script_text or "").strip()
+    if not text:
+        return []
+    parts = re.split(r"[。！？!?？！;；,，\n\r]+", text)
+    return [p.strip() for p in parts if p.strip()]
+
+
+def estimate_sentence_timings_by_chars(sentences: list[str], total_duration: float) -> list[dict]:
+    """降级方案：按字数比例估算句子时间（与原前端逻辑一致）."""
+    if not sentences or total_duration <= 0:
+        return []
+    total_chars = sum(len(s.replace(r"\s", "")) for s in sentences)
+    if total_chars == 0:
+        return []
+
+    timings = []
+    acc = 0
+    for i, sent in enumerate(sentences):
+        chars = len(sent.replace(r"\s", ""))
+        start = (acc / total_chars) * total_duration
+        end = ((acc + chars) / total_chars) * total_duration
+        timings.append(
+            {
+                "index": i,
+                "text": sent,
+                "start_time": round(start, 2),
+                "end_time": round(end, 2),
+            }
+        )
+        acc += chars
+    return timings
+
+
+def probe_audio_duration(audio_data: bytes, timeout: int = 10) -> float:
+    """用 ffprobe 读取音频字节流的时长（秒）.
+
+    Returns:
+        时长（秒），失败返回 0.0
+    """
+    if not audio_data:
+        return 0.0
+    tmp_path: Optional[str] = None
+    try:
+        with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
+            tmp.write(audio_data)
+            tmp_path = tmp.name
+        result = subprocess.run(
+            [
+                "ffprobe",
+                "-v",
+                "error",
+                "-show_entries",
+                "format=duration",
+                "-of",
+                "default=noprint_wrappers=1:nokey=1",
+                tmp_path,
+            ],
+            capture_output=True,
+            text=True,
+            timeout=timeout,
+        )
+        stdout = (result.stdout or "").strip()
+        if not stdout:
+            logger.warning("[sentence_timings] ffprobe 无输出: stderr=%s", (result.stderr or "")[:200])
+            return 0.0
+        return float(stdout)
+    except Exception as exc:
+        logger.warning("[sentence_timings] ffprobe 时长探测失败: %s", exc)
+        return 0.0
+    finally:
+        if tmp_path:
+            try:
+                os.unlink(tmp_path)
+            except Exception:
+                pass
+
+
+def compute_sentence_timings(audio_data: bytes, script_text: str, total_duration: float) -> list[dict]:
+    """基于 TTS 音频的静音检测，精确计算每句文案的起止时间.
+
+    使用 ffmpeg silencedetect 检测静音段，将静音点与句子边界对齐。
+    比字数比例估算准确得多。
+
+    Args:
+        audio_data: TTS 音频二进制数据（MP3）
+        script_text: 文案全文
+        total_duration: 音频总时长（秒）
+
+    Returns:
+        list[{"index": int, "text": str, "start_time": float, "end_time": float}]
+    """
+    sentences = split_script_into_sentences(script_text)
+    if not sentences:
+        return []
+
+    tmp_path: Optional[str] = None
+    try:
+        with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as tmp:
+            tmp.write(audio_data)
+            tmp_path = tmp.name
+
+        result = subprocess.run(
+            [
+                "ffmpeg",
+                "-i",
+                tmp_path,
+                "-af",
+                "silencedetect=noise=-25dB:d=0.3",
+                "-f",
+                "null",
+                "-",
+            ],
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+        stderr = result.stderr or ""
+
+        silence_ends = []
+        for match in re.finditer(r"silence_end:\s*([\d.]+)", stderr):
+            t = float(match.group(1))
+            if 0 < t < total_duration:
+                silence_ends.append(t)
+
+        if len(silence_ends) < len(sentences) - 1:
+            logger.warning(
+                "[sentence_timings] 静音点不足(%d < %d)，降级为字数比例估算",
+                len(silence_ends),
+                len(sentences) - 1,
+            )
+            return estimate_sentence_timings_by_chars(sentences, total_duration)
+
+        n_boundaries = len(sentences) - 1
+        boundaries = []
+        used_indices = set()
+
+        for i in range(n_boundaries):
+            expected_pos = (i + 1) / len(sentences) * total_duration
+            best_idx = None
+            best_dist = float("inf")
+            for j, t in enumerate(silence_ends):
+                if j in used_indices:
+                    continue
+                dist = abs(t - expected_pos)
+                if dist < best_dist:
+                    best_dist = dist
+                    best_idx = j
+            if best_idx is not None:
+                used_indices.add(best_idx)
+                boundaries.append(silence_ends[best_idx])
+
+        boundaries.sort()
+
+        timings = []
+        prev_end = 0.0
+        for i, sent in enumerate(sentences):
+            start = prev_end
+            end = boundaries[i] if i < len(boundaries) else total_duration
+            timings.append(
+                {
+                    "index": i,
+                    "text": sent,
+                    "start_time": round(start, 2),
+                    "end_time": round(end, 2),
+                }
+            )
+            prev_end = end
+
+        return timings
+
+    except Exception as exc:
+        logger.warning("[sentence_timings] 静音检测异常，降级为字数比例估算: %s", exc)
+        return estimate_sentence_timings_by_chars(sentences, total_duration)
+    finally:
+        if tmp_path:
+            try:
+                os.unlink(tmp_path)
+            except Exception:
+                pass

