The difference in bad channels, does explain it.
They must have the same bad channels, the fact that it works for the good folder is only chance.
This is the long explanation: one difference between the good and the bad folders is that in the first, one SSP component is removed, and in the second no component is removed when doing EOG artifact removal. When computing the inverse model the noise covariance is loaded, and as expected, the rows and columns for the bad channels (PO3 and POz) are zeros. Then, when computing the inverse model, the existing projectors (in the channel file) are applied to the head model and to the noise covariance. In the case of the "good" folder, the SSP projector was computed with POz (channel 30) as good channel, then when this projector is applied on the noise covariance, it removes the zeros on the row and column (30), as consequence the following steps work.