RandLib 2 is a C++20, header-only library. The probability model is familiar, but namespace, layout, engines, fitting, and several type-level APIs intentionally changed. Migrate one distribution at a time and regenerate any seed-based fixtures.

Change	1.x	2.x
Language	C++17	C++20
Headers	`distributions/.../*.h`	`randlib/.../*.hpp` or `randlib/randlib.hpp`
Namespace	Mostly global	`randlib`
Linkage	Compiled library	Header-only `RandLib::RandLib`; optional explicit-instantiation archive
Sampling engine	Engine stored inside each distribution	Any caller-owned standard URBG passed to `operator()` or `Sample`
Reseeding	`const` member changed hidden state	Non-const compatibility member; prefer reseeding the external engine
Bulk arguments	Primarily `std::vector`	`std::span` plus contiguous-range conveniences
Fitting	Mutating `X.Fit(sample)`	Non-mutating `fit<Distribution>(sample)` returning `FitResult`
Errors	General standard exceptions	`randlib::fit_error` with `FitErrorCode` for fitting failures
Support category	Unscoped values	`enum class randlib::Support`
Inheritance	Concrete classes could be subclassed	Concrete distribution classes are `final`
Engine selection	`RANDLIB_GENERATOR` and engine preprocessor macros	Explicit engine type; `DefaultEngine` is Xoshiro256++
Reproducibility	v1.1 algorithms	Corrected streams; no bit-for-bit compatibility promise

Includes and namespace

Before:

#include "distributions/univariate/continuous/NormalRand.h"
 
NormalRand<double> normal(0, 1);

After:

#include <randlib/continuous/normal.hpp>
 
randlib::NormalRand<double> normal(0, 1);

The umbrella <randlib/randlib.hpp> is convenient for applications. Libraries sensitive to compile time can include per-distribution headers.

Engines and sampling

Before:

NormalRand<double> normal(0, 1, 42);
double x = normal.Variate();
normal.Reseed(7);

After:

randlib::NormalRand<double> normal(0, 1);
std::mt19937_64 engine(42);
double x = normal(engine);
engine.seed(7);

Bulk output accepts a span-compatible contiguous range:

std::vector<double> values(1'000);

normal.Sample(engine, values);

make_streams(seed, count) creates deterministic per-worker engines. The CMake and preprocessor RANDLIB_GENERATOR selection mechanism is removed. The legacy JKiss/JLKiss engines remain under randlib::legacy only for investigations that require old engine types.

Fitting

Before:

NormalRand<double> normal;

normal.Fit(values);

After:

auto result = randlib::fit<randlib::NormalRand<double>>(values);

const auto normal = result.distribution;

Use fit_moments, fit_bayes, fit_best, and goodness_of_fit for named workflows. Legacy member functions remain deprecated until v3, but new code should not depend on mutation. Catch fit_error when a fitting failure needs to be distinguished from invalid construction parameters.

Other source changes

Support values require qualification, for example randlib::Support::Finite.
Concrete distributions are final; compose a distribution instead of inheriting from it.
Factories replace constructors whose boolean flags were ambiguous.
Internal numerical helpers moved under randlib::math::detail and are not API.
Bernoulli sampling no longer shares cached bits, and discrete-uniform sampling uses unbiased rejection. These fixes, together with full-mantissa uniforms and engine corrections, change draws for the same seed.
Reseed is non-const because mutation is visible. It only affects the compatibility engine owned by a distribution; explicit external engines are the preferred v2 design.

Do not update a reproducibility fixture by accepting new values blindly. Where the fixture represents a probability function rather than a sample stream, follow the accuracy evidencepolicy".