后端生成闸+多项修复;前端全站更新;QA 审计与报告

后端:
- 新增 celery_health 生成前置闸——无 worker 在线时图片/视频生成入口
  一律 503,防"提交到 ARK 后无人轮询、结果悬空+额度冻结"的数据丢失
- 拼接导出:帧率跟随源众数、字幕逐句重映射、原声/BGM 混音修复
- 资金核算审计脚本 + genesis 账目回填 + 滞留预留清理命令
- 接入 yunqi provider 与豆包 TTS 模型(catalog/migrations/bootstrap 命令)

前端:全站页面更新(pipeline/library/products/projects/team/account 等),
新增共享 pager 分页组件

QA:刷新 function-audit 全量输出,新增 full-qa 报告
文档:BP 产品介绍资料、design/CLAUDE.md

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

core/backend/apps/ai/services.py

@@ -10,7 +10,7 @@ from django.db import transaction
 from django.utils import timezone
 
 from apps.ai.models import AITask, ModelConfig
-from apps.ai.providers import VolcanoArkProvider
+from apps.ai.providers import TtsNotConfigured, VolcanoArkProvider, VolcanoTtsProvider, YunqiProvider
 from apps.assets.models import Asset, AssetFile
 from apps.assets.storage import TosStorage
 from apps.billing.services.ledger import charge_reserved_credit, release_credit, reserve_credit
@@ -22,6 +22,7 @@ from apps.projects.models import (
     ScriptVersion,
     StoryboardFrame,
     StoryboardVersion,
+    Timeline,
     VideoSegment,
     VideoSegmentVersion,
 )
@@ -36,6 +37,13 @@ def get_default_model(capability: str) -> ModelConfig:
     )
 
 
+def get_image_provider(model_config: ModelConfig):
+    """生图按 provider 分流:yunqi 走 OpenAI 兼容网关(gpt-image-2),其余沿用火山 ARK。"""
+    if model_config.provider.name == "yunqi":
+        return YunqiProvider(base_url=model_config.provider.base_url or None)
+    return VolcanoArkProvider(base_url=model_config.provider.base_url or None)
+
+
 def estimate_cost(model_config: ModelConfig) -> Decimal:
     return model_config.unit_price if model_config.unit_price > 0 else Decimal("1.0000")
 
@@ -48,7 +56,9 @@ def build_script_prompt(*, project, user_prompt: str, selling_point_ids: list[st
     selling_text = "\n".join(f"- {item.title}: {item.detail}" for item in selling_points)
     system = (
         "你是电商短视频脚本导演。请为 9:16 竖屏带货短视频生成 60 秒脚本，"
-        "拆成 4 个 15 秒段落。每段包含旁白、画面描述、商品露出方式和转场建议。"
+        "拆成 4 个 15 秒段落。严格按以下格式输出，段落之间空一行，不要输出其他内容：\n"
+        "镜头1\n旁白：这一镜要念出来的口播文案（一两句话）\n画面：这一镜的画面描述、商品露出方式和转场建议\n\n"
+        "镜头2\n旁白：…\n画面：…（依此类推到镜头4）"
     )
     user = f"""
 商品标题：{product.title}
@@ -65,6 +75,42 @@ def build_script_prompt(*, project, user_prompt: str, selling_point_ids: list[st
     return [{"role": "system", "content": system}, {"role": "user", "content": user}]
 
