Before

# LogNormal currently falls back to the generic helper:
# logccdf(dist, value) internally computes log1mexp(logcdf(dist, value))
import pymc as pm
from pymc.logprob.basic import logccdf

dist = pm.LogNormal.dist(mu=0, sigma=1)
logccdf(dist, 2.0)  # goes through _logccdf_helper -> log1mexp(normal_lcdf(...))

After

# LogNormal uses direct closed-form log survival function:
# normal_lccdf(mu, sigma, log(t))
import pymc as pm
from pymc.logprob.basic import logccdf

dist = pm.LogNormal.dist(mu=0, sigma=1)
logccdf(dist, 2.0)  # directly calls normal_lccdf(mu, sigma, log(t))

Context for the issue:

LogNormal currently falls back to the generic _logccdf_helper which computes log1mexp(logcdf). While log1mexp is numerically stable (Mächler two-branch formula), LogNormal has a direct closed-form log survival function:

log S(t) = normal_lccdf(mu, sigma, log(t))

This reuses the existing normal_lccdf from dist_math.py, consistent with how Normal.logccdf already works. It is a genuinely different function call, not just the logcdf intermediate minus a log1mexp wrapper.

Context

Refer to the related discussion on Discourse:
https://discourse.pymc.io/t/consider-adding-numerically-stable-logccdf-log-survival-function-to-exponential-weibull-and-other-distro/17603

The log-likelihood for survival models with right-censored observations includes a log S(t_i) term for every censored observation. pymc/logprob/censoring.py calls _logccdf_helper for this.