Like bootstrap_W but NOAA sites are fitted with nonstationary
GEV (linear mu trend), extracting (mu0, log_sigma, xi) per replicate.
ADCIRC sites use the standard stationary GEV.
This is the primary bootstrap approach used in the manuscript.
Note
A small fraction of bootstrap resamples may produce degenerate data
that causes warnings or convergence failures in extRemes::fevd.
Failed fits are recorded in n_failures and excluded via
pairwise-complete covariance estimation (failure rates are typically
well below 1 percent).
See also
bootstrap_W for the stationary version.