NEWS

n4m 1.0.21.9006

Added native dispatch for nine sample splitters and 22 seeded, train-only X-to-X augmentations. The augmentation contract preserves sample order and feature width; it does not transform targets or serialize fitted state.
This development package carries libn4m ABI 2.11. The pls4all R subset is also versioned 1.0.21.9006 so R-universe does not reuse an ABI-2.10 package version for a newly built binary.

Affine MethodResult regressors can be promoted by the C++ core to a predict-only, N4MM-serializable model. Sixteen verified methods use this common path in the Python and R estimator interfaces; the JavaScript/WASM binding can replay the native model.
R formula estimators use native prediction for qualified affine results and retain N4MM bytes for saveRDS()/readRDS() across processes.
R and Python expose the same generic native preprocessing pipeline for the fifteen supported operator kinds. Fitted linear-chain state can be exchanged as versioned N4MP bytes, with native plan inspection and strict decoding. Branching, variable selection and product-level trained pipeline archives are not implied by this low-level format.

group_sparse_pls now applies groupwise proximal shrinkage to the coefficients used for prediction. Earlier builds accepted group_lambda but left predictions unchanged.
fused_sparse_pls consequently applies l1_lambda to predictive coefficients before its existing fusion smoothing. Predictions with a positive penalty can differ from 1.0.21.9003.
These are documented post-SIMPLS approximations, not the sgPLS::gPLS estimator or a full fused-lasso optimization.