fix(core): 资产详情大图随数据显示,缩略图已出大图不再卡转圈

- 三视图/立绘大图改为跟数据(url)走,出图完成即显示,不再被在途轮询的
  busy 标志挡住——根治「小缩略图已出、大图一直转圈」
- 一并带上工作区其余改动(script_agent / services / tests / ai-tools / actor-library / App)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/backend/apps/ai/script_agent.py

@@ -72,6 +72,12 @@ _OUTPUT_PROTOCOL = """
 1. 先用 **1 句中文口语**告诉用户你正在做什么(≤40 字,例:「在为这款保温杯生成 4 镜痛点脚本…」),让用户看到进展;
 2. 紧接着输出**且仅输出一个** ```json 代码块,内容为符合技能契约(铁律1)的 ScriptDraft 对象;
 3. json 代码块之后**不要再写任何文字**。
+
+### 字段名锚定(硬性 · 下游靠它取数,跑偏即数据全空)
+- 分镜数组的键名**必须**叫 `segments`(禁止用 scenes / script / shots / 分镜 等同义词)。
+- 每镜口播键名**必须**叫 `narration`(禁止用 voiceover / audio / line)。
+- 每镜画面键名**必须**叫 `visual`,且为**一句话字符串**(禁止用 scene / screen / 画面,也禁止写成 {setting,camera,...} 对象)。
+- 即使你额外附带了 scenes / shots 等创作结构,也**必须同时**给出标准 `segments` 数组,并把口播填进 `narration`、画面填进 `visual`,否则视为不合格。
 """
 
 
@@ -215,31 +221,81 @@ def _nearest_duration(value) -> int:
 # 与其逐一追变体,不如「优先键命中 → 否则按关键词模糊匹配」通用解析。SKIP 掉明显的非内容键,
 # 避免误抓(shotNo/duration/bgMusic/note 等)。
 _PICK_SKIP_KEYS = {
-    "shotno", "shotsize", "shotsizetype", "duration", "starttime", "endtime", "timerange",
-    "bgmusic", "bg_music", "music", "sound", "sfx", "note", "notes", "tips", "index",
-    "role", "speaker", "entity_refs", "product_exposure", "id", "no", "transition",
+    "shotno", "shotsize", "shotsizetype", "duration", "duration_seconds", "starttime", "endtime",
+    "timerange", "time_range", "bgmusic", "bg_music", "music", "sound", "sfx", "note", "notes",
+    "tips", "index", "role", "speaker", "entity_refs", "product_exposure", "id", "no", "transition",
+    # 标题/编号类:含 scene 字样会误命中画面 fuzzy,显式跳过(注意:不跳裸 scene,它常=画面)
+    "scene_id", "sceneid", "scene_no", "sceneno", "scene_number", "scenenumber",
+    "scene_title", "scenetitle", "title", "scene_title_type",
 }
-_VISUAL_EXACT = ("visual", "visual_prompt", "visual_description", "scene", "screen_description", "screendescription", "screen", "picture", "shot_description")
-_VISUAL_FUZZY = ("scene", "screen", "visual", "picture", "画面", "镜头描述", "分镜画面", "描述")
-_NARRATION_EXACT = ("narration", "voiceover", "voice_over", "line", "caption", "subtitle", "speech", "口播", "旁白")
-_NARRATION_FUZZY = ("narrat", "voice", "旁白", "口播", "台词", "dialog", "caption", "subtitle", "字幕", "speech", "monolog")
+_VISUAL_EXACT = ("visual", "visual_prompt", "visual_description", "screen_description", "screendescription", "screen", "picture", "shot_description", "scene", "画面")
+_VISUAL_FUZZY = ("visual", "screen", "picture", "scene", "画面", "镜头描述", "分镜画面")
+_NARRATION_EXACT = ("narration", "voiceover", "voice_over", "vo", "line", "caption", "subtitle", "speech", "口播", "旁白")
+_NARRATION_FUZZY = ("narrat", "voice", "旁白", "口播", "台词", "dialog", "caption", "subtitle", "字幕", "speech", "monolog", "audio")
+
+
+def _flatten_text(val) -> str:
+    """把 字符串/字典/列表 里的文本拍平成一句。模型常把 visual 写成 {setting,camera,key_shots}
+    对象、把口播写成数组,这里统一抽成纯文本,避免结构化值被当空丢弃(全空根因之一)。"""
+    if isinstance(val, str):
+        return val.strip()
+    if isinstance(val, dict):
+        return " · ".join(p for p in (_flatten_text(v) for v in val.values()) if p)
+    if isinstance(val, (list, tuple)):
+        return " ".join(p for p in (_flatten_text(v) for v in val) if p)
+    return ""
 
