Changelog¶

This changelog starts from v1.2.0.

v1.2.1 - 2026-06-06¶

Added score-test support for variant-level .bed annotation directories, including BED-only directories that use score-stat HDF5 row data for variant coordinates.
Added estest test --perturb-annot to the wrapped score-test CLI.
Exported read_parquet_sumstats_multi from the top-level graphld package.

Changed LDClumper.clump() and run_clump() to return the original input rows in original order with a boolean is_index column; graphld clump still writes only retained index variants.
Harmonized gaussian_likelihood_hessian() with gaussian_likelihood_gradient() by adding trace_estimator, preserving diagonal_method as a deprecated keyword alias, and defaulting diagonal-only Hessian output to xdiag.
Defined precomputed surrogate HDF5 maps in full LDGM row coordinates, with GraphREML translating them back to the active post-merge coordinate system.
Updated BLUP so public heritability is interpreted as total heritability for the analyzed variant scope and distributed across matched LDGM effect dimensions.
Removed GraphLD's package-import mutation of global NumPy floating-point warning settings.

Fixed graphld CLI option precedence, .vcf.gz summary-stat dispatch, multi-trait parquet REML output naming, and simulate annotation-column forwarding.
Fixed graphREML filtering and validation bugs affecting per-block max-chi-square filtering, binary annotation parameter alignment, and missing sample-size fallback behavior.
Fixed PrecisionOperator subset PCG solves, inverse-diagonal paths, invalid diagonal update atomicity, selection-cache invalidation, copied solver state, and stale factor refresh across shared matrix aliases.
Fixed simulation reproducibility and phasing issues by making seeded effect-size draws block-specific, loading worker LDGMs with metadata-row populations, and phasing annotation-based simulated effects/noise back to the annotation allele convention.
Fixed score-test jackknife block boundaries, deterministic annotation-file ordering, empty trait-group handling, gene-set identifier detection, and genome-scale nearest-gene overflow.
Fixed core I/O edge cases for VCF sample/FORMAT parsing, metadata-row variant partitioning, ID-only SNP-list merges, order-independent BED interval annotation, load_annotations(file_pattern=...), position-only annotation loading, and load_ldgm(snps_only=True) variant metadata filtering.
Hardened parquet summary-stat validation so incomplete trait BETA/SE pairs and missing variant identifiers fail with clear errors.
Fixed surrogate-marker generation for blocks with no candidates by writing the initialized identity map.

Removed unsupported or stale public surfaces: the broken create-geneset-annot console script, the duplicate graphld.heritability_testing module, and the obsolete graphld.genesets_gene_hdf5 workflow. Gene-level HDF5 conversion remains available through estest convert.

Initial baseline for this changelog.