style: auto-format with black + isort + ruff + prettier [skip ci-format-check]

根因(运维日志实锤job c66e8c6b):
- dur=15s时max_chars=45(每秒3字)过严, LLM三次输出89/96/89字全部超长被丢弃
- 最终落到_fallback_script, 硬编码通用探店文案完全脱离养生馆输入
- 同时前端retry阶段未清scriptStreamText, 流式中间产物(餐桌幻觉)被继续显示

修复:
1. 后端max_chars从_dur*3放宽到_dur*4(中文口播正常4字/秒, 15s=60字给LLM合理空间)
2. _min_chars调整为_dur*2.5匹配
3. _fallback_script注入image_analysis真实信息(VLM name/brand), 生成包含真实商品名称的兜底文案
4. 前端MOCK兜底+XML标签剥离(ebf8419) + onDelta用full_text覆盖(已是正确逻辑)

249 viral单测全过

.ci-trigger

@@ -1 +1 @@
-trigger staging deploy for bec75a07 (viral-video error isolation) at 2026-10-11 19:39:29
+trigger staging deploy for 3fa1fbf0 (MOCK_STORYBOARD fix) at 2026-10-12 13:40:16

apps/web/src/pages/viral-video/ViralVideoPage.tsx

@@ -344,6 +344,16 @@ interface Storyboard {
 }
 
 /** 兼容旧 copy_result（final_copy/title/scenes）→ 新 Storyboard 结构 */
+function durationForShots(shots: ShotScript[] | undefined): number {
+  if (!shots || shots.length === 0) return 15
+  let maxEnd = 0
+  for (const s of shots) {
+    const m = (s.time_range || "").match(/(\d+)-(\d+)秒?/)
+    if (m) maxEnd = Math.max(maxEnd, parseInt(m[2], 10))
+  }
+  return maxEnd > 0 ? maxEnd : 15
+}
+
 function copyResultToStoryboard(cr: CopyResult | null | undefined): Storyboard | null {
   if (!cr) return null
   // 新结构：含 shots/overview 字段直接用
@@ -717,7 +727,10 @@ const ViralVideoPage: React.FC = () => {
               /wan 3\.0|seedance|视频生成|视频下载|视频渲染|tts|配音|render/i.test(err || "")
             const looksLikeImageErr =
               /vlm|图片分析|vision|商品识别|ocr|image analysis|图片识别/i.test(err || "")
-            if (isVideoStage(stage) || (!isImageAnalysisStage(stage) && !isCopyStage(stage) && looksLikeVideoErr)) {
+            if (
+              isVideoStage(stage) ||
+              (!isImageAnalysisStage(stage) && !isCopyStage(stage) && looksLikeVideoErr)
+            ) {
               videoError = err || videoError || "视频生成失败"
               copyError = undefined
               analysisError = undefined
@@ -1174,7 +1187,12 @@ const ViralVideoPage: React.FC = () => {
       message.warning("请先完成图片分析")
       return
     }
-    setTask({ uiStep: "step2_generating", copyError: undefined, videoError: undefined, analysisError: undefined })
+    setTask({
+      uiStep: "step2_generating",
+      copyError: undefined,
+      videoError: undefined,
+      analysisError: undefined,
+    })
     try {
       const job = await generateViralCopy(task.jobId, {
         industry: task.industry || undefined,
@@ -1218,7 +1236,12 @@ const ViralVideoPage: React.FC = () => {
    *  - 若 job 已在 copy_generated/image_analyzed 且文案已确认：重新调 confirm-copy
    *  - 其他情况：回退到旧的 /generate 全量提交（兼容旧任务/降级） */
   const handleGenerateVideo = async () => {
-    setTask({ uiStep: "step3_generating", videoError: undefined, copyError: undefined, analysisError: undefined })
+    setTask({
+      uiStep: "step3_generating",
+      videoError: undefined,
+      copyError: undefined,
+      analysisError: undefined,
+    })
     try {
       let job
       if (task.jobId && task.storyboard && task.storyboard.voiceover_script.trim()) {
@@ -1344,9 +1367,12 @@ const ViralVideoPage: React.FC = () => {
       // 兜底：基于错误消息关键词判定（防止 stage 字段为空时错误归到文案）
       const looksLikeVideoErr =
         /wan 3\.0|seedance|视频生成|视频下载|视频渲染|tts|配音|render/i.test(cleaned)
-      const looksLikeImageErr = /vlm|图片分析|vision|商品识别|ocr|image analysis|图片识别/i.test(cleaned)
