bgev_mle() was rewritten. It now uses a multistart Nelder-Mead search on a reparametrised scale (log(sigma), log(delta)), seeded by quantile matching plus loose data-driven box starts. Estimation is restricted to delta > 0 (bimodality requires it, and for delta < 0 the likelihood is unbounded at x = mu); the distribution functions still accept the full delta > -1.
The return value changed: instead of the old DEoptim object ($optim$bestmem), bgev_mle() now returns $par, $se, $loglik and diagnostics. This is a breaking change.
Standard errors (se) are returned from the inverse observed-information Hessian, for an admissible (regular) optimum only; near the parameter-dependent support boundary they are not reliable and are returned as NA.
New diagnostics on every fit: convergence, agree, admissible (a positive-definite-Hessian acceptance gate that rejects spurious optima), boundary, and optimum.
likelihood = "grouped_likelihood" added for discrete or rounded data, using the interval likelihood F(x + h/2) - F(x - h/2); tied data under the continuous density now raises a warning.