========================================================================
单路径消融聚合：suf_only / nec_only / gated_full（30集×5种子）
========================================================================

## 全局指标（每集先跨5种子取均值，再跨30集宏平均）

Variant                  Acc mean   AUC mean
Sufficient-only            0.7880     0.7700
Necessary-only             0.7813     0.7665
Gated full (BLN)           0.8653     0.8806

## 配对 Wilcoxon 符号秩检验（跨30集配对，双尾）

ACC Gated full (BLN)   vs Sufficient-only   : Δ=+7.73pp  win/loss/tie=30/0/0  p=0.00000
ACC Gated full (BLN)   vs Necessary-only    : Δ=+8.41pp  win/loss/tie=30/0/0  p=0.00000
ACC Sufficient-only    vs Necessary-only    : Δ=+0.68pp  win/loss/tie=20/10/0  p=0.10040
AUC Gated full (BLN)   vs Sufficient-only   : Δ=+11.06pp  win/loss/tie=30/0/0  p=0.00000
AUC Gated full (BLN)   vs Necessary-only    : Δ=+11.41pp  win/loss/tie=30/0/0  p=0.00000
AUC Sufficient-only    vs Necessary-only    : Δ=+0.35pp  win/loss/tie=14/15/1  p=0.30437

## 门控权重分布

alpha_s: mean=0.913 median=0.945 std=0.081 min=0.610 max=0.999
alpha_n (=1-alpha_s): mean=0.087 median=0.055

per-dataset table written: ablation_singlepath_table.csv