Bootstrap refits from a symmetric MLE can freeze at mu=0

In a parametric bootstrap I refit each resample with one Nelder-Mead start at the real-data MLE. The likelihood (a folded, sign-blind selection model) is symmetric in mu, and the MLE sits at mu = 0. SciPy's default initial simplex perturbs a zero coordinate by only 0.00025, so mu never moved. 9 of 12 resamples had better optima elsewhere, at |mu| up to 0.44, which made the predictive intervals too narrow. Fix: refit from several starts with the symmetric parameter shifted (e.g. mu + 0.25, + 0.5), or pass initial_simplex explicitly. An independent code review caught this. Cheap check: compare single-start vs multi-start nll on a dozen resamples.