 
+def parse_segment_fields(block: str) -> tuple[str, str]:
+    """从一镜文本里拆出(旁白, 画面)。
+
+    模型按 build_script_prompt 的格式输出「旁白：…/画面：…」标签行时精确拆分;
+    自带脚本/旧格式没有标签则两个字段都用整段(保持旧行为),字幕/故事板各自兜底。
+    """
+    narration_lines: list[str] = []
+    visual_lines: list[str] = []
+    current: list[str] | None = None
+    for raw in (block or "").splitlines():
+        line = raw.strip()
+        if not line:
+            continue
+        matched = re.match(r"^(旁白|口播|台词|文案)\s*[:：]\s*(.*)$", line)
+        if matched:
+            current = narration_lines
+            if matched.group(2):
+                current.append(matched.group(2))
+            continue
+        matched = re.match(r"^(画面|镜头描述|视觉|画面描述)\s*[:：]\s*(.*)$", line)
+        if matched:
+            current = visual_lines
+            if matched.group(2):
+                current.append(matched.group(2))
+            continue
+        if re.match(r"^(镜头|分镜|场)\s*\d+", line):
+            continue  # 「镜头N」标题行不计入任何字段
+        if current is not None:
+            current.append(line)
+    narration = " ".join(narration_lines).strip()
+    visual = " ".join(visual_lines).strip()
+    if not narration and not visual:
+        return block.strip(), block.strip()
+    return narration or visual, visual or narration
+
+
 def split_script_into_segments(content: str, count: int = 4) -> list[str]:
     """把一段脚本稳健地拆成 `count` 个分镜文本,保证每镜都非空、且所有内容都被分配到某一镜。
 
@@ -124,7 +170,7 @@ def create_ai_task(*, project, user, task_type: str, model_config: ModelConfig,
     return task
 
 
-def generate_project_script(*, project, user, user_prompt: str, selling_point_ids: list[str] | None = None) -> ScriptVersion:
+def generate_project_script(*, project, user, user_prompt: str, selling_point_ids: list[str] | None = None, source: str = "ai") -> ScriptVersion:
     model_config = get_default_model(ModelConfig.Capability.TEXT)
     if model_config is None:
         raise ValueError("no active text model configured")
@@ -162,16 +208,17 @@ def generate_project_script(*, project, user, user_prompt: str, selling_point_id
                 task=task,
                 title="AI 脚本",
                 content=content,
-                source="ai",
+                source=source if source in ("ai", "theme", "manual") else "ai",
                 is_adopted=False,
             )
             for index, segment_text in enumerate(split_script_into_segments(content)):
+                narration, visual = parse_segment_fields(segment_text)
                 ScriptSegment.objects.create(
                     script_version=script,
                     sort_order=index,
                     duration_seconds=15,
-                    narration=segment_text,
-                    visual_prompt=segment_text,
+                    narration=narration,
+                    visual_prompt=visual,
                 )
 
