NEWS

PhysioML 0.4.2

A vignette carries one task end to end on synthetic or bundled data, offline, and is built and run by R CMD check.
Runnable @examples added or corrected across 20 help pages. Each runs offline in seconds, writes nothing outside tempdir(), and is executed by R CMD check; anything needing a device, a download or an optional backend is fenced with the reason stated.
The README's quick start runs as written: it attaches the package, builds its own inputs, and uses only hard dependencies.

Made the foundation-cache symlink-rejection test portable: the test now skips when the build filesystem cannot create a symlink (e.g. the Windows r-universe binary builder), where file.symlink() fails silently so no unsafe condition exists to assert on. The security assertion itself is unchanged and still runs wherever symlinks work.

Adds governed R torch to TorchScript to Python PyTorch ONNX export with canonical hash-bound model cards and pre-landing checker/logit parity gates.
Adds persistent raw-payload ONNX import and CPU-only ONNX Runtime prediction with provider fallback disabled and exact Dataset contract validation.

Adds peDataset() for exact case-aware windowing, zero-based class targets, and train-only z-score or robust normalization.
Adds compact EEGNet, CNN1D, causal TCN, and LSTM modules through the optional R torch backend.
Adds CPU trainModel() workflows through luz with case-disjoint validation, caller RNG restoration, and mandatory canonical EEGNet max-norm projection.
Adds contract-checked predictModel() logits, probabilities, classes, and regression responses without changing batch-normalization buffers.

Adds exact PhysioExperiment input-axis and identity contracts.
Adds canonical catch22 features through the optional Rcatch22 backend.
Adds fitted, reusable ROCKET and MiniRocket transforms through the optional aeon backend.
Adds peReducedFeatures() for case-by-feature design matrices with transform provenance.