v8.2.3

Added

• A release updates the bio.tools entry.
  build-wheels.yml calls the new biotools.yml workflow after the PyPI
  upload, which sends docs/biotools.json with scripts/update_biotools.py
  and the BIOTOOLS_TOKEN secret. biotools.yml also runs on
  workflow_dispatch, without building wheels.
  tests/test_release_metadata.py fails while CITATION.cff or
  docs/biotools.json names a version other than the one in pyproject.toml.

Changed

• docs/biotools.json names release 8.2.3 instead of 8.0.4, lists BGEN v1.2
  input, and its example command reads the default .npy kinship file.
• The formal models follow the agent-formal-verify 0.2 layout. The two Lean
  models share one Lake project in lean/, and each model's deliberate bugs
  live in a .mutations file beside it instead of in constants and branches
  inside the model. docs/TESTING.md §4 has one row per model and the
  commands to rerun them. lean/check_mutations.py and
  docs/formal-verification.md are removed.
• CITATION.cff names release 8.2.2 instead of 8.0.2, and the release
  checklists in docs/DEVELOPMENT.md and docs/DEPLOYMENT.md include updating
  it.
• The README pipeline diagram and docs/architecture.png show the -loco
  path: per-chromosome kinship, eigendecomposition and association.

Fixed

• The NumPy SNP-statistics fallback computes variances in double precision for
  float32 genotypes, as the native kernel does and as the fallback's means
  already did. Float32 variances used to differ from the float64 result by up
  to 2e-5 relative at 1,940 samples. Runs that load the native kernel are
  unchanged, and so are float64 inputs. The fallback's temporary memory for a
  float32 chunk rises from 1.75 to 2.75 times the chunk.
• The test suite no longer leaves jamma.jlinalg loaded with the native
  extension in a session that set JAMMA_FORCE_NUMPY_FALLBACK. In the weekly
  sanitizer run every test after tests/test_force_numpy_fallback.py saw the
  extension at call time, which hid a failing test_abi_version. The four
  tests/test_snp_stats.py deselections in sanitizers.yml are gone.
• The NumPy SNP-statistics fallback gives the same means and variances for the
  same genotypes whatever their memory order. A C-ordered matrix and a
  Fortran-ordered PLINK read used to differ by about 1e-15 in the variance.
  The native kernel was already order-independent.
• The weekly sanitizer run passes again (#492). The stale-ABI rebuild test no
  longer inherits JAMMA_FORCE_NUMPY_FALLBACK, and the GEMMA equivalence
  report is skipped when the AddressSanitizer runtime is loaded, where it no
  longer fits its 300 s timeout.