Changes in version 0.1.3 (2026-09-16)                  

  - Explicitly initialize conditional outcome pointers for strict
    compiler diagnostics.

  - Corrected survival preprocessing to log positive event/censoring
    times once, matching the original supplementary C code. Survival
    fits must be rerun; survival_scale = "identity" explicitly
    reproduces the historical raw-time implementation. The CV
    partitioning and selection sampler order are unchanged.
  - Added optional posterior_draws() with subgroup coefficient intervals
    and posterior predictive intervals. All active coefficients,
    including intercepts and clinical effects, use the original pMOM
    prior. Clinical effects remain always included; automatic clinical
    variable selection is not implemented.
  - Conditional sampling reaugments binary and censored responses and
    reports classical split R-hat. Model averaging retains the fitted
    selection sampler's Laplace approximation; it is not an exact
    model-weight calculation.
  - Survival posterior prediction uses time-scale medians as point
    summaries because inverse-gamma variance mixtures need not have
    finite time-scale means.
  - Added independent analytic/integration checks and survival-scale
    regressions.

                        Changes in version 0.1.2                        

Native-code reliability

  - Fixed premature unprotection of R return objects in the fitting and
    prediction entry points.
  - Fixed an allocation leak in binary cross-validation prediction.
  - Initialized cross-validation work arrays and moved input-sized
    buffers off the C stack.
  - Added dedicated valgrind CI and macOS ARM vignette checks under
    ASAN/UBSAN.

User interface

  - Added the validated imr_data class for training and prediction
    inputs.
  - Added an imr() formula/data method while preserving the original
    interface.
  - Formula predictions reuse training transformations, factor levels
    and contrasts; subject IDs are excluded from dot-formula expansion.
  - Unsupported no-intercept and offset formulas now fail explicitly
    instead of silently fitting a different model.
  - Added fitted-object validation, fit comparison, posterior
    uncertainty summaries, credible intervals and parameter-specific
    trace plots.
  - predict.imr() now returns values at full numeric precision.

Reproducibility

  - Expanded the standalone replication script to execute every R
    command shown in the manuscript and to fail when full-run reference
    results drift.