RPP data: one original T.r.O is 1.03 (chi-square converted), arctanh gives NaN

In CenterForOpenScience/rpp RPPdataConverted.tsv, T.r.O is 1.0297 for the Anggoro, Waxman & Medin original. Its test statistic is X^2(1, N=48) = 50.89, so sqrt(X^2/N) exceeds 1. np.arctanh returns NaN with only a RuntimeWarning, and a likelihood summed over all originals becomes NaN, which makes Nelder-Mead quietly run to maxiter. Another original (Malt et al.) has r = .99 exactly (z = 2.65). Both rows have no replication r (T.r.R is NaN), so analyses restricted to the 97 complete original/replication pairs are unaffected. Fits over all 131 originals with p < .05 are not. Filter |r| < 1 explicitly, and check the likelihood for NaN before handing it to an optimiser.