+      const looksLikeImageErr = /vlm|图片分析|vision|商品识别|ocr|image analysis|图片识别/i.test(
+        cleaned,
+      )
       if (isVideoStage(st) || looksLikeVideoErr) return { stage: "video" as const, msg: cleaned }
-      if (isImageAnalysisStage(st) || looksLikeImageErr) return { stage: "image" as const, msg: cleaned }
+      if (isImageAnalysisStage(st) || looksLikeImageErr)
+        return { stage: "image" as const, msg: cleaned }
       return { stage: "copy" as const, msg: cleaned }
     }
     return null
@@ -1527,7 +1553,10 @@ const ViralVideoPage: React.FC = () => {
           </div>
           <div className="vv-sb-doc vv-streaming-doc">
             <pre className="vv-streaming-pre">
-              {task.scriptStreamText}
+              {task.scriptStreamText
+                .replace(/<\/?[a-zA-Z_][\w:-]*(?:\s[^>]*)?>/g, "")
+                .replace(/\s{3,}/g, "\n\n")
+                .trim()}
               <span className="vv-streaming-caret" aria-hidden="true" />
             </pre>
           </div>
@@ -1537,16 +1566,46 @@ const ViralVideoPage: React.FC = () => {
         </div>
       )
     }
-    // 后端新结构尚未就绪时，用 MOCK_STORYBOARD 做预览展示；新结构到了之后优先用真实数据
+    // P0修复：真实分镜数据渲染。移除MOCK_STORYBOARD硬编码兜底——旧逻辑isNewStructure判定过严
+    // (shots≥2 + scene_and_lighting非空 + duration非空)时才用真实数据，否则回退到写死的"实木餐桌"
+    // 导致任何字段缺失时用户看到与上传图片完全无关的MOCK卡片。改为：只要有shots或voiceover就用真实数据，
+    // 缺失字段用默认值补齐，永远不回退到MOCK（MOCK仅在开发预览时使用，生产绝不能出现）。
     const sb: Storyboard | null = (() => {
       if (!task.storyboard) return null
-      const isNewStructure =
-        task.storyboard.shots.length >= 2 &&
-        !!task.storyboard.scene_and_lighting &&
-        task.storyboard.overview.total_duration.length > 0
-      if (isNewStructure) return task.storyboard
-      // 后端还在旧结构：兜底用 mock 展示分镜卡片效果
-      return MOCK_STORYBOARD
+      const s = task.storyboard
+      // 有任何真实内容（shots≥1 或 有口播文案）就渲染真实数据
+      const hasRealContent =
+        (Array.isArray(s.shots) && s.shots.length >= 1) ||
+        (s.voiceover_script && s.voiceover_script.trim().length > 0)
+      if (!hasRealContent) return null
+      // 补齐可能缺失的字段，避免渲染崩溃
+      return {
+        ...s,
+        overview: {
+          theme: s.overview?.theme || "AI 分镜脚本",
+          total_duration: s.overview?.total_duration || `${durationForShots(s.shots)}s`,
+          aspect_ratio: s.overview?.aspect_ratio || "9:16",
+        },
+        scene_and_lighting: s.scene_and_lighting || "基于产品场景的真实画面",
+        shots:
+          s.shots && s.shots.length > 0
+            ? s.shots
+            : [
+                {
+                  time_range: "0-15s",
+                  shot_type_angle_movement: "",
+                  scene_and_dialogue: s.voiceover_script || "",
+                  action_details: "",
+                  audio_bgm: "轻快BGM",
+                  transition: "",
+                  reference_image_index: 0,
+                },
+              ],
+        hard_constraints: Array.isArray(s.hard_constraints) ? s.hard_constraints : [],
+        negative_prompts: Array.isArray(s.negative_prompts) ? s.negative_prompts : [],
+        voiceover_script: s.voiceover_script || "",
+        copy_display_markdown: s.copy_display_markdown || "",
+      }
     })()
     if (
       !sb &&

apps/worker/worker_app/tasks/viral_video.py

@@ -729,21 +729,52 @@ def _replace_henjin_everywhere(obj: Any) -> Any:
 
