Strange behaviour of Regression fit when lifetime data order changes

[21ch216]

This affects the branch refactoring

Here is a sample code to reproduce the issue :

time, event, entry, *covar = load_insulator_string()
covar = np.column_stack(covar)
ph = ProportionalHazard(Gompertz()).fit(time, event, entry, model_args=(covar,))

Then if in utils.data, IndexedData.union is like :

def union(*others: Self) -> Self:
    return IndexedData(
        np.concatenate(
            [other.values for other in others],
            axis=0,
        ),
        np.concatenate([other.index for other in others]),
    )

the estimation of the parameters fails, the log-likelihood diverges in the optimization steps. After some inspections, I figured out that the values stored in lifetime_data.rc where the same than those in data._time.D_RC but in a different order. Thus, after reordering index and values created in union below, the error has disappeared. Here is the union working fine :

def union(*others: Self) -> Self:
    values = np.concatenate(
        [other.values for other in others],
        axis=0,
    )
    index = np.concatenate([other.index for other in others])
    sort_ind = np.argsort(
        index
    )
    return IndexedData(values[sort_ind], index[sort_ind])

why does the order of values matter ? is it due to scipy operations ? is the fault of np.concatenate ?