TBH I suspect we could get away without using getBoundingClientRect at all here - and that would avoid some bugs we probably have right now in case the user has the graph inside some sort of transform.

But in the meantime, to be consistent in our usage, couldn't we switch from outerContainer: fullLayout._paperdiv to outerContainer: fullLayout._paper.node() like we have everywhere else this gets used - ie use _paper (the main <svg>, that always seems to have a meaningful size) rather than _paperdiv (its enclosing <div>, that for whatever weird CSS reason sometimes has no size)... also just pass around a node, don't d3.select it here only to .node() it here