                 Changes in version 0.4.3 (2026-09-27)                  

  - Restore Windows builds by removing an unnecessary platform-specific
    R header.

  - Emit each tidy self-correlation edge once across processing panels.

  - Bound neighbor counts by available candidates and prevent integer
    overflow in top-k buffers.

  - Preserve independent dimension names for rectangular correlation
    matrices.

  - Require unique, non-missing column names for named edge outputs.

  - Use bounded weighted deviations to improve numerical stability under
    extreme rescaling while preserving the documented estimator.

  - Initialize optional SIMD dispatch safely when panel workers start
    concurrently.

  - Validate the estimator against an independent R reference and remove
    the incompatible WGCNA equality assertion and unused suggested
    dependency.

  - Expand regression coverage and align sparse help examples and
    runtime-control documentation.

                        Changes in version 0.4.2                        

  - Prepared the source tree for CRAN review: corrected package metadata
    and the source-level CITATION evaluation.
  - Updated sparse Matrix coercion to avoid the deprecated direct
    triplet-to-dgCMatrix route.
  - Undefined self-correlations for constant, all-missing, or ineligible
    columns now remain NA rather than being forced to 1.
  - Positive-integer arguments now reject non-integral values instead of
    silently truncating them.
  - Applied min_overlap consistently to fully finite dense fast paths as
    well as NA-aware paths.
  - Fixed the NA-aware sparse self-KNN path, including overlap-specific
    denominators and stored non-finite values.
  - Added internal engine helpers for controlled parallel loops, thread
    limits, runtime SIMD detection, blocked panel traversal, tiny top-k
    maintenance, sparse dgCMatrix summaries, and safe interrupt checks.
  - Renamed the mixed dense/sparse native KNN entry point to
    C_bicor_knn_spdem.
  - Kept the public API focused on bicor() and bicor_knn() while
    strengthening the native execution layer.
  - Preserved exact bicor and exact KNN semantics while improving
    maintainability and long-running job interruptibility.

                        Changes in version 0.3.5                        

  - Fixed dense bicor(x, y) matrix indexing for rectangular x/y inputs.
  - Added native mixed dense/dgCMatrix support for bicor_knn(x, y).
  - Made dense bipartite KNN factor-level handling match the sparse
    path.
  - Linked BLAS portably through $(BLAS_LIBS) $(FLIBS) on Unix-like and
    Windows builds.
  - Updated the optional WGCNA parity test to use pairwise-complete
    observations and a single WGCNA thread.

                        Changes in version 0.3.3                        

  - Prepared package metadata, documentation, tests, and vignette for
    repository review.
  - Added runnable examples for exported functions.
  - Clarified signed-threshold semantics in bicor_knn().
  - Clarified KNN edge direction and bipartite behavior.
  - Honored pairwise.complete.obs in the native KNN interface.
  - Removed generated vignette artifacts from the source tree.
  - Removed hard-coded local macOS SDK paths from build flags.
  - Set an explicit C++ standard and capped default OpenMP usage to at
    most two threads.