 
 def _pick_field(seg: dict, exact: tuple[str, ...], fuzzy: tuple[str, ...]) -> str:
-    """从一镜里取某类文本字段:先按优先键精确命中,再按关键词在剩余键里模糊匹配(跳过非内容键)。"""
+    """从一镜里取某类文本字段:先按优先键精确命中,再按关键词在剩余键里模糊匹配(跳过非内容键)。
+    值允许是 字符串/对象/数组(嵌套结构拍平成一句),不再只认裸字符串。"""
     for key in exact:
-        val = seg.get(key)
-        if isinstance(val, str) and val.strip():
-            return val.strip()
+        if key in seg:
+            txt = _flatten_text(seg.get(key))
+            if txt:
+                return txt
     for key, val in seg.items():
         kl = str(key).lower()
-        if kl in _PICK_SKIP_KEYS or not (isinstance(val, str) and val.strip()):
+        if kl in _PICK_SKIP_KEYS:
             continue
         if any(f in kl for f in fuzzy):
-            return val.strip()
+            txt = _flatten_text(val)
+            if txt:
+                return txt
     return ""
 
 
+# 模型给分镜数组的键名五花八门(segments/scenes/script/shots…),且常同时给一个**空的**
+# segments 骨架 + 真内容放在 scenes 里。所以不能「见 segments 是 list 就用」,要在所有候选里
+# 挑「能解析出最多非空旁白/画面」的那个。segments(契约本名)排第一,同分时优先。
+_SEGMENT_ARRAY_KEYS = (
+    "segments", "shots", "scenes", "script", "shot_list", "shotlist", "shotList",
+    "scene_list", "scenes_list", "storyboard", "分镜", "镜头",
+)
+
+
+def _resolve_segments(draft: dict) -> list:
+    """在所有候选数组键里挑内容最丰富的分镜数组(按非空旁白/画面条数打分)。"""
+    best: list = []
+    best_score = -1
+    for key in _SEGMENT_ARRAY_KEYS:
+        arr = draft.get(key)
+        if not isinstance(arr, list) or not arr:
+            continue
+        dict_items = [s for s in arr if isinstance(s, dict)]
+        if not dict_items:
+            continue
+        score = sum(
+            1
+            for s in dict_items
+            if _pick_field(s, _NARRATION_EXACT, _NARRATION_FUZZY)
+            or _pick_field(s, _VISUAL_EXACT, _VISUAL_FUZZY)
+        )
+        if score > best_score:  # 严格大于 → 同分保留更靠前的键(segments 优先)
+            best, best_score = arr, score
+    return best
+
+
 def normalize_draft(raw_text: str, *, aspect_ratio: str, total_duration: int) -> dict:
     """把模型输出抽成 JSON 并按铁律1契约规范化。宽容:小问题就地修,不轻易抛错。"""
     blob = _extract_json(raw_text)
@@ -262,11 +318,8 @@ def normalize_draft(raw_text: str, *, aspect_ratio: str, total_duration: int) ->
         draft.setdefault("total_duration", basic.get("totalDuration") or basic.get("total_duration"))
         draft.setdefault("aspect_ratio", basic.get("aspectRatio") or basic.get("aspect_ratio"))
         draft.setdefault("theme", basic.get("theme"))
-    if not isinstance(draft.get("segments"), list):
-        for alias in ("shots", "scenes_list", "shotList"):
-            if isinstance(draft.get(alias), list):
-                draft["segments"] = draft[alias]
-                break
+    # 在所有候选键里挑内容最丰富的分镜数组(空 segments 骨架会被更丰富的 scenes/shots 顶替)
+    draft["segments"] = _resolve_segments(draft)
 
     draft["aspect_ratio"] = (draft.get("aspect_ratio") or aspect_ratio or "9:16").strip()
     dur = _nearest_duration(draft.get("total_duration") or total_duration)

core/backend/apps/ai/services.py

@@ -819,6 +819,34 @@ def build_storyboard_frame_prompt_refs(project, version, segment, refs: list[dic
     return "\n".join(line for line in lines if line)
 