tests/unit/test_ai_avatar_emotion_tts_lipsync.py

@@ -262,8 +262,8 @@ def test_smart_cover_selects_best_frame_and_persists():
     score_patch.assert_called_once()
     # 验证使用了增大的轮询参数
     call_kwargs = mk.extract_frames.call_args
-    assert call_kwargs.kwargs.get("poll_interval") == 1.0 or call_kwargs[1].get("poll_interval") == 1.0
-    assert call_kwargs.kwargs.get("max_poll_attempts") == 15 or call_kwargs[1].get("max_poll_attempts") == 15
+    assert call_kwargs.kwargs.get("poll_interval") == 2.0 or call_kwargs[1].get("poll_interval") == 2.0
+    assert call_kwargs.kwargs.get("max_poll_attempts") == 30 or call_kwargs[1].get("max_poll_attempts") == 30
 
 
 def test_smart_cover_returns_empty_when_mediakit_unavailable():
@@ -334,8 +334,8 @@ def test_extract_frames_uses_extended_poll_params():
         cov.select_best_cover_frame("https://other/avatar.mp4", max_frames=3)
 
     call_kwargs = mk.extract_frames.call_args
-    assert call_kwargs.kwargs.get("poll_interval") == 1.0 or call_kwargs[1].get("poll_interval") == 1.0
-    assert call_kwargs.kwargs.get("max_poll_attempts") == 15 or call_kwargs[1].get("max_poll_attempts") == 15
+    assert call_kwargs.kwargs.get("poll_interval") == 2.0 or call_kwargs[1].get("poll_interval") == 2.0
+    assert call_kwargs.kwargs.get("max_poll_attempts") == 30 or call_kwargs[1].get("max_poll_attempts") == 30
     assert call_kwargs.kwargs.get("max_retries") == 1 or call_kwargs[1].get("max_retries") == 1
 
 

tests/unit/test_ai_avatar_render_service.py

@@ -519,8 +519,8 @@ class TestAiAvatarRenderService:
         # 不应执行渲染逻辑
         mock_db.commit.assert_not_called()
 
-    def test_execute_render_success_creates_clip_record(self):
-        """execute_render 完成后自动创建成片记录到成片库."""
+    def test_execute_render_completed_does_not_auto_persist(self):
+        """execute_render 完成后不自动入库成片库（改为用户点「完成」时由 finalize_job 入库）."""
         from app.services.ai_avatar_render_service import AiAvatarRenderService
 
