Changes in version 0.1.0                        

  - couple() and tune_cca() gain method = "pcr" for principal components
    regression — an OLS alternative to CCA that maps predictor PC
    amplitudes to predictand PC amplitudes. Regularization is the
    predictor truncation (k_pred); the CCA-style coupling k/k_cca is
    inert for PCR (a warning is emitted if passed). CCA-specific
    diagnostics (get_canonical_*) are not defined for PCR couplings.
  - patterns() accepts multi-attribute stars objects for combined
    (multivariate) EOF analysis — e.g. joint temperature + precipitation
    downscaling with physically consistent predictions. Requires scale =
    TRUE.
  - The whole stack is multivariate-aware: project_patterns(),
    reconstruct(), couple()/predict(), tune_eof(), tune_cca(), and plot
    methods. CV metrics gain per-variable columns plus a pooled score.
  - Breaking change: EOF attributes in patterns$eofs are now named after
    the variable (previously "weight").
  - Cells with non-finite standardized anomalies (climatological sd
    of 0) are now excluded like NA cells.

Initial release.

Core Features

  - patterns() for EOF extraction from stars spatiotemporal data, with
    area weighting, standardization, varimax rotation, and monthly
    anomalization
  - couple() and predict() for CCA-based statistical downscaling
  - common_patterns() for joint EOF decomposition across multiple
    datasets
  - tune_eof() and tune_cca() for cross-validated hyperparameter
    selection with contiguous temporal folds
  - Scree plots with North et al. (1982) error bars and modified Rule N
    significance testing
  - Teleconnection maps with FDR-corrected significance contours
  - Full integration with stars, sf, and the tidyverse

Architecture

  - S3 classes: patterns, coupled_patterns, common_patterns, cv_folds
  - Plot, print, summary, and predict methods for all major classes
  - Automatic IRLBA truncated SVD for large datasets
  - Unit preservation throughout the pipeline via the units package