             stage, _ = ProjectStage.objects.get_or_create(project=project, stage=ProjectStage.Stage.SCRIPT)
@@ -283,7 +330,7 @@ def generate_base_asset(*, project, user, kind: str, prompt: str) -> BaseAssetGr
     )
     reservation = task.credit_reservation
     try:
-        provider = VolcanoArkProvider(base_url=model_config.provider.base_url or None)
+        provider = get_image_provider(model_config)
         response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=prompt)
         media = provider.extract_first_media_url(response)
         with transaction.atomic():
@@ -408,7 +455,7 @@ def _storyboard_frame_worker(task_id, version_id, segment_id, user_id) -> None:
         task.status = AITask.Status.SUBMITTED
         task.save(update_fields=["status", "updated_at"])
         try:
-            provider = VolcanoArkProvider(base_url=model_config.provider.base_url or None)
+            provider = get_image_provider(model_config)
             frame_prompt = task.request_payload.get("prompt") or build_storyboard_frame_prompt(project, version, segment)
             response = provider.image_generation(
                 model=model_config.name,
@@ -416,12 +463,9 @@ def _storyboard_frame_worker(task_id, version_id, segment_id, user_id) -> None:
                 prompt=frame_prompt,
             )
             media = provider.extract_first_media_url(response)
-            task.status = AITask.Status.SUCCEEDED
-            task.response_payload = response
-            task.actual_cost = task.estimated_cost
-            task.completed_at = timezone.now()
-            task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
-            charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
+            # 注意顺序:task 是 poll 端的「占位锁」,必须等帧真正落库后才置 SUCCEEDED。
+            # 旧实现先置 SUCCEEDED 再上传 TOS(数秒)最后建帧,中间窗口 poll 会判「无在途且帧缺失」
+            # 为同一镜重复起线程 → 重复帧 + 重复扣费(实测 4 帧出 6 帧)。
             asset = _store_generated_media(
                 team=project.team,
                 user=user,
@@ -432,13 +476,22 @@ def _storyboard_frame_worker(task_id, version_id, segment_id, user_id) -> None:
                 category=Asset.Category.SCENE,
                 asset_type=Asset.Type.IMAGE,
             )
-            StoryboardFrame.objects.create(
-                storyboard=version,
-                script_segment=segment,
-                asset=asset,
-                sort_order=segment.sort_order,
-                prompt=segment.visual_prompt,
-            )
+            with transaction.atomic():
+                task.status = AITask.Status.SUCCEEDED
+                task.response_payload = response
+                task.actual_cost = task.estimated_cost
+                task.completed_at = timezone.now()
+                task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
+                charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
+                # 幂等守卫:该镜已有帧(任何残余竞态/双 poll)就不再建,保持一镜一帧
+                if not StoryboardFrame.objects.filter(storyboard=version, script_segment=segment).exists():
+                    StoryboardFrame.objects.create(
+                        storyboard=version,
+                        script_segment=segment,
+                        asset=asset,
+                        sort_order=segment.sort_order,
+                        prompt=segment.visual_prompt,
+                    )
         except Exception as exc:  # noqa: BLE001 — 失败回滚额度,标记任务失败供 poll 上报
             task.status = AITask.Status.FAILED
             task.error_message = str(exc)
@@ -470,53 +523,64 @@ def generate_storyboard_frame(*, project, user) -> dict:
         _finalize_storyboard(project, version)
         return {"status": "succeeded", "done": total, "total": total, "version_id": str(version.id)}
 
-    # 该版本内是否已有帧在后台生成中(RESERVED/SUBMITTED 的故事板任务即为「占位锁」)。
-    # 仅算「近 3 分钟内」的任务:若进程/线程意外中断留下僵尸任务,超时后不再视为在生成,允许重新发起。
+    # 每镜独立「占位锁」(该镜有 CREATED/RESERVED/SUBMITTED 的任务=在生成中)。
+    # 旧实现是版本级单锁、一次只生成一帧;改为缺哪几镜就同时起哪几镜的线程。
+    # ★ 实测注记(2026-06-10):4 线程并发时当前 Seedream 端点在服务侧排队,单帧 25s→83-105s,
+    #   整版总时长 ≈ 串行(117s)。瓶颈是 ARK 端点并发配额而非本机;并行无额外成本,
+    #   配额提升后自动受益。可用 settings.STORYBOARD_MAX_PARALLEL 调并发(1=回到串行)。
+    # 仅算「近 3 分钟内」的任务:线程意外中断留下的僵尸任务超时后不再占锁,允许重新发起。
+    from django.conf import settings as dj_settings
+    STORYBOARD_MAX_PARALLEL = int(getattr(dj_settings, "STORYBOARD_MAX_PARALLEL", 4))
     stale_cutoff = timezone.now() - timedelta(minutes=3)
-    inflight = AITask.objects.filter(
-        project=project,
-        task_type=AITask.Type.STORYBOARD,
-        status__in=[AITask.Status.CREATED, AITask.Status.RESERVED, AITask.Status.SUBMITTED],
-        request_payload__storyboard_version=str(version.id),
-        created_at__gte=stale_cutoff,
-    ).exists()
-    if inflight:
-        return {"status": "generating", "done": done, "total": total, "version_id": str(version.id)}
+    inflight_segment_ids = {
+        str(v)
+        for v in AITask.objects.filter(
+            project=project,
+            task_type=AITask.Type.STORYBOARD,
+            status__in=[AITask.Status.CREATED, AITask.Status.RESERVED, AITask.Status.SUBMITTED],
+            request_payload__storyboard_version=str(version.id),
+            created_at__gte=stale_cutoff,
+        ).values_list("request_payload__storyboard_segment", flat=True)
+        if v
+    }
 
