bgen-limix 4.15.0

Public native BGEN layout-2 encoding: bounded GP/HP float32/64 batches, signed hard calls, specialized legacy genotype blocks and fused fixed-stride variant records. Includes GIL-free Python APIs, checked C interfaces, explicit missingness, stable quantization and deterministic stored zlib framing. Preserves the 4.14 reader APIs.

Validated all 21 C tests across Linux/macOS/Windows Release+Debug, Linux sanitizers, S3 integration, and 1,212 Python tests on each of four wheel platforms (5 optional tests skipped). Public PyPI installation on CPython 3.12 verified all five encoder APIs and the full Python suite.

Source: cdb373e.
Trusted publishing: https://github.com/jpfeuffer/bgen-limix/actions/runs/39433422681
PyPI: https://pypi.org/project/bgen-limix/4.15.0/

Wheels: CPython 3.11+ abi3 for Linux x86_64/aarch64 (manylinux_2_28), macOS arm64 (11.0+), Windows AMD64. Other platforms use the source distribution. Runtime nanobind-backend dependency is resolved from PyPI. Full changes and API contracts are in CHANGELOG.md and README.md.

4.14.0

4.14.0 - 2026-10-19

Added

• Public BgenFile.read_dimensions(offsets, *, ploidy_source="actual", probability_model="auto", max_ploidy=None) returning simple GP, HP and decode
  capacity maxima. Actual-ploidy and declared-header strategies use native raw/
  zlib prefix reads and validated full zstd blocks; layout 1 has fixed intrinsic
  dimensions without genotype payload access. Explicit model/ploidy assumptions
  avoid unnecessary discovery and must be checked during full decoding.
• Optional max_actual_ncomb on modern batch reads for a single-pass compact
  capacity, with no width pre-scan and no declared-width probability scratch.
  Validate actual sample widths, including missing samples, before writing each
  variant; report effective actual row widths in batch ncombs.

Fixed

• Allow compact reads with loose declared maximum ploidy even when unused
  declared genotype counts overflow: compute and validate actual ploidy ranges
  before combination counts. Default single/batch decoding remains unchanged.

4.13.0

4.13.0 - 2026-10-19

Added

• Modern BgenFile.read_genotypes_batch(offsets, max_ncomb=None, precision=64)
  returning GenotypeBatchResult: owned contiguous probabilities and columnar
  phase, ploidy, missingness, allele/bit counts, declared variant widths and
  structural per-sample probability lengths. Supports both layouts, every
  supported compression method, float32/64 and zero-width phased records.
• Public native structured float32/64 batch and metadata APIs sharing the
  single-record header/decoding core, with checked dimensions, bounded
  one-variant scratch, explicit padding capacity and NaN padding.
• BgenFile.read_genotype_metadata and get_ncombs for discovery without
  probability decoding. Compressed layout-2 blocks still decompress; automatic
  batch sizing scans widths before the decode pass.

Fixed

• Support layout-1 raw/zlib and raw layout-2 records in existing native and legacy
  Python batch/count APIs, preserving probability-only legacy return values and
  avoiding metadata work on paths that do not request it.
• Reject invalid batch offsets, insufficient/overflowing widths and use after
  reader closure explicitly. Preserve requested ordering and duplicates, exact
  layout-1 rounding, original missingness and result ownership after close.

Performance

• Keep probability-only and width-only paths free of new metadata allocation
  and copying. Reuse native decode/decompression scratch and initialize padded
  output with lazy contiguous NaN fills rather than slower per-row tail loops.
• Extend the reproducible comparison tool to float32 single reads, wider padded
  batches and actual installed Python packages, using pre-change 4.12.0 as the
  baseline for this API work.

bgen-limix 4.12.0

Changelog

4.12.0 - 2026-10-16

Fixed

• Read uncompressed layout-2 genotype blocks using their encoded byte lengths,
  rather than a fixed six-bytes-per-sample assumption. Validate sample counts,
  ploidy ranges, bit depths and probability payload boundaries.
• Support layout-1 metafile creation and partition reads for raw and
  zlib-compressed files, preserving original genotype offsets and variant
  metadata and validating each record's sample count.
• Populate layout-1 native ploidy and missingness metadata, preventing NULL
  buffer access from Python genotype/probability reads. Layout-1 samples are
  diploid and unphased; all-zero probability triplets are missing. Preserve
  independently rounded probabilities as q / 32768 without renormalization.
• Accept valid phased layout-2 records with zero ploidy for every sample,
  returning probability arrays of shape (nsamples, 0) while retaining
  missingness and ploidy metadata.
• Remove eight-bit truncation of unphased genotype combination loop counts,
  widen probability bit-offset arithmetic, and compute implied probabilities
  using integer residuals before floating-point conversion.
• Reject unsupported header flags, truncated records and decompressed-length
  mismatches explicitly.
• Expose native unsigned 32-bit sample and variant counts without signed
  narrowing in Python, and report the installed bgen-limix distribution
  version through cbgen.__version__.

Added

• Public layout properties on BgenFile and legacy bgen_file, backed by
  the native parsed header.
• Generated native and Python regression tests covering raw/zlib/zstd blocks,
  1--32-bit probabilities, phase, multiple alleles, variable and zero ploidy,
  layout-1 indexing, both floating-point precisions, missingness and malformed
  inputs.
• A reproducible native performance comparison tool,
  scripts/benchmark_decode.py, with deterministic synthetic fixtures.

Performance

• Fuse fixed-ploidy validation with metadata copying and precompute genotype
  combination counts outside per-sample decode loops, retaining bounds checks
  without the initial correctness patch's normal-use slowdown.
• On an Apple M2 Max, matched Release builds with 20,000 samples and 128
  variants measured end-to-end unphased Python reads approximately 4--6% faster
  than 4.11.0 for the tested 8/16/32-bit compressed workloads. Phased 8-bit
  reads were essentially unchanged; combination-only scans retained about
  0--3% validation overhead. These are local warm-cache measurements, not
  guarantees for other systems or datasets.