Unlike ComputeBounds, this method is used to calculate the expected size of this node when collapsed, so that Collapse can call CollapseChild and FinishCollapse with the intended collapsed bounds. The size of each child (that is not skipped by ComputeCollapsedSizeSkip) is given by the GoObject.SelectionObject, except for nested subgraphs, where ComputeCollapsedSize is called recursively. This method does not augment the size to include the CollapsedTopLeftMargin or CollapsedBottomRightMargin.
Note that the standard behavior of this method makes sure the resulting size is wide enough to hold each child node and tall enough to hold each child node--it is not just the size of any CollapsedObject. The reason is that the standard behavior wants to insure that all links connected to child nodes will end at or in the collapsed subgraph, not potentially far away from the subgraph.