test(eval): verify scoring recovers after benchmark import failure

tests/eval/test_bench_loader.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import importlib
+import json
 import sys
 from collections.abc import Generator
 from pathlib import Path
@@ -91,3 +92,29 @@ def test_unsupported_benchmark_is_rejected(loader: ModuleType) -> None:
 def test_missing_benchmark_file_is_rejected(loader: ModuleType) -> None:
     with pytest.raises(SystemExit, match="benchmark implementation not found"):
         loader.load_bench("gpqa")
+
+
+def test_scoring_can_retry_after_benchmark_import_failure(loader: ModuleType, tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    score = importlib.import_module("score")
+    dependency = "_deepinfer_score_retry_dependency"
+    monkeypatch.delitem(sys.modules, dependency, raising=False)
+    write_bench(loader, "gpqa", f"import {dependency}\nfrom gpqa.bench import *\n")
+    data = tmp_path / "questions.jsonl"
+    data.write_text(json.dumps({"task_id": "q1", "question": "Choose B.", "answer": "B"}) + "\n", encoding="utf-8")
+    output = tmp_path / "run"
+    output.mkdir()
+    (output / "meta.json").write_text(json.dumps({"bench": "gpqa", "data_file": str(data), "n_sample": 1}), encoding="utf-8")
+    (output / "raw_rank0.jsonl").write_text(
+        json.dumps({"task_id": "q1", "sample_idx": 0, "raw_text": "ANSWER: B"}) + "\n", encoding="utf-8"
+    )
+    with pytest.raises(ModuleNotFoundError, match=dependency):
+        score.score_run(str(output), judge="rule", verbose=False)
+    assert not (output / "summary.json").exists()
+    monkeypatch.setitem(sys.modules, dependency, ModuleType(dependency))
+    assert score.score_run(str(output), judge="rule", verbose=False) == 0
+    row = json.loads((output / "scored.jsonl").read_text(encoding="utf-8"))
+    summary = json.loads((output / "summary.json").read_text(encoding="utf-8"))
+    assert row["pred"] == "B"
+    assert row["correct"] is True
+    assert summary["coverage_complete"] is True
+    assert summary["num_rollouts"] == 1