feat(bench): HMMT/AIME small-subset harness + answer extraction tests

tests/test_bench_harness.py

@@ -0,0 +1,27 @@
+from scripts.bench_hmmt import extract_final_answer, score_subset
+
+
+def test_extract_final_answer_picks_last_boxed():
+    text = "Long reasoning... \\boxed{42} done."
+    assert extract_final_answer(text) == "42"
+
+
+def test_extract_final_answer_falls_back_to_last_number():
+    text = "...therefore the answer is 17."
+    assert extract_final_answer(text) == "17"
+
+
+def test_extract_final_answer_returns_empty_when_no_number():
+    assert extract_final_answer("no answer here") == ""
+
+
+def test_score_subset_counts_correct():
+    items = [
+        {"question": "q1", "answer": "42"},
+        {"question": "q2", "answer": "100"},
+    ]
+    predictions = ["The answer is 42.", "Final: 99"]
+    score = score_subset(items, predictions)
+    assert score.correct == 1
+    assert score.total == 2
+    assert abs(score.accuracy - 0.5) < 1e-6