裂变补图：中转站波动导致单张失败时只重生失败分镜

实测发现 codeproxy 偶发不稳，单套中某张图重试3次仍失败会记 error。
本次加按需补图：只重生标 error 的分镜，成功的图与 token 不浪费。

- 019迁移+模型：fission_tasks 加 source_task_id(新架构不扇出GenerationTask,
  补图需复刻首轮上下文:产品/operator key/张数,故落库回查)
- fission_image_retry：只补失败 role(引擎 target_role 单张重生,R2零改动),
  merge 回 images_json,全恢复则 status=image_done
- tasks.py：retry_fission_images Celery task,key归属源任务operator(基石B),
  幂等锁防重复补图
- api：POST /fission/{id}/notes/{seq}/retry-images,无失败分镜守卫 queued=false
- 前端：FissionNoteCard 失败提示改"重试补图"按钮+loading,父页轮询刷新

实测task5 seq1(hook图首轮codeproxy重试3次失败)补图跑通：只补hook(1024×1536
去AI化)、applied_proof原图不动、recovered=1 still_failed=0、status=image_done;
apiports503回落codeproxy gpt-5.5强档兜底;seq0无失败图守卫queued=false

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

backend/app/workers/tasks.py

@@ -286,3 +286,72 @@ def run_fission_pipeline(self, fission_id: int, source_task_id: int) -> dict:
     finally:
         _r.delete(lock_key)
         db.close()
+
+
+@celery_app.task(
+    bind=True,
+    name="app.workers.tasks.retry_fission_images",
+    max_retries=1,
+    default_retry_delay=20,
+    queue="generation",
+)
+def retry_fission_images(self, fission_id: int, note_id: int) -> dict:
+    """裂变补图：对某一套只重生标了 error 的失败分镜（中转站波动兜底）。
+
+    只接收 id（基石B），不接收 key。复用源产品参考图与 image_count。
+    解密 key→构建 clients→委托 fission_image_retry.retry_fission_note_images。
+    """
+    logger.info("retry_fission_images start: fission_id=%s note_id=%s", fission_id, note_id)
+    lock_key = f"fission:retry:lock:{note_id}"
+    import redis as _redis
+    from app.core.config import get_settings as _gs
+    _r = _redis.from_url(_gs().REDIS_URL, decode_responses=True)
+    if not _r.set(lock_key, "1", nx=True, ex=1800):
+        logger.warning("retry_fission_images 跳过(已在跑): note_id=%s", note_id)
+        return {"note_id": note_id, "status": "skipped_locked"}
+
+    db = _get_db()
+    try:
+        from app.models.fission import FissionTask
+        from app.models.task import GenerationTask
+        from app.models.product import Product
+        from app.workers.pipeline_steps import (
+            decrypt_user_key, build_clients_and_clear_key, build_product_dict,
+        )
+        from app.services.fission_image_retry import retry_fission_note_images
+
+        ft = db.query(FissionTask).filter(FissionTask.id == fission_id).first()
+        if not ft:
+            return {"note_id": note_id, "status": "not_found"}
+        if not ft.source_task_id:
+            return {"note_id": note_id, "status": "no_source_task"}
+        src = db.query(GenerationTask).filter(
+            GenerationTask.id == ft.source_task_id,
+        ).first()
+        if not src:
+            return {"note_id": note_id, "status": "source_task_gone"}
+        product_row = db.query(Product).filter(Product.id == src.product_id).first()
+        if not product_row:
+            return {"note_id": note_id, "status": "no_product"}
+
+        # key 归属源任务 operator，与首轮生图一致（基石B：不接收明文 key）
+        plain_key = decrypt_user_key(db, src.operator_id, ft.workspace_id)
+        clients = build_clients_and_clear_key(plain_key)
+        plain_key = None  # 用完即清，基石B
+
+        product = build_product_dict(product_row)
+        image_count = max(1, (src.image_count or 3))
+        result = retry_fission_note_images(
+            db, clients, ft, product, image_count, note_id,
+        )
+        result["status"] = "completed"
+        return result
+    except Exception as exc:  # noqa: BLE001
+        logger.error("retry_fission_images failed: note_id=%s err=%s", note_id, exc)
+        exhausted = self.request.retries >= self.max_retries
+        if exhausted:
+            return {"note_id": note_id, "status": "failed", "error": str(exc)}
+        raise self.retry(exc=exc)
+    finally:
+        _r.delete(lock_key)
+        db.close()