v0.2.5

• DE(): test all non-reference levels of the first non-intercept term by default
• Documentation fixes

v0.2.4

• harmonize(): match Harmony2 more closely

v0.2.3

• Fix performance regression in hvg()

v0.2.2

• In k-nearest neighbors, skip blocks of 64 cells where even the nearest one does not beat the current worst neighbor
• Speed up hvg() for CSR count matrices

v0.2.1

• Fix UMAP correctness issues vs. umap-learn and speed up gradient descent
• Speed up PaCMAP and LocalMAP

v0.2.0

• Harmony: several-fold speedup; remove original=True
• Add Python 3.15 support

v0.1.8

• Port scikit-misc to pure Cython and drop it as a dependency

v0.1.7

• hvg(): min_cells now considers total, not per-dataset/batch, detection rate
• harmonize(): reduce memory usage on large datasets by capping chunks at 4096
• harmonize(): fix buffer overflow
• harmonize(): fix for tau > 0
• normalize(): vectorize log1p()
• hvg(): fix when specifying exclude
• pca(): compute standard deviation as double for accuracy
• find_doublets(): compute doublet pair probabilities directly from integer detection counts
• plot_heatmap(): flip heatmap so rows are plotted top to bottom
• plot_volcano(): plot zero p-values at the smallest float64 instead of dropping them
• de(): fix multiple edge-case bugs
• Create internal threadpoolctl version to avoid threadpoolctl 3.7.0's MKL BLAS oversubscription, and drop as a dependency
• New brisc_blas() function to list which BLAS implementation brisc is using
• Fix bugs in plotting functions
• Fix memory corruption when concatenating arrays with mixed data types
• Fix polars deprecation warnings
• Avoid implicit float-to-double promotions
• Stricter warning checks
• Bump minimum required polars version from 1.41.0 to 1.43.0

v0.1.6

Fix ~10x slowdowns on NUMA nodes with fragmented memory: numa_zeros() no longer requests transparent huge pages with MADV_HUGEPAGE when the kernel would compact memory synchronously to satisfy the request (transparent_hugepage/defrag = always, madvise or defer+madvise). With enabled=always the kernel still uses free huge pages; on fragmented nodes it now falls back to local 4 KB pages instead of stalling in compaction and placing huge pages on remote NUMA nodes.

v0.1.5

• Fix integer column handling in Pseudobulk._create_design_matrix(), used in de() and regress_out()