     pending = [s for s in segments if s.id not in done_segment_ids]
-    segment = pending[0]
-    # 单帧失败次数上限,避免持续失败时无限重试
-    failed_for_segment = AITask.objects.filter(
-        project=project,
-        task_type=AITask.Type.STORYBOARD,
-        status=AITask.Status.FAILED,
-        request_payload__storyboard_segment=str(segment.id),
-    ).count()
-    if failed_for_segment >= 2:
-        last = AITask.objects.filter(project=project, task_type=AITask.Type.STORYBOARD, status=AITask.Status.FAILED,
-                                     request_payload__storyboard_segment=str(segment.id)).order_by("-created_at").first()
-        return {"status": "failed", "done": done, "total": total, "version_id": str(version.id),
-                "error": last.error_message if last else "storyboard frame failed"}
+    # 单帧失败次数上限,避免持续失败时无限重试;任一镜到上限即整版上报失败
+    for segment in pending:
+        failed_for_segment = AITask.objects.filter(
+            project=project,
+            task_type=AITask.Type.STORYBOARD,
+            status=AITask.Status.FAILED,
+            request_payload__storyboard_segment=str(segment.id),
+        ).count()
+        if failed_for_segment >= 2:
+            last = AITask.objects.filter(project=project, task_type=AITask.Type.STORYBOARD, status=AITask.Status.FAILED,
+                                         request_payload__storyboard_segment=str(segment.id)).order_by("-created_at").first()
+            return {"status": "failed", "done": done, "total": total, "version_id": str(version.id),
+                    "error": last.error_message if last else "storyboard frame failed"}
 
+    spawnable = [s for s in pending if str(s.id) not in inflight_segment_ids]
+    slots = max(0, STORYBOARD_MAX_PARALLEL - len(inflight_segment_ids))
     model_config = get_default_model(ModelConfig.Capability.IMAGE)
-    task = create_ai_task(
-        project=project,
-        user=user,
-        task_type=AITask.Type.STORYBOARD,
-        model_config=model_config,
-        request_payload={
-            "model": model_config.name,
-            "endpoint": model_config.endpoint,
-            "prompt": build_storyboard_frame_prompt(project, version, segment),
-            "storyboard_version": str(version.id),
-            "storyboard_segment": str(segment.id),
-        },
-    )
-    threading.Thread(
-        target=_storyboard_frame_worker,
-        args=(str(task.id), str(version.id), str(segment.id), str(user.id)),
-        daemon=True,
-    ).start()
+    for segment in spawnable[:slots]:
+        task = create_ai_task(
+            project=project,
+            user=user,
+            task_type=AITask.Type.STORYBOARD,
+            model_config=model_config,
+            request_payload={
+                "model": model_config.name,
+                "endpoint": model_config.endpoint,
+                "prompt": build_storyboard_frame_prompt(project, version, segment),
+                "storyboard_version": str(version.id),
+                "storyboard_segment": str(segment.id),
+            },
+        )
+        threading.Thread(
+            target=_storyboard_frame_worker,
+            args=(str(task.id), str(version.id), str(segment.id), str(user.id)),
+            daemon=True,
+        ).start()
     return {"status": "generating", "done": done, "total": total, "version_id": str(version.id)}
 
 
@@ -656,16 +720,17 @@ def poll_video_segment(*, video_segment: VideoSegment, user) -> VideoSegmentVers
     if video_segment.status == VideoSegment.Status.FAILED:
         return None
 
