math: add the flt (binary32) engine (bnd::math::flt::)

[NiceAndPeter]

Phase 4a of the math-engine plan — the flt (binary32) engine

A third math engine alongside cordic (integer) and dbl (binary64): a small, reproducible libm in float, callable as bnd::math::flt::fn side-by-side with the others in one binary. For single-precision-only FPUs (Cortex-M4F etc.) and size/speed where double-grade precision isn't needed.

Same determinism contract as dbl

New cmath_float.hpp mirrors the double engine in single precision: own fixed polynomials evaluated with the correctly-rounded std::fma(float), Cody-Waite range-reduction constants derived at compile time (mask the low mantissa bits of the float-rounded constant — no external codegen), and correctly-rounded std::sqrt(float). So flt is bit-identical on every IEEE-754 binary32 platform compiled without -ffast-math.

Validated empirically before wiring

Measured the float cores against libm: trig ≈ 0.6 ULP, atan ≈ 0.8, log ≈ 1, exp a few ULP — holding across the shared ±2²⁰ domain (the constexpr split keeps float reduction float-grade even at the edge). So flt reuses the same input envelopes — the same programs compile on every engine. (cos uses proper quadrant reduction, not sin(x+π/2): shifting the float input would lose bits.)

Details

• A third value set: float ≠ double ≠ cordic; snapped results differ by up to a few notches on fine grids, coincide on coarse grids and algebraically-exact inputs.
• flt:: has the full public-shaped API mirroring dbl:: (auto-deduced grids, domain static_asserts, expected-returning tan/signed-sqrt/pow). Gated behind !BND_MATH_NO_FP; a flt:: call in the FP-free build is a compile error.
• flt::store widens float→double for real(f64-backed) bounds or snaps via the rational path otherwise.

Tests / docs

test_math_engines.cpp now instantiates cordic + dbl + flt in one TU, checks they agree on exact inputs, and pins 10 float-engine golden values (determinism). Three-engine docs table + flt section. Single header regenerated.

Deferred to Phase 4b

BND_MATH_FLOAT (make flt the unqualified default) and the f32 storage flag + real→f64 rename.

Note on the earlier design question: the prototype showed native float can range-reduce over the full ±2²⁰ domain (float-grade), so I kept the shared domain rather than imposing tighter flt-only limits — simpler and preserves "the same program compiles on every engine." Happy to add artificial tighter limits if preferred.

Verified: default 404/404, CORDIC 442/442, all single-header smokes build.

🤖 Generated with Claude Code