It appears to me that the flux induced in the core from the other two – still active – phases which obviously pass through the core section of the “inactive” phase are enough to generate an almost normal voltage in both the primary and secondary coils of that phase. This could be because the flux density of the VT is reasonably low to give an accurate transformation ratio over its entire range. The problem may be reduced by using a heavier load on the secondary so that the higher current will show more “droop” when the voltage fails. However this may mean less sensitivity to abnormally high voltage. I must admit that we usually supply 3 separate single phase transformers for this type of application, each connected between phase and neutral thus eliminating the possibility of phase interaction.