Internal helper computing:
$$\mathrm{Var}(\hat{f}_k) = \frac{\sigma^2_k}{W_k}$$
where \(W_k = \sum_i w_{i,k} \cdot C_{i,k}^\alpha\). This is consistent
with the WLS weight \(w_{i,k} / C_{i,k}^{2-\alpha}\) used in .lm_ata()
and follows Mack (1993)'s distribution-free standard-error derivation
for the chain ladder reserve estimator.
Paradigm pairing: the package keeps two natural analytical variance
helpers, one per paradigm-target pair: .mack_f_var() (CL / Mack 1993
applied to f-factor) and .ed_g_var() (ED / Buehlmann-Straub 1970
applied to g-intensity). They share the underlying volume-weighted
variance idea (\(\sigma^2_g = \sigma^2_f\) via \(g_k = f_k - 1\)),
but operate on different Link columns (f reads loss_to/loss_from;
g reads loss_delta/premium_from) and produce differently-named
output columns (f_var / g_var), so are kept as separate helpers.
Also used by fit_ratio() for the CL component.