-    task = video_segment.versions.order_by("-created_at").first()
-    ai_task = None
-    if task:
-        ai_task = task.task
+    # ★ 先找「在途任务」再回退旧版本的任务。旧实现反过来:重跑时段上已有(旧)版本,
+    # 取到旧版本挂的已成功任务 → 短路返回旧版,在途的新任务永远没人轮询,
+    # 段永远卡「生成中」、新视频取不回来(实测重跑卡 40 分钟,ARK 侧其实早已生成完)。
+    ai_task = video_segment.project.ai_tasks.filter(
+        task_type=AITask.Type.VIDEO_SEGMENT,
+        request_payload__video_segment_id=str(video_segment.id),
+        status__in=[AITask.Status.SUBMITTED, AITask.Status.POLLING],
+    ).order_by("-created_at").first()
     if ai_task is None:
-        ai_task = video_segment.project.ai_tasks.filter(
-            task_type=AITask.Type.VIDEO_SEGMENT,
-            request_payload__video_segment_id=str(video_segment.id),
-            status__in=[AITask.Status.SUBMITTED, AITask.Status.POLLING],
-        ).order_by("-created_at").first()
+        latest_version = video_segment.versions.order_by("-created_at").first()
+        ai_task = latest_version.task if latest_version else None
     if ai_task is None:
         raise ValueError("no active video generation task")
 
@@ -679,9 +744,11 @@ def poll_video_segment(*, video_segment: VideoSegment, user) -> VideoSegmentVers
     response = provider.poll_video_task(endpoint=ai_task.model_config.endpoint, provider_task_id=ai_task.provider_task_id)
     remote_status = response.get("status")
     if remote_status in {"queued", "running", "processing"}:
-        ai_task.status = AITask.Status.POLLING
-        ai_task.response_payload = response
-        ai_task.save(update_fields=["status", "response_payload", "updated_at"])
+        # 仍在生成:只在状态首次进入 POLLING 时落一次库。旧实现每次 poll(5s 一次)都把完整
+        # response JSON 回写远程 MySQL——纯浪费写带宽,终态时反正会存完整 payload。
+        if ai_task.status != AITask.Status.POLLING:
+            ai_task.status = AITask.Status.POLLING
+            ai_task.save(update_fields=["status", "updated_at"])
         return None
     if remote_status in {"failed", "expired", "cancelled"}:
         ai_task.status = AITask.Status.FAILED
@@ -706,23 +773,33 @@ def poll_video_segment(*, video_segment: VideoSegment, user) -> VideoSegmentVers
         category=Asset.Category.VIDEO_CLIP,
         asset_type=Asset.Type.VIDEO,
     )
-    ai_task.status = AITask.Status.SUCCEEDED
-    ai_task.response_payload = response
-    ai_task.actual_cost = ai_task.estimated_cost
-    ai_task.completed_at = timezone.now()
-    ai_task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
-    charge_reserved_credit(reservation=ai_task.credit_reservation, actual_amount=ai_task.actual_cost)
-    version = VideoSegmentVersion.objects.create(
-        video_segment=video_segment,
-        task=ai_task,
-        asset=asset,
-        prompt=ai_task.request_payload.get("prompt", ""),
-        is_adopted=True,
-    )
-    video_segment.adopted_version = version
-    video_segment.status = VideoSegment.Status.SUCCEEDED
-    video_segment.error_message = ""
-    video_segment.save(update_fields=["adopted_version", "status", "error_message", "updated_at"])
+    # 终态化必须持锁原子做:两个并发 poll(前端 5s 静默轮询 × 提交后轮询/worker)同时走到这里时,
+    # 旧实现会同 task 建两个版本 + charge_reserved_credit 双扣费(实测 03:08:25 同秒双版本)。
+    # select_for_update 锁 task 行,后到者看到 SUCCEEDED 直接回已有版,不再建版/扣费。
+    with transaction.atomic():
+        locked_task = AITask.objects.select_for_update().get(id=ai_task.id)
+        if locked_task.status == AITask.Status.SUCCEEDED:
+            existing = video_segment.versions.filter(task=locked_task).order_by("-created_at").first()
+            if existing is not None:
+                return existing
+        locked_task.status = AITask.Status.SUCCEEDED
+        locked_task.response_payload = response
+        locked_task.actual_cost = locked_task.estimated_cost
+        locked_task.completed_at = timezone.now()
+        locked_task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
+        charge_reserved_credit(reservation=locked_task.credit_reservation, actual_amount=locked_task.actual_cost)
+        version = VideoSegmentVersion.objects.create(
+            video_segment=video_segment,
+            task=locked_task,
+            asset=asset,
+            prompt=locked_task.request_payload.get("prompt", ""),
+            is_adopted=True,
+        )
+        video_segment.versions.exclude(id=version.id).update(is_adopted=False)
+        video_segment.adopted_version = version
+        video_segment.status = VideoSegment.Status.SUCCEEDED
+        video_segment.error_message = ""
+        video_segment.save(update_fields=["adopted_version", "status", "error_message", "updated_at"])
     return version
 
