TrioHap 0.5.0 release validation, 25 October 2026

All 165 unit tests pass with plotting and alignment dependencies installed.
Reproduce: python -m unittest discover -s tests -v

Original SNP/VCF/gVCF inference modules are byte-identical to the 0.4.0 release.
The 100-family SNP and 100-family gVCF synthetic validation suites still pass.
New sequence tests cover actual indexed BAM/CRAM input, trio transmission,
missing evidence, SNP/sequence overlap, symbolic uncertainty and report output.

PER3 source import reconstructs all 462 assembly haplotypes independently from
Minigraph-Cactus and PGGB: 462/462 full-window sequences agree exactly, yielding
12 full-window classes and eight repeat structures. Direct SV evidence tests
verify identical complete support matrices for all 12 classes and all 231
original diploid genomes under the two real decomposed VCF encodings.

The new synthetic PER3 experiment uses 12 invented trios spanning all 12
classes, seed 73031. All 12 parental pairs are uniquely recovered from explicit
SV VCFs and idealized 150bp BAM reads. All 12 trimmed individual-VCF comparisons
match joint results; all 12 missing-father checks preserve unknown evidence.
Reproduce: python scripts/validate_per3.py --bam
Machine-readable results: validation/per3_summary.json
Rendered synthetic sequence example: demo-per3-output/.

Local 0.5.0 packaging checks passed strict twine metadata validation, isolated
installed-wheel tests (165 discovered, 149 passed, 16 optional-dependency skips),
independent source-archive rebuild, and original plus structural-only CLI/graph
smokes. Optional BAM/CRAM tests run separately with alignment dependencies
installed. GitHub CI repeats these checks on the final
source snapshot; its outcome is recorded in GitHub.
Reproduce from a clean checkout with an empty dist/ directory:
           python -m pip install '.[dev,plots,alignments]'
           python -m build
           python -m twine check --strict dist/*
           python scripts/validate_distribution.py --dist-dir dist --tag v0.5.0 --plots

These checks use synthetic family data and idealized alignments. They do not
measure clinical sensitivity, specificity, mapping bias or calibration. The
PER3 source repository was read only; its data attribution remains in
examples/per3/README.md and SOURCE_LICENSE.md. This record describes validation;
publication is verified separately through the release workflow and PyPI.

Previous release validation record follows.

TrioHap 0.4.0 validation, 25 October 2026

95 unittest cases pass, including 34 gVCF tests, 19 individual-VCF tests, and
exact SCN5A catalogue/source-byte provenance reproduction after checkout.
Reproduce: python -m unittest discover -s tests -v

New gVCF experiment: 100 synthetic families, seed 64031, eight haplotypes and
40 SNPs per panel. All expected parental backgrounds recovered. All 100
individual-gVCF, 100 joint-gVCF, and 100 mixed-input comparisons give the same
rankings and weights as equivalent explicit VCFs. All 100 dropped-reference-
block checks preserve unknown coverage. Reference bases vary within blocks;
ALT dictionaries and child phase orientations vary across families.
Reproduce: python scripts/validate_gvcf.py
Machine-readable results: validation/gvcf_summary.json
Rendered synthetic gVCF example: demo-gvcf-output/.

Deterministic synthetic experiment: 100 families, seed 4927, eight panel
haplotypes and 40 SNP markers per family. All 100 expected maternal/paternal
pairs recovered; all 100 joint/individual results agree exactly. Independent
per-file ALT encodings and child phase orientations are varied. Fifty families
include an inherited allele absent from every panel haplotype. All 100 missing
parent-record checks preserve unknown status and match a reduced joint input.
Reproduce: python scripts/validate_synthetic.py
Machine-readable results: validation/synthetic_summary.json
These are functional checks under the implemented model, not clinical accuracy.

Joint and individual CLI example reports/plots render successfully; their
rankings, phase, markers, summaries and optional Unfazed annotations agree.
Outputs: demo-output/ and demo-individual-output/.

Reimporting the included SCN5A source subset reproduces catalogue.json exactly.
Earlier SCN5A source VCF/table verification: 20,835 allele cells; 45 REF alleles
match the supplied reference window. Synthetic panel-derived integration
recovers the expected maternal eight-member SNP class and paternal H05.
The SCN5A audit graph contains no patient placement. Full-variant classes remain
unresolved where SNP signatures coincide. No clinical trios have been tested.

GitHub Actions configuration covers Python 3.10-3.13 unit tests and both
synthetic experiments, plus separate VCF/gVCF report-rendering checks.
Its remote outcome is recorded by GitHub. Current gVCF scope is conventional
INFO/END blocks; newer sample-specific FORMAT/LEN is explicitly unsupported.

PyPI packaging preparation (distribution name triohap, version 0.4.0):
Wheel and source archive pass python -m twine check --strict. All 95 unit
tests pass against an isolated installed wheel outside the source checkout.
The source archive independently rebuilds to the same Python module bytes;
that rebuilt wheel installs separately and passes the VCF/gVCF CLI checks.
Installed-wheel graph generation, console entry point, --version, expected
parental assignments, and output files pass. A mismatched v9.9.9 release tag
is rejected. No distribution was uploaded to PyPI during these checks.
Reproduce: python -m build
           python -m twine check --strict dist/*
           python scripts/validate_distribution.py --dist-dir dist --tag v0.4.0 --plots
