上线版: 产品表单统一+form嵌套修复+用户管理+部署+三套叙事

- 产品编辑入口统一走 ProductFormFull(卖点/风格/人群/品牌词全字段);
  修复开任务页 <form> 套 <form> 致"编辑产品"报错、改不了、跳回首个产品
- dashboard 入口卡片对齐实际路由: 系统管理(/config) 与 工作配置(/settings) 分开;
  settings ?tab=products 直达改用挂载后读 URL, 消除 hydration mismatch
- 新增用户管理(users API/admin service/改密页) + alembic 022/023/024
- 上线部署: Dockerfile / docker-compose.prod+https / nginx https / .env.example
- A8 三套正交叙事(痛点/场景/成分背书) + beige 调色去AI化 + 飞轮 text_import 高权重信号

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

backend/app/workers/replenish_task.py

@@ -34,6 +34,7 @@ def replenish_text_candidates(self, task_id: int) -> dict:
     from app.workers.pipeline_steps import (
         load_task_and_product,
         decrypt_user_key,
+        decrypt_codeproxy_key,
         build_clients_and_clear_key,
         build_product_dict,
         load_flywheel_context,
@@ -57,8 +58,10 @@ def replenish_text_candidates(self, task_id: int) -> dict:
 
         workspace_id = task.workspace_id
         plain_key = decrypt_user_key(db, task.operator_id, workspace_id)
-        clients = build_clients_and_clear_key(plain_key)
+        alt_key = decrypt_codeproxy_key(db, task.operator_id, workspace_id)
+        clients = build_clients_and_clear_key(plain_key, alt_key)
         plain_key = None  # 用完即清除，基石B
+        alt_key = None
         product_dict = build_product_dict(product)
         flywheel_fragment, _ = load_flywheel_context(
             db, workspace_id, task.product_id, product_dict
@@ -71,7 +74,7 @@ def replenish_text_candidates(self, task_id: int) -> dict:
         current = existing
         while current < target and rounds < MAX_REPLENISH_ROUNDS:
             rounds += 1
-            run_text_generation(
+            _, _, _, fail_reason, _ = run_text_generation(
                 db, clients, task, product_dict, flywheel_fragment,
                 _push_event_sync, workspace_id, seq,
             )
@@ -85,6 +88,11 @@ def replenish_text_candidates(self, task_id: int) -> dict:
                 "补充轮 %d: task_id=%s 库内合格=%d/%d",
                 rounds, task_id, current, target,
             )
+            # 评分通道不可用时再补也是空烧 token(每轮还各 35s backoff)——立即停,
+            # 等通道恢复用户手动重试。区别于"质量不达标"(那个值得多补几轮)。
+            if fail_reason == "scoring_unavailable":
+                logger.error("补充中止: task_id=%s 评分通道不可用,停止空跑重试", task_id)
+                break
 
         return {
             "task_id": task_id,