0.4.4

[0.4.4] - 2026-10-14

Changed

• Every routine follows the Fortran 90 line by line, including the order of
  the argument checks, so some invalid inputs (e.g., a parameter of −∞)
  report a different error.

• The integer quantile of the discrete distributions brackets the answer by
  doubling.

• Error messages are more descriptive and print floating-point values in
  short form.

Improved

• Fortran 90 reference tables for every kernel, regime, error exit, cdf*
  call and status, and for the iteration traces of dinvr and dzror;
  tests/regenerate/coverage.sh checks that they execute every reachable
  line of cdflib.f90, and tests/regenerate/rust_branch_coverage.sh that
  they take both sides of every branch of the port except those listed with
  a reason.

• The code comments reproduce the Fortran 90 comments verbatim.

• The documentation states the behavior at NaN, infinite and endpoint
  arguments, the tolerances of the searches, and the precision limits of
  the noncentral distributions and of the closed-form densities and masses.

• CI also runs Miri, the tests with the MSRV, semver checks, the Fortran
  coverage check, and the regeneration of the reference tables.

Fixed

• The reference is the current upstream cdflib.f90, which calls
  gamma_user instead of the intrinsic Γ (fixed upstream after our report).
  The port reproduces every reference table bit for bit, in debug and
  release builds, on macOS on Apple silicon.

• Inputs on which the Fortran loops forever return NaN or an error, and the
  noncentral distributions panic where the Fortran integers would overflow.

• Constructors and searches reject parameters on which the Fortran returns
  meaningless values, and several inputs no longer panic.

• cdf, ccdf and the inverses are exact at the ends of the support, and
  a NaN argument gives NaN.

• Densities, masses, moments and entropies no longer overflow, underflow or
  lose precision for extreme parameters.

• try_gamma reports Underflow instead of Overflow below about −171.6.