We study order selection for a finite-order stationary ARFIMA process by minimizing a nested folded-concave penalization of the integrated Toeplitz profile Whittle criterion. The AR and MA orders are encoded through suffix groups, so that exact recovery of the zero tails is equivalent to exact recovery of the orders. The analysis is carried out on a weighted sieve neighborhood whose geometric weights keep growing-lag perturbations from pushing AR or MA roots toward the unit circle. Under finite true orders, coprime root-separated short-memory polynomials, an interior stationary memory parameter, growing sieve orders, suitable tuning, and independent sub-Gaussian innovations, we prove existence of an oracle local minimizer, rather than global optimality of the nonconvex criterion. This local minimizer selects the true AR and MA orders with probability tending to one, sets all inactive coefficients exactly to zero, and achieves the active-coordinate rate \( \Op(L_n n^{-1/2}) \), where the stochastic scale \(L_n\) is governed by the local mismatch \(Δ\). The proof combines weighted root-separation, matched finite-section Toeplitz control, finite-section trace-bias control, a sub-Gaussian quadratic-form argument for the stationary observed vector, and a local domination step for folded-concave penalties.