NEWS

AgriFusionR 0.1.1

Fixed the stacked ensemble's default base learners, which listed glmnet, the package name, where the registered learner is enet. The name was dropped silently, so stack ran without the elastic net and nothing said so. An unregistered base learner now warns instead of disappearing.
Added a demo covering the whole package: demo("AgriFusionR").

Eight more learners: glm, knn, xgboost, cubist, enet, svm, gam, and stack, a stacked ensemble weighted by non-negative least squares, with stack_weights() to report the contributions. Ten in total.
Learners are tested for signal recovery and row alignment, so an adapter that runs but predicts nonsense fails rather than passing quietly.
Three more explanation methods: accumulated local effects (Apley and Zhu 2020), ICE curves, and exact tree SHAP via treeshap.
Five remote data sources: CHIRPS, Daymet, WorldClim, SoilGrids, and SRTM elevation with slope and aspect. CHIRPS defaults to the ClimateSERV backend, because the CHC Cloud Optimized GeoTIFF path failed with GDAL tile-read errors on point queries.
Visualisation layer in base graphics, so it adds no dependency: plot() methods for agri_project, agri_resample and agri_model, plus plot_map() (prediction, residual and error surfaces), plot_effect() (ALE, PDP, ICE) and plot_uncertainty() (interval coverage). Colour scales come from hcl.colors() and are safe for colour vision deficiency.
219 tests; R CMD check clean.

Pre-release development.

agri_project() keyed on management unit and season, with role detection, a duplicate-key validator and a provenance ledger.
Source and learner registries (register_source(), register_learner()) as the extension mechanism.
growing_degree_days() and phenology_windows() for thermal-time staging; indicative thermal parameters for six crops in crop_parameters().
build_features() aggregating daily layers over phenological stages, with calendar and whole-season alternatives for comparison, plus heat, frost and dry-spell counters.
check_project() including a leakage guard that refuses covariate windows reaching past the harvest they predict.
resample_scheme() with spatial blocking, leave-location-out, forward-season and buffered variants; spatial blocking is the default.
train_model() reporting spatial and random cross-validation side by side so the optimism of random folds is quantified.
uncertainty() giving split-conformal prediction intervals with a non-circular coverage estimate.
explain() with out-of-fold permutation importance and partial dependence.
report() writing a model card whose limitations section is generated from the model's own diagnostics.
Offline, deterministic demonstration system with a known data-generating process; 111 tests.