HomoeoGWAS v2.1.0

HomoeoGWAS v2.1.0 adds a heteroscedasticity-robust bootstrap null for group
omniB, chunk-parallel streaming scans and agent-facing audit and summary
tools.

Highlights:

• interact.calibration.null_variance: smooth_pc4 (canonical group omniB).
  The parametric bootstrap draws null phenotypes whose residual variance is a
  smooth function of the leading genotype principal components, fitted by
  IRLS. When residual variance tracks population structure, the homoscedastic
  bootstrap can be anti-conservative; smooth_pc4 restored familywise
  calibration in the tested settings. On real data it can reject fewer units
  than the homoscedastic null. Results record
  smooth_variance_parametric_bootstrap_minp_plus_one and the fitted
  weights. The low-level config default stays homoscedastic; report it as a
  sensitivity analysis. smooth_pc4 requires bootstrap checkpointing.
• scan.n_jobs for streaming single-locus scans (plain and LOCO). Chunks keep
  the serial chunk boundaries and batches and are scored by forked workers;
  each worker uses one native thread, and the decompressed summary statistics
  are byte-identical to the serial scan run with one BLAS thread (verified on
  1.9 million wheat markers). Each worker holds about
  chunk_size × n_samples × 20 bytes (about 3.3 GB at the default chunk size
  and 827 samples); lower scan.chunk_size or scan.n_jobs if memory is
  limited. A worker that dies stops the scan with an error. scan.n_jobs has
  no effect on in-memory or GPU scans. Gzip output is a standard multi-member
  gzip stream.
• MCP tools audit_results and summarize_results. run_interaction
  exposes null_variance (default smooth_pc4, with checkpointing enabled in
  the generated config) and run_gwas exposes scan_jobs.
• Batched interact cases share phenotype-independent panel preparation, and
  an opt-in hash-keyed GRM cache reuses relationship matrices across runs with
  identical inputs.
• interact.burden.feature_seed separates gene-feature randomness from the
  bootstrap seed (default: the bootstrap seed).

Behaviour changes:

• In group-unit (and joint) families, a homoeolog group is eligible only when
  every declared pair edge is estimable; groups with an unestimable edge are
  marked unestimable instead of being scored on a partial edge set, which can
  reduce the number of tested groups. Edge-unit families are unchanged.
• Partial omniB response errors fail closed.

Reproducibility note: REML variance components depend on the BLAS thread count
in the last digits (about 1e-6 in PVE), which propagates to per-marker
statistics. Bit-identical reproduction requires the same thread settings; the
formal interaction path fixes one native thread per worker.

Upgrade guidance: existing configs run unchanged. To use the robust null, add
null_variance: smooth_pc4 and checkpoint: {enabled: true, root: <dir>}
under interact.calibration.