Incorrect control over false positives #518

I've been testing the bootstrap_isc and I'm puzzled by the false-positive rate (FPR). I made null data with only standard normal noise for each subject and voxel.
With subject-wise sampling the FPR is lower than expected at around 0.02
With element-wise sampling, the FPR is around 0.09 (inflated, as expected from Chen et al, 2016).

SW false positive rate (alpha=0.05) 0.02
EW false positive rate (alpha=0.05) 0.09

Code is here

Am I doing something wrong?