Choose arithmetic independently of linear-algebra providers:

Option	Default	Public types
`UNI20_ENABLE_FLOAT80`	`AUTO`	`float80`, `complex160`
`UNI20_ENABLE_FLOAT128`	`OFF`	`float128`, `complex256`
`UNI20_ENABLE_MPREAL`	`OFF`	`mpreal`, `mpcomplex` (also `complex<mpreal>`)
`UNI20_BACKEND_MPLAPACK`	`OFF`	MPLAPACK kernels for enabled families

float32, float64, complex64, and complex128 are unconditional. Complex format names count both components: complex128 contains two binary64 components, not two binary128 components.

Requests and results

All three scalar options accept:

ON: require usable support; missing dependencies or incompatible formats are configuration errors.
OFF: omit the optional family and its provider kernels.
AUTO: detect usable support without downloading a scalar provider; unavailable support leaves the family disabled.

For binary128, AUTO does not enable a family whose dependency policy forces fetching (UNI20_USE_SYSTEM_MPLAPACK=OFF). If combined mpreal/binary128 conversions are unavailable, an explicit ON request takes priority over AUTO; with both families automatic, mpreal takes priority.

float80 requires native x87 extended-format long double: 64 significand bits and the corresponding exponent range. Storage may include padding. It is not emulated on other platforms. Disabling the alias does not disable the C++ builtin type.

The arbitrary-precision family requires installed GMP, MPFR >= 4.2, and MPC. Real and complex scalars are enabled together. Working precision remains an explicit per-value or per-operation choice, not a CMake setting.

Binary128 currently obtains its scalar type, math, and usage requirements from MPLAPACK >= 3.0.0. Thus UNI20_ENABLE_FLOAT128=ON can fetch that dependency even with UNI20_BACKEND_MPLAPACK=OFF. This does not register Uni20's MPLAPACK kernels: its binary128 BLAS/LAPACK traits remain disabled. Extracting a standalone scalar provider is a possible future implementation improvement.

Requests are not capability macros. C++ code tests the generated 0/1 values UNI20_HAS_FLOAT80, UNI20_HAS_FLOAT128, and UNI20_HAS_MPREAL. Provider integration tests use UNI20_HAS_MPLAPACK_FLOAT80, UNI20_HAS_MPLAPACK_FLOAT128, and UNI20_HAS_MPLAPACK_MPREAL. These are derived results, not independent configuration inputs. The configure summary and build information report arithmetic and kernel providers separately.

Selecting MPLAPACK

UNI20_BACKEND_MPLAPACK=ON requests kernels for all enabled optional families. There are no per-precision backend switches or separate MPC switch. Missing required package components fail configuration instead of silently dropping requested kernels. Requesting the backend with no enabled family also fails. Binary128 GEMM also uses the existing UNI20_BACKEND_BLAS wrapper switch (on by default); disabling that switch leaves its LAPACK kernels available. MPFR and binary80 adapters do not have this additional GEMM requirement.

# Scalars, tensor storage and supported CPU operations, including LU; no MPLAPACK kernels.
cmake -S . -B build_codex/scalars -DUNI20_ENABLE_MPREAL=ON
 
# Binary128 and arbitrary precision, plus native float80 where available.
cmake -S . -B build_codex/wide \
  -DUNI20_ENABLE_FLOAT128=ON -DUNI20_ENABLE_MPREAL=ON \
  -DUNI20_BACKEND_MPLAPACK=ON
 
# Only arbitrary-precision kernels, without a binary128 compiler requirement.
cmake -S . -B build_codex/arbitrary \
  -DUNI20_ENABLE_FLOAT80=OFF -DUNI20_ENABLE_FLOAT128=OFF \
  -DUNI20_ENABLE_MPREAL=ON -DUNI20_BACKEND_MPLAPACK=ON

Dependency sourcing remains independent: UNI20_USE_SYSTEM_MPLAPACK=AUTO prefers an installed package and otherwise fetches the pinned release; ON requires an installed package; OFF fetches. Explicit backend selection may fetch its components even if a scalar was auto-detected.

Scalar availability does not promise every numerical operation supports that type. See binary80, binary128, and arbitrary-precision coverage. For example, binary80 complex LU and decompositions remain disabled for providers with unsafe generic complex division.

Updating older build trees

The old UNI20_ENABLE_MPFR, UNI20_ENABLE_MPC, UNI20_ENABLE_MPLAPACK, UNI20_ENABLE_MPLAPACK_BINARY80, and UNI20_ENABLE_MPLAPACK_MPFR options are removed, not silently interpreted. Remove those cache entries or start a fresh build tree. In particular, the old UNI20_ENABLE_MPLAPACK=ON meant binary128; its replacement is -DUNI20_ENABLE_FLOAT128=ON -DUNI20_BACKEND_MPLAPACK=ON.