test: harden baseline_cache guard to per-call unit_id check

tests/unit/test_no_dead_perquestion_paths.py

@@ -10,6 +10,7 @@ app.harness.validate（gate 块核心残留点）。
 """
 
 import inspect
+import re
 
 from app.harness import batching, inference, pools
 
@@ -26,9 +27,20 @@ def test_no_parallel_perquestion_split_helpers():
     )
     # 旧逐题分桶残留：正确率不应按 qid 直接取
     assert "correctness.get(qid)" not in src, "batching 仍有逐题分桶残留"
-    # gate 块残留探测：validate.py 是逐题 gate 核心残留点，基线缓存键须含 unit_id
+
-    assert "baseline_cache.get(" not in src or "unit_id" in src, "validate baseline_cache 仍按 qid"
+    # gate 块残留探测：validate.py 是逐题 gate 核心残留点。
+    # 逐个 baseline_cache.get(...) 调用点都必须按 unit_id 键控、不得回退到裸 qid/question_id，
+    # 否则视为逐题残留复活。跨行实参用容纳换行的懒惰匹配抓全整个调用文本。
+    baseline_get_calls = re.findall(r"baseline_cache\.get\((?:[^()]|\n)*?\)", src)
+    # 兜底：正则必须真抓到 validate.py 的调用，否则 for 空转又成 tautology
+    assert baseline_get_calls, "守卫失效：未抓到任何 baseline_cache.get 调用"
+    for call in baseline_get_calls:
+        assert "unit_id" in call, f"baseline_cache.get 未按 unit_id 键控（逐题残留）: {call}"
+        assert "qid" not in call and "question_id" not in call, (
+            f"baseline_cache.get 回退逐题 qid: {call}"
+        )
+
     # 逐题累加残留：n_used 不应按逐题 chunk 长度累加（应按 unit）
     assert "n_used += len(chunk)" not in src, "validate n_used 仍逐题累加（应按 unit）"
     # 正向 unit 化探测：pools/batching/validate/inference 已走 unit 契约
-    assert src.count("build_units") >= 1 or "unit_id" in src, "pools/batching/validate 未走 unit 化"
+    assert src.count("build_units") >= 1, "pools/batching/validate/inference 未走 unit 化聚合"