I found the bookmark-switching regression on the train. dsop beat me to the fix, but I thought this test would be nice since without this test we'll miss it if the feature regresses in the future.