So I have the improved Howland current pump experiment set up and working in LTSpice, and am experimenting using two pumps to feed bandgap references.
One of them is using an NPN transistor to source current into the reference with the other end grounded, while the other is used to sink current from the reference from ground into a PNP connected to Vee.
Interstingly, the NPN version seems stable on its own, while the PNP version oscillates like crazy without 100p or so of capacitance from the PNP emitter to non inverting input.