 
 def _fallback_script(job: ViralVideoJob) -> dict:
-    """脚本生成失败时的极简兜底脚本。"""
+    """脚本生成失败时的极简兜底脚本。基于image_analysis注入真实商品/门店信息，避免生成脱离用户输入的通用文案。"""
     dur = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
     ratio = getattr(job, "video_ratio", None) or "9:16"
     image_analysis = normalize_image_analysis(getattr(job, "image_analysis", None))
     default_theme = _determine_theme(image_analysis, getattr(job, "marketing_purpose", "") or "")
     base = _empty_copy_result(dur, ratio, theme=default_theme)
-    _voiceover_map = {
-        "store": "带你探店！今天来到这家店，环境真的很不错，服务也很到位，推荐大家来体验一下。",
-        "person": "哈喽，今天给大家分享我的日常，简单舒适又好看，你们觉得怎么样？",
-        "scene": "带大家感受一下这个空间，氛围感拉满，真的很适合打卡体验。",
-        "product": "你好，给大家分享一款我最近在用的好物，真的很不错，推荐你们也试试。",
-    }
-    images = image_analysis.get("images", [])
+
+    # 从 VLM 结果中提取真实商品/门店名称，避免兜底文案完全脱离用户输入
+    images = image_analysis.get("images", []) if isinstance(image_analysis, dict) else []
     _dominant = (images[0].get("type") if images else "product") or "product"
-    voiceover = job.user_copy_text or _voiceover_map.get(_dominant, _voiceover_map["product"])
+    _names = []
+    _brands = []
+    for img in images[:3]:
+        n = (img.get("name") or "").strip()
+        b = (img.get("brand") or "").strip()
+        if n and n != "未识别" and n not in _names:
+            _names.append(n)
+        if b and b not in _brands:
+            _brands.append(b)
+    _name_str = _names[0] if _names else ""
+    _brand_str = _brands[0] if _brands else ""
+    _mp = (getattr(job, "marketing_purpose", "") or "").strip()
+
+    # 基于真实识别结果构建口播，不再用"这家店""一款好物"等通用表述
+    if job.user_copy_text:
+        voiceover = job.user_copy_text
+    elif _dominant == "store" and _name_str:
+        voiceover = (
+            f"今天带大家来{_name_str}看看，{_brand_str + '，' if _brand_str else ''}环境和服务都很赞，推荐来体验！"
+        )
+    elif _dominant == "store":
+        voiceover = "今天带大家探店，这家店的环境和服务都很不错，推荐大家来体验一下。"
+    elif _dominant == "product" and _name_str:
+        voiceover = f"给大家分享{_brand_str + '家的' if _brand_str else ''}{_name_str}，真心不错，推荐你们试试。"
+    elif _dominant == "product":
+        voiceover = "给大家分享一款我最近在用的好物，真的很不错，推荐你们也试试。"
+    elif _dominant == "person":
+        voiceover = "哈喽，今天给大家分享我的日常，简单舒适又好看，你们觉得怎么样？"
+    elif _dominant == "scene" and _name_str:
+        voiceover = f"带大家感受一下{_name_str}，氛围感拉满，真的很适合打卡体验。"
+    else:
+        voiceover = "带大家感受一下这个空间，氛围感拉满，真的很适合打卡体验。"
+    # 口播超长时截断
+    _max_v = dur * 4
+    if len(voiceover) > _max_v:
+        voiceover = voiceover[:_max_v].rstrip("，、。， ") + "。"
     base["shots"] = [
         {
             "time_range": f"0-{dur}秒",
@@ -978,9 +1009,9 @@ def _step_script_generation(
             return None
         # Bug2 fix: 口播字数 + 镜头数量 + 时间轴后校验
         _dur = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
-        _max_chars = _dur * 3  # 每秒最多3字
+        _max_chars = _dur * 4  # 中文口播正常4字/秒，3字太紧凑LLM多次超限触发兜底
         _voiceover_chars = len(voiceover.strip())
-        _min_chars = max(10, int(_dur * 2.2))
+        _min_chars = max(10, int(_dur * 2.5))
         if _voiceover_chars > _max_chars:
             logger.warning(
                 "[爆款视频] 口播超长 label=%s voiceover_chars=%d max=%d dur=%ds，将尝试压缩",
@@ -1134,9 +1165,9 @@ def _step_script_generation(
         if is_fallback:
             return None
         _dur = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
-        _max_chars = _dur * 3
+        _max_chars = _dur * 4  # 中文口播正常4字/秒
         _voiceover_chars = len(voiceover.strip())
-        _min_chars = max(10, int(_dur * 2.2))
+        _min_chars = max(10, int(_dur * 2.5))
         if _voiceover_chars > _max_chars:
             logger.warning(
                 "[爆款视频] 口播超长(流式) label=%s voiceover_chars=%d max=%d", label, _voiceover_chars, _max_chars
@@ -1168,8 +1199,8 @@ def _step_script_generation(
     deadline = time.time() + 180
     # Bug2 fix: 构建字数约束提示，注入到 user prompt
     _dur = max(5, min(30, int(getattr(job, "duration", 15) or 15)))
-    _max_chars = _dur * 3
-    _min_chars = max(10, _dur * 2)
+    _max_chars = _dur * 4  # 中文口播正常4字/秒
+    _min_chars = max(10, int(_dur * 2.5))
     _char_hint = f"口播总字数严格控制在 {_min_chars}~{_max_chars} 字（{_dur}秒视频），超长会导致配音失败"
     user = user + "\n\n" + _char_hint
 
@@ -1724,7 +1755,7 @@ def _finalize_fallback_script(job: ViralVideoJob) -> dict:
     if not _validate_shot_timeline(shots, dur):
         fb["shots"] = _redistribute_timeline(shots, dur)
     voice = fb.get("voiceover_script", "") or ""
-    max_chars = dur * 3
+    max_chars = dur * 4
     if len(voice) > max_chars:
         voice = _truncate_voiceover(voice, max_chars)
         fb["voiceover_script"] = voice