Learned replacement values for cached states designed to suppress the influence of erased spans without full recomputation.