Changes in version 0.3.0 (2026-09-21)                  

  - All three component links (cloglog, logit, and log) now support EM,
    direct Marquardt Newton-Raphson, and hybrid fitting.
  - Added analytic observed-likelihood derivatives for logit and log.
    The hybrid uses the chosen link throughout, and final inference uses
    its unmodified observed Hessian. Log-link proposals outside the
    probability domain are rejected by backtracking. No new dependencies
    or public controls were added.
  - Scaled the direct line-search rounding allowance to the summed
    likelihood to prevent false failures near a root on large samples.
  - Expanded derivative, EM/direct/hybrid agreement, covariance,
    rare-probability, and domain-backtracking tests; updated examples
    and algorithm documentation.
  - Version 0.2.0 Win-builder results do not validate this extension.
    Repeat R-release and R-devel checks on the 0.3.0 archive before
    submission.

                        Changes in version 0.2.0                        

  - Added direct observed-likelihood Marquardt Newton-Raphson and
    EM-initialized hybrid fitting through exposure_fit(). Pure EM
    remains available through exposure_em(); no ES algorithm is
    included.
  - Verified all three EM component links, their full Newton M-steps,
    Louis information, and the cloglog observed derivatives against
    independent numerical differentiation and likelihood optimization.
  - EM termination now checks observed likelihood change and the
    unmodified score-curvature measure as well as the selected change
    criterion. Direct termination consistently applies the three change
    and curvature checks. Final observed scores and Newton decrements
    are returned for all methods.
  - Reject indefinite covariance information and constant cloglog
    exposure proportions. Validate integer limits, named controls,
    logical settings, finite covariates, response shape, and unsupported
    formula offsets.
  - Stabilized structural-zero exposure calculations and rare-event
    direct derivatives. The EM Q ascent check now allows rounding error
    proportional to objective size; tiny M-step changes also require a
    small score.
  - Fixed missing-exposure prediction for logit, domain checking with
    mixed missing and invalid predictions, and na.exclude restoration in
    standard prediction, fitted-value, and residual methods.
  - Hybrid warm-up is documented as an upper iteration limit and
    respects each phase's trace setting. Documentation now explains
    inference assumptions, convergence limitations, returned fields, and
    the absence of a proven hybrid runtime advantage.
  - Added self-contained numerical safeguard tests. The restricted-data
    ABCD comparison remains in the local source tree but is excluded
    from source archives; the installed smoke-test example is portable.