Changes in version 0.0.0.9000                      

  - Bundled the published 41-node X-CNV classifier as a validated
    runtime-neutral R data object, and made it the default when an
    annotation resource directory does not supply an override. Runtime
    prediction remains independent of XGBoost.
  - Bundled a compact real ClinVar 72461/SOX2 example with its
    case-local X-CNV annotations and objective GENCODE/ClinGen content
    evidence.
  - Corrected the published-resource readers: LJB26 START == END records
    are one-base points rather than empty BED intervals, LJB scores are
    selected by their explicit site index, and the headerless
    merged-sample file retains its first observation.
  - Added an installable R package API for reading CNV tables,
    validating resource bundles, computing X-CNV annotations, predicting
    MVP scores, and writing the legacy table format.
  - Added a small deterministic fixture and tinytest coverage, including
    an independent R overlap oracle for the production DuckDB inequality
    join.
  - Removed bedtools and XGBoost from the installed runtime. A
    deterministic staging script converts a separately obtained XGBoost
    model to a validated, tabular xcnv-tree-v1 artifact; the package
    evaluates that artifact in R.
  - Added an independently authored ACMG/ClinGen 2019 constitutional-CNV
    scoring contract with provenance-required evidence rows, published
    score ranges, criterion-group caps, and five-tier classification. A
    DuckDB-backed content provider derives only sections 1, 2A, and 3
    from explicit gene and ClinGen dosage relations; case, phenotype,
    inheritance, partial-gene, and population evidence remain
    curator/provider inputs.
  - Recorded the pinned upstream compatibility scope and current
    resource limitations.