A reaction to autoflight disengagement perhaps? In all honesty I don't know, but the deviations around the normal prior to manual input still seem to be less than those encountered on the actual trace - it's just that because the AoA ended up at such an extreme value, the vertical element of the graph is compressed in it's early stages.