         mock_db = _make_mock_db()
@@ -549,36 +549,139 @@ class TestAiAvatarRenderService:
             patch.object(svc, "_upload_to_oss", side_effect=lambda path, key: f"https://oss/{key}"),
             patch("subprocess.run") as mock_run,
             patch("tempfile.TemporaryDirectory") as tmpdir_mock,
-            patch(
-                "app.services.ai_avatar_cover_service.generate_smart_cover", return_value="https://oss/smart_cover.jpg"
-            ),
+            patch("packages.domain.generated_video.GeneratedVideo.create") as gv_create,
+        ):
+            import subprocess as _sp
+
+            mock_run.return_value = _sp.CompletedProcess(args=[], returncode=0, stdout="", stderr="")
+            tmpdir_mock.return_value.__enter__ = MagicMock(return_value="/tmp/testdir")
+            tmpdir_mock.return_value.__exit__ = MagicMock(return_value=False)
+
+            svc.execute_render("render-ok")
+
+        # 状态应为 completed，但没有自动入库
+        assert mock_job.status == "completed"
+        gv_create.assert_not_called()
+        assert mock_job.output_video_url.startswith("https://oss/")
+
+    def test_finalize_job_persists_to_library(self):
+        """finalize_job 在用户点「完成」后写入成片库，thumbnail_url 使用 job.output_cover_url."""
+        from app.services.ai_avatar_render_service import AiAvatarRenderService
+
+        mock_db = MagicMock()
+        mock_job = _make_mock_render_job(
+            job_id="render-finalize",
+            status="completed",
+            output_video_url="https://oss/ai-avatar/render-finalize/output.mp4",
+            output_cover_url="https://oss/cover.jpg",
+            output_duration=12.0,
+        )
+
+        # get_render_job → db.query(AiAvatarRenderJob).filter().first() 返回 mock_job
+        # finalize 幂等检查 → db.query(GeneratedVideoModel).filter().first() 返回 None（未入库）
+        def _query_side_effect(model):
+            q = MagicMock()
+            if model.__name__ == "AiAvatarRenderJob":
+                q.filter.return_value.first.return_value = mock_job
+            else:
+                # GeneratedVideoModel
+                q.filter.return_value.first.return_value = None
+            return q
+
+        mock_db.query.side_effect = _query_side_effect
+
+        svc = AiAvatarRenderService(mock_db)
+
+        with (
             patch("packages.domain.generated_video.GeneratedVideo.create") as gv_create,
             patch(
                 "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository"
             ) as repo_cls,
         ):
-            import subprocess as _sp
-
-            mock_run.return_value = _sp.CompletedProcess(args=[], returncode=0, stdout="", stderr="")
-            import tempfile as _tf
-
-            tmpdir_mock.return_value.__enter__ = MagicMock(return_value="/tmp/testdir")
-            tmpdir_mock.return_value.__exit__ = MagicMock(return_value=False)
-
             mock_clip = MagicMock()
-            mock_clip.id = "clip-001"
+            mock_clip.id = "clip-new"
+            mock_clip.thumbnail_url = "https://oss/cover.jpg"
             gv_create.return_value = mock_clip
             mock_repo = MagicMock()
+            mock_repo.create.return_value = mock_clip
+            mock_repo.get.return_value = mock_clip
             repo_cls.return_value = mock_repo
 
-            svc.execute_render("render-ok")
+            video = svc.finalize_job("render-finalize", "user-1")
+            assert video.id == "clip-new"
+            gv_create.assert_called_once()
+            call_kwargs = gv_create.call_args.kwargs
+            assert call_kwargs["file_url"].endswith("output.mp4")
+            assert call_kwargs["thumbnail_url"] == "https://oss/cover.jpg"
+            assert call_kwargs["generation_task_id"] == "render-finalize"
+            mock_repo.create.assert_called_once()
 
