Mirror chain trimmed to 6 mainstream sources; manual-load guidance

16 -> 6 (daemon default, daocloud, 1ms.run, 1panel, rat.dev, ustc):
the long tail never delivered anyway. All-6 failure now prints an
explicit self-load recipe (save/load via an egress machine, or pull
when the network recovers) instead of a terse RuntimeError.

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

evalharness/sandbox/docker.py

@@ -45,11 +45,13 @@ def ensure_image(img: str) -> None:
         return
     except RuntimeError as e:
         raise RuntimeError(
-            f'sandbox image {img!r} is not available: not local, and every '
-            f'mirror failed. {str(e)[:200]}. '
-            'Fix: pull/build it manually (for bigcodebench the official image '
-            'is bigcodebench/bigcodebench-evaluate:latest), then re-run with '
-            '--rescore to score the cached predictions.') from e
+            f'沙箱镜像 {img!r} 不可用：本地没有，全部镜像源尝试失败。'
+            f'{str(e)[:150]}\n'
+            '需要自己加载镜像（任选其一），然后重跑加 --rescore：\n'
+            f'  A. 有外网的机器:  docker pull {img} && docker save {img} | gzip > img.tgz\n'
+            f'     拷回本机后:    docker load < img.tgz\n'
+            f'  B. 本机网络恢复后: docker pull {img}\n'
+            '预测都在 checkpoint 里，镜像就位后纯判分即可。') from e
 
 
 @register_sandbox('docker')

evalharness/sandbox/prefetch.py

@@ -19,28 +19,18 @@ from ..data.dataset import Dataset
 # CN mirrors tried in order before/alongside the daemon's configured mirrors.
 # Some namespaces (e.g. swebench/*) are blocked by individual CN mirrors, so we
 # fall through: daemon default -> 1ms.run -> baidubce -> sjtug.
-# CN + international mainstream fallbacks. History on this host: daocloud
-# (139s) and 1ms.run (342s + trickle) have delivered this 9GB image; the
-# international entries (Google's hub cache, AWS ECR public) need real
-# egress -- usually blocked here, kept for networks that have it.
+# Six mainstream fallbacks (trimmed from 16 per user call: the long tail
+# died anyway). History on this host: daocloud and 1ms.run have each
+# delivered the 9GB bigcodebench image; the rest are the usual suspects.
 _CN_MIRROR_FALLBACKS = [
     '{img}',  # daemon default (uses its own registry-mirrors config)
     'docker.m.daocloud.io/{img}',
     'docker.1ms.run/{img}',
     'docker.1panel.live/{img}',
     'hub.rat.dev/{img}',
-    'docker.xuanyuan.me/{img}',
-    'dockerhub.icu/{img}',
-    'dockerproxy.com/{img}',
-    'hub-mirror.c.163.com/{img}',
-    'mirror.baidubce.com/{img}',
     'docker.mirrors.ustc.edu.cn/{img}',
-    'docker.mirrors.sjtug.sjtu.edu.cn/{img}',
-    'docker.nju.edu.cn/{img}',
-    'docker.mirrors.tuna.tsinghua.edu.cn/{img}',
-    'mirror.gcr.io/{img}',       # Google's docker-hub cache (intl)
-    'public.ecr.aws/docker/{img}',  # AWS ECR public (library/* coverage)
 ]
+
 # 用户/环境可整体覆盖（逗号分隔模板，{img} 占位）
 import os as _os