The conclusion of the conversation represented by #300 (there and in slack) is that storing type as an attribute of these data frames is too brittle. The row subsetting and binding those data frames go through makes it hard to retain the attributes correctly. I'm closing that PR in deference to some larger rethinking of how type gets passed around.

Probably also what's going on in #291. That also seems like a case where the specific combination of packages to be installed from various sources was unlucky.