-        assert mock_job.status == "completed"
-        gv_create.assert_called_once()
-        call_kwargs = gv_create.call_args
-        assert "https://oss/" in call_kwargs.kwargs["file_url"]
-        assert call_kwargs.kwargs["user_id"] == "user-1"
-        mock_repo.create.assert_called_once_with(mock_clip)
+    def test_finalize_job_idempotent_when_already_persisted(self):
+        """finalize_job 重复调用：幂等检查命中后直接返回已有记录，不再 create."""
+        from app.services.ai_avatar_render_service import AiAvatarRenderService
+
+        mock_db = MagicMock()
+        mock_job = _make_mock_render_job(
+            job_id="render-finalize-2",
+            status="completed",
+            output_video_url="https://oss/output.mp4",
+            output_cover_url="https://oss/cover.jpg",
+            output_duration=12.0,
+        )
+        existing_model = MagicMock()
+        existing_model.id = "clip-existing"
+        existing_model.thumbnail_url = "https://oss/cover.jpg"
+
+        def _query_side_effect(model):
+            q = MagicMock()
+            if model.__name__ == "AiAvatarRenderJob":
+                q.filter.return_value.first.return_value = mock_job
+            else:
+                q.filter.return_value.first.return_value = existing_model
+            return q
+
+        mock_db.query.side_effect = _query_side_effect
+
+        svc = AiAvatarRenderService(mock_db)
+
+        with (
+            patch("packages.domain.generated_video.GeneratedVideo.create") as gv_create,
+            patch(
+                "packages.adapters.sqlalchemy_impl.generated_video_repository.SQLAlchemyGeneratedVideoRepository"
+            ) as repo_cls,
+        ):
+            mock_existing = MagicMock()
+            mock_existing.id = "clip-existing"
+            mock_repo = MagicMock()
+            mock_repo.get.return_value = mock_existing
+            repo_cls.return_value = mock_repo
+
+            video = svc.finalize_job("render-finalize-2", "user-1")
+            assert video.id == "clip-existing"
+            gv_create.assert_not_called()
+            mock_repo.create.assert_not_called()
+
+    def test_finalize_job_requires_completed_status(self):
+        """finalize_job 在非 completed 状态下抛异常."""
+        from app.services.ai_avatar_render_service import AiAvatarRenderError, AiAvatarRenderService
+
+        mock_db = MagicMock()
+        mock_job = _make_mock_render_job(
+            job_id="render-pending",
+            status="processing",
+            output_video_url="",
+            output_cover_url="",
+            output_duration=0.0,
+        )
+        filter_mock = MagicMock()
+        filter_mock.first.return_value = mock_job
+        query_mock = MagicMock()
+        query_mock.filter.return_value = filter_mock
+        mock_db.query.return_value = query_mock
+
+        svc = AiAvatarRenderService(mock_db)
+        with pytest.raises(AiAvatarRenderError):
+            svc.finalize_job("render-pending", "user-1")
 
     def test_execute_render_clip_failure_does_not_affect_render(self):
         """成片创建失败不影响渲染任务标记为成功."""
@@ -610,7 +713,6 @@ class TestAiAvatarRenderService:
             patch.object(svc, "_upload_to_oss", side_effect=lambda path, key: f"https://oss/{key}"),
             patch("subprocess.run") as mock_run,
             patch("tempfile.TemporaryDirectory") as tmpdir_mock,
-            patch("app.services.ai_avatar_cover_service.generate_smart_cover", side_effect=RuntimeError("DB error")),
         ):
             import subprocess as _sp
 
@@ -619,7 +721,7 @@ class TestAiAvatarRenderService:
             tmpdir_mock.return_value.__exit__ = MagicMock(return_value=False)
             svc.execute_render("render-clip-fail")
 
-        # 即使成片创建失败，渲染任务仍应标记为 completed
+        # 渲染任务仍应标记为 completed（不入库不影响渲染成功）
         assert mock_job.status == "completed"
 
     def test_error_exception_has_code(self):

tests/unit/test_lipsync_routes.py

@@ -91,16 +91,27 @@ class TestSchemaValidation:
         assert req.voice_id == "longxiaochun_v3"
         assert req.script_text == "大家好，欢迎来到直播间"
 
-    def test_invalid_video_url_not_mp4(self):
+    def test_invalid_video_url_unsupported_ext(self):
         from app.schemas.lipsync import CreateLipsyncJobRequest
 
-        with pytest.raises(ValueError, match="MP4"):
+        # 不支持的扩展名（.txt）应报错
+        with pytest.raises(ValueError, match="格式不支持"):
             CreateLipsyncJobRequest(
-                video_url="https://example.com/video.mov",
+                video_url="https://example.com/video.txt",
                 voice_id="longxiaochun_v3",
                 script_text="测试文本",
             )
 
