I first want to applaud the author for a) doing methods-sanity checks and b) publishing the analysis code.
In the article you describe accuracy as:
"Tracker accuracy was calculated for each participant by the averaging of all horizontal
and vertical distances between validation target locations and the gaze samples obtained
when a participant was fixating these targets, excluding missing data. The resulting values
were averaged across all participants"
This implies a distance measure which is usually bounded to be positive. But looking at your code, if I understand everything correctly, it seems you subtracted sample-coordinate from target coordinate  (which can be positive and negative). (I might have missed an absolute value or square calculation somewhere, but could not find one. If I missed this please disregard my comment below.)
This is problematic because it allows for signed distance (too far to the right or left of target). Imagine you have only two calibration targets. The left one 1° too far to the left (thus -1°) the right one 1° too far to the right (thus +1°). In your calculation this would average to a mean error of 0°. But I think it should be a mean calibration error of 1°.
You could recalculate the accuracy numbers using a positive distance measurement. In principle horizontal + vertical angle could be combined using spherical-coordinates, but this is a personal preference.
Hope this comment helps you improving this work. If I missed or missunderstood something I'm happy to edit this comment and put it right. I think even though the eyetracker is not available anymore, this preprint is well-cited and often downloaded. I think its worth to update it if needed.
With kind regards.