This is a known issue (internal ticket number AP-21471). KNIME expects every data type to indicate its capabilities and requirements via Java interfaces. The Process Mining extension PM4KNIME doesn’t do that, so it looks like (for example) the PetriNetCell data type is compatible with any and all values. When the column type for a mixed column like column1 above is computed, the PetriNetCell type looks like a perfect fit, so it it picked.
While the cell type definition is unconventional, our code should not be as easily confused by extensions. We already had the ticket in our backlog, but your report has bumped up its urgency. Thanks for your report!