feat(harness): add Action Recognition training experiment

- PerCategoryPoolStrategy: filter test pool by task_types
- RunConfig: add run_holdout_eval toggle (default true)
- load_config: fix YAML task_types list-to-tuple conversion
- Runner: conditionally skip _holdout_four_way when disabled
- CLI: add --no-run-holdout-eval flag
- New config/train_action_recognition.yaml (3 epochs, per_category)
- New scripts/train_action_recognition.sh (baseline + seed + train)

app/harness/pools.py

@@ -205,6 +205,7 @@ def _q_to_dict(q: GeneratedQuestion) -> dict:
         "answer": q.answer,
         "source_nodes": list(q.source_nodes),
         "difficulty": q.difficulty,
+        "family": q.family,
         "skill_target": q.skill_target,
         "difficulty_steps": q.difficulty_steps,
     }
@@ -228,6 +229,7 @@ def _dict_to_q(d: dict) -> GeneratedQuestion:
         answer=d["answer"],
         source_nodes=tuple(d.get("source_nodes", ())),
         difficulty=d.get("difficulty", "medium"),
+        family=d.get("family"),
         skill_target=d.get("skill_target"),
         difficulty_steps=d.get("difficulty_steps"),
     )
@@ -595,12 +597,15 @@ class PerCategoryPoolStrategy:
             all_train.extend(train)
             all_val.extend(val)
 
-        # Phase 4: test 池（从外部目录加载，无则空）
+        # Phase 4: test 池（从外部目录加载，无则空；按 task_types 过滤）
         test: list[GeneratedQuestion] = []
         if config.test_questions_dir is not None:
             from app.question_gen import load_benchmark
 
             test = load_benchmark(config.test_questions_dir)
+            if config.task_types is not None:
+                allowed = set(config.task_types)
+                test = [q for q in test if q.task_type in allowed]
 
         # Phase 5: 计算 baseline_val_accuracy
         val_correct = sum(1 for q in all_val if correctness.get(q.question_id, False))