 
+def _is_transient_error(exc: Exception) -> bool:
+    """网络抖动/超时类瞬时错误(可重试),区别于内容审核拦截、参数非法等确定性失败。
+    中转站(tokenssr 等)偶发 Read timeout / 连接重置会无谓掐掉单帧,这类才重试。"""
+    msg = str(exc).lower()
+    return any(
+        k in msg
+        for k in ("timed out", "timeout", "connection", "reset by peer", "temporarily",
+                  "bad gateway", "502", "503", "504", "remotedisconnected", "max retries")
+    )
+
+
+def _call_image_with_retry(fn, *, attempts: int = 3, base_delay: float = 2.0):
+    """对一次出图网络调用做有界重试:仅瞬时错误重试(指数退避),确定性失败立即抛出。
+    出图无副作用(失败=没拿到图),重试安全;成功一次即返回。"""
+    import time
+
+    last: Exception | None = None
+    for i in range(attempts):
+        try:
+            return fn()
+        except Exception as exc:  # noqa: BLE001
+            last = exc
+            if i == attempts - 1 or not _is_transient_error(exc):
+                raise
+            time.sleep(base_delay * (i + 1))
+    raise last  # 理论不可达(循环内已 return/raise)
+
+
 def _storyboard_frame_worker(task_id, version_id, segment_id, user_id) -> None:
     """后台线程:真正调 ARK 生成一帧故事板图并落库。每次 poll 不阻塞在此——HTTP 永远秒回。"""
     import threading  # noqa: F401 — 仅标注此函数运行在独立线程

core/backend/apps/ai/tests.py

@@ -80,6 +80,47 @@ class NormalizeDraftTests(SimpleTestCase):
         self.assertEqual(draft["segments"][0]["narration"], "一句完整口播")
         self.assertEqual(draft["segments"][0]["dialogue"], [])  # 字符串不进结构化对白
 
+    def test_empty_segments_skeleton_yields_to_richer_scenes(self):
+        """真实回归(全空根因):模型把内容放 scenes/voiceover/visual,却另给一个**空 segments 骨架**。
+        必须挑内容最丰富的数组(scenes),而不是见 segments 是 list 就用→落 4 个空镜。"""
+        raw = json.dumps({
+            "scenes": [
+                {"sceneIndex": 1, "sceneTitle": "通勤", "voiceover": "手机一卡效率掉线", "visual": "地铁口拿出亮银色机身"},
+                {"sceneIndex": 2, "sceneTitle": "办公", "voiceover": "A19多任务很跟手", "visual": "办公桌俯拍切换应用"},
+            ],
+            "segments": [{"index": 0, "role": "钩子", "narration": "", "visual": ""}, {"index": 1, "role": "痛点", "narration": "", "visual": ""}],
+            "total_duration": 30,
+        }, ensure_ascii=False)
+        draft = normalize_draft(raw, aspect_ratio="9:16", total_duration=30)
+        self.assertEqual(draft["segments"][0]["narration"], "手机一卡效率掉线")
+        self.assertEqual(draft["segments"][0]["visual"], "地铁口拿出亮银色机身")
+
+    def test_script_key_with_visual_object_is_flattened(self):
+        """真实回归:数组键叫 script、visual 写成 {setting,camera,key_shots} 对象。
+        必须认出 script 数组并把 visual 对象拍平成一句,而非留空。"""
+        raw = json.dumps({
+            "script": [
+                {"scene_id": 1, "scene_title": "通勤", "voiceover": "选手机看重流畅好看",
+                 "visual": {"setting": "地铁口", "camera": "竖屏手持跟拍", "key_shots": ["拿出亮银色机身"]}},
+            ],
+            "total_duration": 15,
+        }, ensure_ascii=False)
+        draft = normalize_draft(raw, aspect_ratio="9:16", total_duration=15)
+        seg = draft["segments"][0]
+        self.assertEqual(seg["narration"], "选手机看重流畅好看")
+        self.assertIn("地铁口", seg["visual"])
+        self.assertIn("竖屏手持跟拍", seg["visual"])
+
+    def test_audio_field_fills_narration(self):
+        """真实回归:shots 用 audio 放口播(GPT 变体),旁白曾因 audio 不在词典而留空。"""
+        raw = json.dumps({
+            "shots": [{"scene": 1, "visual": "咖啡厅办公把玩机身", "audio": "这款亮银色是我的高光决定"}],
+            "total_duration": 15,
+        }, ensure_ascii=False)
+        draft = normalize_draft(raw, aspect_ratio="9:16", total_duration=15)
+        self.assertEqual(draft["segments"][0]["narration"], "这款亮银色是我的高光决定")
+        self.assertEqual(draft["segments"][0]["visual"], "咖啡厅办公把玩机身")
+
     def test_canonical_flat_schema_still_works(self):
         """契约内的扁平 schema(segments/narration/visual)不受兼容改动影响。"""
         raw = json.dumps({