Import ftz for reproducible numerical types and math. The native dependency supplies SIMD registers and masks. Import ftz.controls on its own to manage the floating-point environment without pulling in numerical types or vector operations.

Path	Responsibility
cxx/modules/ftz.ccm	m32, h32, conversions, operators, admission and exported math
cxx/modules/ftz.controls.ccm	Thread FP controls, scoped restoration and external-call boundaries
cxx/ftz/math.h	Scalar and register-pack host math implementation
cxx/ftz/simd.h	The custom-element specialization for native::simd and packed math
hlsl/ftz/ftz32.h	HLSL 2021 scalar wrapper, conversions and operators
hlsl/ftz/math.h	Shader math overloads
shared/ftz	Arithmetic graphs, coefficients and bit operations consumed by host and shader implementations

The C++ headers are inputs to the named module, installed so downstream builds can generate their own compatible BMIs. They do not form a separate header-only public API. The shader headers are the public ftz::hlsl interface; applications own their entry points and choose the compiled policy with defines.

native owns native registers, masks, ISA values and generic wide forwarding. FTZ supplies an element customization once for each arithmetic policy; that customization composes with the supported native architectures. FTZ imports native operations and scalar storage, and owns the polynomial graphs and range policies required by its arithmetic contract. The dependency remains one way. Array math kernels expose independent register chains without depending on the wide container itself.

Manual and hardware policies describe the same arithmetic results under their respective admitted environments, with NaN sign and payload left unspecified. Hardware policy removes flushing which its established FP environment already performs. Boundary repairs remain where hardware differs in underflow rounding or signed-zero behavior. Shared integer helpers that implement these production repairs belong here. Coefficient checks, numerical regressions and code-generation probes exercise those definitions from tests.

HLSL remains at language version 2021. Its portable integer implementation uses 32-bit words; FTZ_SHADER_INT64 selects the alternative where native 64-bit integer arithmetic is available. Host code uses its native implementation rather than adopting that shader limitation.

The arithmetic guide defines the public obligations; building covers installed module dependencies.