allelix 2.3.1

v2.3.1

on Python PyPI

Fixed

• Cross-build enrichment is now fail-closed. gnomAD v4.1 exome
  frequencies and coordinate resolution, AlphaMissense, and CADD are
  GRCh38-only in the current cache model. GRCh37 and GRCh36 analyses
  now skip all three sources and omit their provenance instead of
  allowing an rsID or coordinate collision to attach a valid GRCh38
  value to the wrong assembly or allele. ClinVar continues to dispatch
  to its build-specific GRCh37 or GRCh38 cache.

• GWAS odds ratios are no longer confused with beta coefficients.
  The effect-size magnitude modifier is applied only when the
  confidence-interval field positively identifies a bare odds ratio.
  Beta coefficients, unit-bearing values, missing or ambiguous effect
  types, and non-positive effects now receive no modifier. On the
  user1190 genome with the 2026-07-23 GWAS Catalog snapshot, this
  changes the default-filter GWAS output from 94 rows to 11; counts
  drift with future database snapshots, while the effect-type rule is
  the stable regression invariant. Beta-specific thresholds remain
  deliberately deferred until effect units can be modeled safely.

• Position-based build detection requires meaningful evidence. A
  detected build is authoritative only with at least three concordant
  seed-SNP matches and an 80% winning ratio. A 4/5 signal is confident;
  3/4 and 2/2 signals are tentative. The CLI and HTML report distinguish
  tentative evidence from a real detection even when the fallback build
  happens to be the same value. The existing GRCh36 fail-safe remains:
  a tentative GRCh36 winner stays on GRCh36 so no modern position cache
  is queried, while the diagnostics still label the evidence tentative.

• SNPedia genotype matching honors source orientation. Cache parser
  format 7 carries SNPedia's plus, minus, or unknown orientation into
  runtime matching. Known minus-strand rows are complemented to forward
  orientation. Unknown-orientation palindromic A/T and C/G genotypes
  abstain rather than risk a false match. Rebuild the SNPedia cache with
  allelix db update after upgrading. GWAS risk-allele strand
  normalization remains deferred pending a forward-strand data model.

• Slow-suite ClinVar drift checks can no longer pass by skipping. The
  workflow downloads the ClinVar cache explicitly, requires it in CI,
  selects the drift test by its exact node ID, emits JUnit XML, and fails
  unless the named drift test appears as executed and passed.

• Annotated VCF output declares every emitted contig. Before writing,
  Allelix scans the input records and synthesizes a missing
  ##contig=<ID=...> declaration for each raw CHROM value. Existing
  declarations and record chromosome names remain unchanged, including
  mixed bare and chr-prefixed names, so bcftools view no longer warns
  about undeclared contigs created by an incomplete input header.

Documentation

• Synchronized the README, security policy, example-report notes,
  GRCh36 guidance, test-data catalog, edge-case expectations, and full
  release protocol with v2.3.1 behavior. The real-data wrong-allele gate
  now requires a non-zero GRCh38 direct-match denominator and separately
  asserts that GRCh37 enrichment is zero. The protocol covers all nine
  parser formats and treats any release-gate skip as a failure.