Changes in version 0.1.0 (2026-10-18)                  

  - Added selectable ewma, arima, arfima, and structural baseline
    engines. ARIMA, ARFIMA, and structural specifications can be
    selected on the initial training window and are refitted causally at
    rolling origins.

  - Replaced duplicated output aliases with one compact structure.
    Fitted objects now expose only forecast, evaluation, calibration,
    model, and metadata; every forecast model owns its summary, samples,
    and pred_fun. Predictions return forecast plus metadata.

  - Made NAMESPACE reproducible with roxygen export/import tags,
    preserving hand-written help pages; renamed unused legacy helpers to
    avoid false S3 method declarations and added registration tests.

  - Isolated test fixtures inside each test so restored workspace
    objects such as an older fit cannot override fixtures during
    devtools::test().

  - Expanded marginal-function tests across all models and horizons;
    separated Windows source-loader limitations from installed-package
    parallel tests. CI now installs test dependencies and enables the
    parallel comparison.

  - Grouped predictive functions in pred_fun with explicit calibrated,
    baseline, and uncalibrated groups on fits and prediction results.

  - Unified fitted and updated forecasts with named model outputs and
    explicit scale metadata. Corrected log-CDF boundary values and
    single-horizon baseline matrix dimensions.

  - Added testthat coverage for the output schema and prediction
    behavior.

  - Added the direct-level engine with EWMA state tracking, direct
    multi-step ELM forecasts, spectral calibration, optional log
    transformation, and named horizon marginal functions.

  - Initial package release of local latent flow forecasting.

  - Registered print, predict, and plot methods for flowz.

  - Complete reference documentation and an executable introductory
    vignette.

  - Regression tests for fitting, prediction, distributions, log-scale
    summaries, input validation, and serialization.

  - Parallel execution by default with at most two automatic workers;
    quiet, non-plotting defaults.

  - Corrected recursive prediction defaults and original-scale summary
    moments for log-transformed fits.

  - Expanded parameter and method help, runnable examples, and a
    practical vignette covering evaluation limits and computational
    settings.