+    def test_mov_video_url_accepted(self):
+        from app.schemas.lipsync import CreateLipsyncJobRequest
+
+        # .MOV 是 iPhone 拍摄的常见容器，h264 编码可直接被 MediaKit 处理
+        req = CreateLipsyncJobRequest(
+            video_url="https://example.com/video.mov",
+            audio_url="https://example.com/audio.mp3",
+        )
+        assert req.video_url.endswith(".mov")
+
     def test_invalid_video_url_empty(self):
         from app.schemas.lipsync import CreateLipsyncJobRequest
 
@@ -159,7 +170,8 @@ class TestSchemaValidation:
             voice_id="longxiaochun_v3",
             script_text="测试文本",
         )
-        assert req.enable_video_loop is False
+        # AI数字人场景文案长度不可控，默认开启视频循环，防止音频长于视频时被截断
+        assert req.enable_video_loop is True
 
     def test_video_url_strip_query_params(self):
         """视频 URL 含查询参数时，扩展名检查应忽略 ? 后面的部分."""

tests/unit/test_lipsync_tts_preview.py

@@ -0,0 +1,169 @@
+"""AI 数字人 对口型 TTS 预合成接口（#1845）单元测试 — 覆盖 LipsyncService.preview_tts 成功/失败路径.
+
+直接调用 LipsyncService.preview_tts()，mock CosyVoiceService / safe_download_bytes / ffprobe，
+验证返回结构、错误码、与共享 sentence_timings 工具的协作。
+"""
+
+import os
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+os.environ.setdefault("JWT_SECRET_KEY", "dev-secret-key-for-testing")
+
+
+def _make_service(
+    *,
+    cosyvoice=None,
+    download_bytes=b"FAKE_MP3_DATA",
+    download_error=None,
+    ffprobe_duration=5.0,
+    timings_result=None,
+):
+    """构造 LipsyncService 并把 CosyVoiceService/safe_download_bytes/probe/compute 全部 mock 掉。"""
+    from app.services.lipsync_service import LipsyncService
+
+    db = MagicMock()
+
+    # 构造唯一的 cosyvoice mock 实例，便于断言
+    _cosy_inst = MagicMock()
+    if cosyvoice is None:
+        _cosy_inst.submit_synthesize_task.return_value = {"audio_url": "https://cosy.example.com/tts.mp3"}
+    elif isinstance(cosyvoice, Exception):
+        _cosy_inst.submit_synthesize_task.side_effect = cosyvoice
+    else:
+        _cosy_inst.submit_synthesize_task.return_value = cosyvoice
+
+    def _fake_get_cosyvoice(self):  # noqa: ARG001
+        return _cosy_inst
+
+    def _fake_resolve_voice_id(self, voice_id, user_id):  # noqa: ARG001
+        return voice_id
+
+    svc = LipsyncService(db=db, client=MagicMock(), voice_clone_repo=MagicMock())
+    svc._cosyvoice = _cosy_inst
+    patch.object(LipsyncService, "_get_cosyvoice", _fake_get_cosyvoice).start()
+    patch.object(LipsyncService, "_resolve_voice_id", _fake_resolve_voice_id).start()
+
+    # mock safe_download_bytes
+    if download_error is not None:
+        patch(
+            "app.services.lipsync_service.safe_download_bytes",
+            side_effect=download_error,
+        ).start()
+    else:
+        patch(
+            "app.services.lipsync_service.safe_download_bytes",
+            return_value=download_bytes,
+        ).start()
+
+    # mock probe_audio_duration（patch 到 lipsync_service 模块的命名空间）
+    patch(
+        "app.services.lipsync_service.probe_audio_duration",
+        return_value=ffprobe_duration,
+    ).start()
+
+    # mock compute_sentence_timings
+    default_timings = [
+        {"index": 0, "text": "你好", "start_time": 0.0, "end_time": 1.5},
+        {"index": 1, "text": "世界", "start_time": 1.5, "end_time": 5.0},
+    ]
+    patch(
+        "app.services.lipsync_service.compute_sentence_timings",
+        return_value=timings_result if timings_result is not None else default_timings,
+    ).start()
+
+    svc.__dict__["_test_cosy"] = _cosy_inst
+    return svc
+
+
+def test_preview_tts_success():
+    """正常路径：TTS 合成成功 → 下载 → ffprobe → 计算 timings，返回完整结构。"""
+    svc = _make_service(ffprobe_duration=5.0)
+    try:
+        result = svc.preview_tts(
+            user_id="user-1",
