Ablation #46 -- Is the 13-perturbation fragility score necessary?
n=412, target=llm_label (PRIMARY), pos_rate=0.614
5-fold stratified CV, seed=42, bootstrap 2000

config                 n_feat                  LogReg                      GB                CatBoost
A4_full                   105    0.7304 [0.678,0.779]    0.6947 [0.644,0.748]    0.6916 [0.641,0.744]
A1_no_pert                 50    0.7591 [0.711,0.807]    0.7417 [0.694,0.791]    0.7324 [0.684,0.782]
A2_pert_only               58    0.5566 [0.501,0.612]    0.5013 [0.445,0.555]    0.5292 [0.469,0.584]
A3_top3                     3    0.6157 [0.562,0.672]    0.6462 [0.593,0.701]    0.6463 [0.592,0.704]

-- Verdict ---------------------------------------------------------------
A4 LogReg (full, 105 feats)            = 0.7304
A1 LogReg (no-pert, ~50 feats)  = 0.7591
Δ (A4 - A1)                            = -0.0288

A3 best AUC (3 features only)          = 0.6463

Rule-of-thumb:
  |Δ| <= 0.02  -> 13 perturbations are COSMETIC (A1 matches A4)
  A3 >= 0.68   -> 3-feature model is COMPETITIVE with 105-feature ensemble
