Changes in version 0.2.0 (2026-06-30)                  

  - Added fit_uddbart_vi() for a lightweight Gaussian variational
    approximation to the unified discrete-time hazard model.
  - Added engine = "vi" to uddbart() for a BART-free UDDBART-VI
    workflow.
  - Made the BART package optional. fit_bart_hazard() now falls back to
    stats::glm(family = binomial) when BART is unavailable.
  - Updated dynamic prediction internals to work with BART, glm
    fallback, and VI hazard backends.
  - Added BART-free tests for the glm fallback and UDDBART-VI workflow.

                        Changes in version 0.1.0                        

  - First release of the U-DDBART-IC framework (version 0.1, engineered
    latent-state features).
  - Pipeline: simulate_uddbart_data(), prepare_uddbart_data(),
    make_followup_grid(), construct_latent_state(), fit_bart_hazard()
    (logit-link BART), compute_survival_from_hazard(),
    interval_survival_likelihood(), uddbart(), and predict.uddbart() for
    the dynamic target pi(t, tau) = 1 - S(t+tau)/S(t).
  - Predictive-performance evaluation with IPCW: brier_score()
    (time-dependent and integrated Brier score), auc_td()
    (cumulative/dynamic time-dependent AUC) and calibration_table().
  - print, summary and plot methods.
  - Bundled anonymised CML molecular-monitoring cohort (cml_long,
    cml_event).
  - Functions that take a seed restore the caller's RNG state on exit.
  - An R Journal manuscript and reproducible replication script are
    provided in paper/ (excluded from the package build).