 
@@ -750,7 +827,7 @@ def generate_standalone_image(*, team, user, prompt: str, mode: str = "image", c
     category = _STANDALONE_CATEGORY.get(mode, Asset.Category.UNCATEGORIZED)
     task_type = _STANDALONE_TASK_TYPE.get(mode, AITask.Type.PRODUCT_IMAGE)
     count = max(1, min(int(count or 1), 4))
-    provider = VolcanoArkProvider(base_url=model_config.provider.base_url or None)
+    provider = get_image_provider(model_config)
     assets: list[Asset] = []
     for index in range(count):
         cost = estimate_cost(model_config)
@@ -798,3 +875,112 @@ def generate_standalone_image(*, team, user, prompt: str, mode: str = "image", c
             release_credit(reservation=reservation, reason=str(exc))
             raise
     return assets
+
+
+# ── 旁白配音(TTS):每镜旁白合成一段语音,导出时作为人声轨混在 BGM 之上 ──
+
+# 音色按「语音合成(经典版)」试用包实测可用清单配置;大模型音色(*_bigtts)需另开通「语音合成大模型」服务,当前账号 403
+VOICEOVER_VOICES = [
+    {"key": "BV700_streaming", "label": "灿灿 · 活力女声"},
+    {"key": "BV034_streaming", "label": "知性姐姐 · 沉稳女声"},
+    {"key": "BV001_streaming", "label": "通用女声"},
+    {"key": "BV056_streaming", "label": "阳光男声"},
+    {"key": "BV102_streaming", "label": "儒雅青年 · 解说男声"},
+    {"key": "BV002_streaming", "label": "通用男声"},
+]
+DEFAULT_VOICEOVER_VOICE = VOICEOVER_VOICES[0]["key"]
+
+
+def synthesize_project_voiceover(*, project, user, items: list[dict], voice_type: str, speed_ratio: float = 1.0) -> dict:
+    """每镜旁白 → **逐句** TTS 配音资产(一句一段音频,带句内起点 offset_ms),映射写入
+    timeline.metadata["voiceover"]。逐句才能支持「拖动字幕块 = 字幕和它的语音一起移动」;
+    一次调用 = 一个 AITask = 计一次费;任何一句失败则整体失败并释放预留(不留半套配音)。"""
+    from apps.projects.services.export import _split_subtitle_text
+
+    texts = []  # (片段 index, 句序 cue, 句文本)
+    for n, item in enumerate(items or []):
+        text = str(item.get("text") or "").strip()
+        if not text:
+            continue
+        idx = int(item.get("index", n))
+        pieces = _split_subtitle_text(text) or [text]
+        for j, piece in enumerate(pieces):
+            texts.append((idx, j, piece))
+    if not texts:
+        raise ValueError("没有可配音的旁白文本")
+    provider = VolcanoTtsProvider()
+    if not provider.configured:
+        raise TtsNotConfigured(
+            "语音合成未配置:请在后端环境变量设置 VOLC_TTS_APPID 和 VOLC_TTS_ACCESS_TOKEN"
+            "(火山引擎控制台 → 语音技术 → 语音合成大模型 → 创建应用)"
+        )
+    voice_type = voice_type or DEFAULT_VOICEOVER_VOICE
+    model_config = get_default_model(ModelConfig.Capability.AUDIO)
+    if model_config is None:
+        raise ValueError("no active audio model configured")
+    task = create_ai_task(
+        project=project,
+        user=user,
+        task_type=AITask.Type.VOICEOVER,
+        model_config=model_config,
+        request_payload={
+            "voice_type": voice_type,
+            "speed_ratio": float(speed_ratio or 1.0),
+            "items": [{"index": idx, "cue": j, "text": text} for idx, j, text in texts],
+        },
+    )
+    reservation = task.credit_reservation
+    try:
+        synthesized = []
