Learning Surrogate Indices from Historical A/Bs: Adversarial ML for Debiased Inference on Functionals of Ill-Posed Inverses