This is not a bug.
The explanation is given in the (annotated) paragraph:
For these comparisons, it is useful to use zero as a reference value (ref_val); providing this reference value [i.e. 0] yields cumulative probabilities [as in
cdf] for the posterior distribution on either side of the value [i.e. vertical line at ref_val=0]. Thus, for the difference in means, 99.4% of the posterior probability is greater than zero…
In other words, 99.4% == values to the right of vertical line - values to the left.