Changes in version 0.1.0 (2026-06-30)                  

Broadcasting

  - Broadcasting a bare length-n vector against a square n-by-n matrix
    is ambiguous (the length matches both dimensions). delarr continues
    to resolve this to row-aligned (one value per row, matching base R
    recycling), but now emits a warning at operation-construction time
    so the choice is explicit. Pass a conformable matrix (e.g. matrix(v,
    n, n, byrow = TRUE)) for column alignment. Silence the warning with
    options(delarr.warn_ambiguous_broadcast = FALSE).

First public release.

Features

  - Lazy delarr arrays with a fused, chunk-aware execution engine.
  - Pipelined verbs: d_map(), d_map2(), d_center(), d_scale(),
    d_zscore(), d_detrend(), d_reduce(), d_where(), and d_aperm(), all
    of which stay lazy until collect().
  - Lazy, block-aware matrix products via d_matmul().
  - Storage backends: delarr_mem() (in-memory matrices and N-d arrays),
    delarr_hdf5() (HDF5 datasets, including N-d), delarr_mmap()
    (memory-mapped 2D matrices), and delarr_backend() for custom pull
    functions.
  - Streaming output to disk with hdf5_writer() and collect(into = ...).
  - block_apply() for chunk-wise computation.
  - Getting-started and advanced vignettes plus benchmark scripts in
    notes/.

Dependencies

  - The core package depends only on rlang. The hdf5r and mmap backends
    are now optional (Suggests); the corresponding constructors raise an
    informative error when the package is not installed.