Complete rewrite with C++ backend (Armadillo + nanobind), matching the R harmony2 package step-by-step. Highlights ~10x faster than v0.1.0: 858k cells in ~36s on Apple M1 Ultra (vs ~340s previously) Matches R harmony2: correlation ≥0.998 across all PCs Minimal dependencies: only numpy at runtime Pre-built wheels for Linux (x86_64, aarch64) and macOS (x86_64, arm64), Python 3.9–3.13 New C++ backend with BLAS-accelerated dense matrix ops (Accelerate on macOS, OpenBLAS on Linux) Custom scatter/gather kernels replace all sparse matrix operations K-means initialization matches R exactly (Gumbel-max cosine-distance sampling) ncores parameter to control BLAS thread count batch_prop_cutoff parameter for underrepresented batch handling Arrowhead matrix inverse for fast single-covariate correction Accepts pandas DataFrame, dict of arrays, or NumPy array for meta_data C++ progress messages routed through Python logging for proper integration with downstream packages (thanks @yakirr) Breaking changes lamb defaults to automatic estimation (was fixed 1). Pass lamb=1 for old behavior. Default parameters changed to match R harmony2: max_iter_kmeans 20→4, epsilon_cluster 1e-5→1e-3, epsilon_harmony 1e-4→1e-2 Only numpy required at runtime (previously pandas, scipy, scikit-learn) See CHANGELOG.md for full details.