=== DO difficulty LLM classification ===
valid: 184

--- overall distribution ---
   28  scalability
   23  reasoning_limitation
   17  generalization
   16  domain_knowledge
   13  semantic_correctness
   12  no_difficulty_extracted
   11  cost_or_efficiency
   11  sim_to_real_gap
   10  lack_or_limited_evaluation
    8  data_scarcity
    8  other
    7  expert_dependency
    7  hallucination
    5  long_horizon
    5  feedback_quality
    3  structural_change