+            voice_id="longxiaochun",
+            script_text="你好，世界",
+            speed=1.0,
+            emotion="natural",
+        )
+        assert result["audio_url"] == "https://cosy.example.com/tts.mp3"
+        assert result["duration"] == 5.0
+        assert isinstance(result["sentence_timings"], list)
+        assert len(result["sentence_timings"]) == 2
+        assert result["sentence_timings"][0]["text"] == "你好"
+        cosy = svc.__dict__["_test_cosy"]
+        cosy.submit_synthesize_task.assert_called_once()
+        kwargs = cosy.submit_synthesize_task.call_args.kwargs
+        assert kwargs["text"] == "你好，世界"
+        assert kwargs["voice_id"] == "longxiaochun"
+    finally:
+        patch.stopall()
+
+
+def test_preview_tts_cosyvoice_error():
+    """CosyVoice 抛错：应该包装成 MediaKitError 抛出。"""
+    from app.services.mediakit_client import MediaKitError
+
+    from packages.application.cosyvoice_service import CosyVoiceError
+
+    svc = _make_service(cosyvoice=CosyVoiceError("cosyvoice down"))
+    try:
+        with pytest.raises(MediaKitError):
+            svc.preview_tts(
+                user_id="user-1",
+                voice_id="longxiaochun",
+                script_text="你好",
+            )
+    finally:
+        patch.stopall()
+
+
+def test_preview_tts_download_fail_still_returns_url():
+    """音频下载失败：不抛错，返回 audio_url + 空 timings，前端仍能继续（降级）。"""
+    svc = _make_service(download_error=RuntimeError("network down"))
+    try:
+        result = svc.preview_tts(
+            user_id="user-1",
+            voice_id="longxiaochun",
+            script_text="你好，世界",
+        )
+        assert result["audio_url"] == "https://cosy.example.com/tts.mp3"
+        assert result["duration"] == 0.0
+        assert result["sentence_timings"] == []
+    finally:
+        patch.stopall()
+
+
+def test_preview_tts_ffprobe_zero_duration():
+    """ffprobe 返回 0：timings 为空，不抛错。"""
+    svc = _make_service(ffprobe_duration=0.0)
+    try:
+        result = svc.preview_tts(
+            user_id="user-1",
+            voice_id="longxiaochun",
+            script_text="你好",
+        )
+        assert result["audio_url"]
+        assert result["duration"] == 0.0
+        assert result["sentence_timings"] == []
+    finally:
+        patch.stopall()
+
+
+def test_preview_tts_no_audio_url_in_response():
+    """CosyVoice 返回无 audio_url：抛 MediaKitError TTSNoAudio。"""
+    from app.services.mediakit_client import MediaKitError
+
+    svc = _make_service(cosyvoice={"audio_url": ""})
+    try:
+        with pytest.raises(MediaKitError) as exc_info:
+            svc.preview_tts(
+                user_id="user-1",
+                voice_id="longxiaochun",
+                script_text="你好",
+            )
+        assert exc_info.value.code == "TTSNoAudio"
+    finally:
+        patch.stopall()

tests/unit/test_sentence_timings.py

@@ -1,4 +1,4 @@
-"""Tests for sentence timing functions in lipsync_tts."""
+"""Tests for sentence timing functions (now in packages/domain/sentence_timings.py)."""
 
 import os
 import subprocess
@@ -6,11 +6,9 @@ import tempfile
 import unittest
 from unittest.mock import MagicMock, patch
 
-from apps.api.app.tasks.lipsync_tts import (
-    _compute_sentence_timings,
-    _estimate_sentence_timings_by_chars,
-    _split_script_into_sentences,
-)
+from packages.domain.sentence_timings import compute_sentence_timings as _compute_sentence_timings
+from packages.domain.sentence_timings import estimate_sentence_timings_by_chars as _estimate_sentence_timings_by_chars
+from packages.domain.sentence_timings import split_script_into_sentences as _split_script_into_sentences
 
 
 class TestSplitScriptIntoSentences(unittest.TestCase):