NEWS

seurust 0.1.5

RowMergeMatrices() scatters CSR rows straight into CSC storage.
ReplaceColsC() copies unchanged columns and splices replacements in one pass.
IntegrateDataC() multiplies anchor weights with a dense anchor buffer instead of a generic sparse matrix product.
fast_dist() writes distances straight into R vectors. Inputs of a few thousand cells stay in column-major order; larger inputs are transposed and computed in parallel.
WriteEdgeFile() and DirectSNNToFile() format edge weights into one buffer. The file text is unchanged. DirectSNNToFile() uses the parallel SNN kernel from 256 cells, so the 400-cell case is parallel.
Returned dgCMatrix objects are filled by assigning S4 slots. That skips the methods::new() validity scan on CSC data that is already valid.

FastExpMean(), FastLogVMR(), SparseRowVar(), and SparseRowVarStd() scan CSC columns directly and accumulate per gene. They no longer build a row index first, which made them slower than Seurat.

fast_dist() accepts integer neighbor ids (previously it panicked unless ids were doubles), keeps the names of the neighbor list like Seurat, and reports out-of-range or NA ids as R errors.
GraphToNeighborHelper() now reads each cell's neighbors from its row, as Seurat does; it previously read columns, which gave different neighbor sets for asymmetric graphs. Ragged rows raise an R error instead of a panic.
DirectSNNToFile() returns the SNN graph as a dgCMatrix (it returned the raw x/i/p/Dim list) and validates nn_ranked like ComputeSNN().
DirectSNNToFile() and WriteEdgeFile() format edge weights like C++ setprecision(15), so edge files are byte-identical to Seurat's. File errors are R errors instead of panics.
Integer matrices and vectors are coerced to double before reaching Rust wherever Seurat's Rcpp signature would coerce them (ComputeSNN(), FastCov(), SparseRowVar2() means, and others).
RowMergeMatrices() still accepts a dgCMatrix (Seurat requires an RsparseMatrix); this is now documented.

FastCov() and FastCovMats() read R memory directly and use a blocked, multi-threaded matrix product (FastCov() only computes the upper triangle). They were several times slower than Seurat and are now faster.
FastRBind() copies columns straight into the result instead of going through element-wise ndarray conversions (about 50x faster than 0.1.2 on large inputs).

Removed the unused Eigen C++ bridge. Source builds no longer need a C++ compiler or RcppEigen, which fixes builds where c++ is Clang without a usable C++ standard library ('cmath' file not found).

Update maintainer contact for CRAN submissions.
Link the Rust static library with Apple ld (-force_load) so macOS source builds succeed.
Carry forward the 0.1.1 packaging fixes for Windows configure and portable Makevars.

Fix Windows source installs for CRAN by adding configure.win / cleanup.win so Makevars.win is generated and the Rust static library is linked.
Remove non-portable .NOTPARALLEL from Makevars templates.
Clarify DESCRIPTION / README wording for CRAN incoming checks.

Initial public release of Rust/extendr native kernels matching Seurat's internal C++ API (LogNorm, FastSparseRowScale, ComputeSNN, IntegrateDataC, RunModularityClusteringCpp, and related helpers).
Offline CRAN builds via vendored Rust crates (src/rust/vendor.tar.xz).
Installable from GitHub Releases, r-universe, and (after acceptance) CRAN.