+        for idx, j, text in texts:
+            audio, duration_ms = provider.synthesize(text=text, voice_type=voice_type, speed_ratio=speed_ratio, uid=str(user.id))
+            synthesized.append((idx, j, text, audio, duration_ms))
+        with transaction.atomic():
+            task.status = AITask.Status.SUCCEEDED
+            task.actual_cost = task.estimated_cost
+            task.completed_at = timezone.now()
+            task.response_payload = {"segments": len(synthesized)}
+            task.save(update_fields=["status", "actual_cost", "completed_at", "response_payload", "updated_at"])
+            charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
+            vo_items = []
+            offset_acc: dict[int, int] = {}  # 同一片段内逐句顺排:句 j 的默认起点 = 前面句时长之和
+            for idx, j, text, audio, duration_ms in synthesized:
+                asset_id = uuid.uuid4()
+                object_key = f"teams/{project.team_id}/projects/{project.id}/voiceover/{asset_id}.mp3"
+                stored = TosStorage().upload_fileobj(fileobj=BytesIO(audio), object_key=object_key, content_type="audio/mpeg")
+                asset = Asset.objects.create(
+                    id=asset_id, team=project.team, created_by=user,
+                    name=f"配音 · 场 {idx + 1} · 句 {j + 1}", asset_type=Asset.Type.AUDIO,
+                    source=Asset.Source.AI_GENERATED, category=Asset.Category.UNCATEGORIZED,
+                    origin_task=task, description=text,
+                )
+                AssetFile.objects.create(
+                    asset=asset, object_key=stored.object_key, bucket=stored.bucket,
+                    content_type=stored.content_type, size_bytes=stored.size_bytes, is_primary=True,
+                )
+                offset_ms = offset_acc.get(idx, 0)
+                offset_acc[idx] = offset_ms + (duration_ms or 0)
+                vo_items.append({
+                    "index": idx, "cue": j, "text": text, "asset": str(asset.id),
+                    "duration_ms": duration_ms, "offset_ms": offset_ms,
+                })
+            timeline, _ = Timeline.objects.get_or_create(
+                project=project, defaults={"name": f"{project.name} Timeline", "duration_seconds": 60}
+            )
+            metadata = dict(timeline.metadata or {})
+            metadata["voiceover"] = {
+                "enabled": True,
+                "voice_type": voice_type,
+                "speed_ratio": float(speed_ratio or 1.0),
+                "items": vo_items,
+            }
+            timeline.metadata = metadata
+            timeline.save(update_fields=["metadata", "updated_at"])
+        return metadata["voiceover"]
+    except Exception as exc:
+        task.status = AITask.Status.FAILED
+        task.error_message = str(exc)
+        task.completed_at = timezone.now()
+        task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
+        release_credit(reservation=reservation, reason=str(exc))
+        raise