NEWS

---

RGlimpse2 0.0.0.9006

• Read R CMD config values without make's stderr, so warnings such as a clock-skew notice about Makeconf on a CI runner can no longer be taken as the compiler command (seen on R-devel for Windows arm64).

RGlimpse2 0.0.0.9005

• Make every build of the phase executable define the same numerical oracle (patch 0009): the in-tree build now pins -fno-fast-math -ffp-contract=off and compiles its AVX2 path without -mfma, as the package build already did, so the in-tree scalar and AVX2 executables, the package's scalar and AVX2 executables, and the pre-0008 sources built the same way all produce byte-identical output on the fixtures. PHASE_SIMD_FLAGS="-mavx2 -mfma" restores upstream's faster but numerically different in-tree build.

RGlimpse2 0.0.0.9004

• Make the phase executables faster and far lighter without changing any output: the imputation forward table is checkpointed every 64 sites and recomputed block by block in cache, the backward pass splits alpha*beta by allele with and/andnot masks, the conditioning bitmatrix is built by transposing rows of a haplotype-major panel copy, the phasing segment transition runs in one pass, per-site transitions are cached, and output rounding no longer allocates. The sparse-PBWT selection compacts its permutation with block moves and, for runs with few target haplotypes, evaluates prefix counts from the run table instead of materialising them, which matters for single-sample runs; the binary reference panel is read through a 16 MB stream buffer so IOPS-throttled storage sees a few large reads per chunk instead of hundreds. Outputs are byte-identical to the previous executables on the AVX2 and portable paths (verified on x86-64 Linux with GCC 13 in the default floating-point environment); per-thread memory drops from about 730 MB to about 30 MB on a 2000-state, 57k-site chunk, while the haplotype-major panel copy adds a second copy of the common-site panel (n_ref_haps x n_com_sites bits, 26 MB for 6198 haplotypes and 35k common sites).
• Remove the phasing kernels' unconditional multiply introduced during the same work after review: a conditional multiply is kept so results do not depend on the denormal mode of the floating-point environment.

RGlimpse2 0.0.0.9003

• Include the pthread header directly in threaded native callers so Rtools ARM64 builds do not depend on a transitive header.

RGlimpse2 0.0.0.9002

• Preserve biallelic symbolic and ambiguous reference records during direct-BAM phasing without attempting a read-level call, and declare GP with VCF Number=G for ploidy-correct likelihood cardinality.

RGlimpse2 0.0.0.9001

• Add rglimpse2_phase_bam() for deterministic direct phasing and imputation from one explicitly indexed BAM or CRAM.

• Add rglimpse2_phase_bams() for one native multi-sample GLIMPSE2_phase --bam-list call with explicit indexed alignments, sample names, and mixed haploid/diploid ploidy. Private BAM/ploidy tables are removed after the call, and the BCF plus its CSI index are published without replacement only after successful staging.

• Make rglimpse2_split_reference() return a typed non_biallelic_reference input error before starting GLIMPSE2 when the requested reference region contains unsplit records.

• Keep biallelic symbolic and other non-observable reference variants in direct-BAM imputation with flat read likelihoods instead of routing them through the SNP caller.

• Ligate overlapping all-haploid chunks using the one-value-per-sample GT stride returned by HTSlib while preserving mixed and diploid phase matching.

• The R package interface is now distributed under GPL-2 or later; bundled upstream components retain their original licences and notices.

• Give every typed operational error the same optional child-process operation and status properties. This preserves the error-value contract when binary build systems install and check separate copies of the package.

• Add a nested R package that builds pinned GLIMPSE2 chunk, split-reference, phase, and ligate executables against the validated htslib contract supplied by Rduckhts.

• Bundle the pinned GRCh37 and GRCh38 autosomal and chromosome X genetic maps, plus explicit zero-recombination coordinate maps for Y non-PAR and mitochondrial sequence, with audited lookup helpers and provenance. Map lookup accepts chromosome names with or without chr and treats M, MT, chrM, and chrMT as mitochondrial aliases.

• Handle zero-span maps safely in phase PBWT grouping. Replace executable test doubles with reproducible vcfppR-generated BCF reference/target pairs that run the real chunk, split-reference, phase, and ligate executables across autosomal, Y, and mitochondrial contig aliases.

• Add stateless wrappers with explicit paths, arguments, seeds, and resources; typed S7 results and operational errors; and typed contract conditions. Split-reference coordinates are canonicalized before output prediction, output paths must be distinct, and ligate lists are validated exactly as the native reader consumes them. Child executables are launched directly with processx argument vectors without an intermediate shell command.

• Build scalar and architecture-specific phase executables and select only runtime-supported AVX2, AVX-512F/BW/VL, or NEON backends.

• Add Unix and Rtools configure paths, pinned SIMDe headers, source-archive drift auditing, and a checksummed Boost 1.90.0 build-only source closure so the executables no longer depend on separately installed Boost headers or libraries. Use parallel native compilation with 2--8 detected make jobs, real executable help/linkage tests, and a synthetic native end-to-end conformance test across all supported SIMD backends.

• Add pkgdown configuration, package-development targets, release checks for Linux, macOS, and Windows, and Linux R-devel coverage. WebAssembly installs retain map and metadata APIs but explicitly omit GLIMPSE2 child executables because that runtime cannot launch the required native processes.