Compare every simd-backed function against Apple's simd and fix the differences found

[ctreffs]

Follow-up to #76. FirebladeMath promises the same results on every platform. This PR checks that systematically and fixes the five places where it wasn't true.

On Apple platforms, Apple's simd library is importable even with the simd trait off. The differential tests run FirebladeMath's portable implementations, which are what Linux and Windows use, and compare each result against simd, which is what FirebladeMath calls when the trait is on.

Coverage: all 108 simd-backed functions

Area	Functions
Matrices	multiply (matrix·matrix, scalar·matrix, vector·matrix, matrix·vector for 2×2/3×3/4×4), determinant, inverse, transpose
Quaternions	multiply (both scalar sides and quaternion), add, subtract, conjugate, inverse, length, normalize, angle, axis, act
Vectors	dot, length, normalize (2D/3D/4D), cross (2D/3D)
Scalars	clamp, fract, mix, rsqrt, sign, smoothstep, step
Conversions	matrix3x3/4x4(from:), quaternion(matrix:), quaternion(angle:axis:), quaternion(from:to:)

All of these are covered in Float and Double.

Inputs: 256 pseudo-random values per test, plus edge cases:

• identity, diagonal, permutation and singular matrices;
• small and large scales;
• zero vectors and zero quaternions, unit axes;
• 180° rotations and opposite vectors;
• NaN, ±∞ and ±0.

Tolerances are rounding-level (Float 2e−6, Double 1e−14; inverses and transcendental functions 2e−4 / 5e−12). Results that should be exact, such as transpose, conjugate, add/subtract, clamp, sign and step, are compared exactly.

Differences found and fixed (portable path only; the simd path is unchanged)

Function	Apple simd	Portable before
normalize(Quat) with zero length, and so act	identity	NaN
clamp(NaN, min, max)	min	NaN
fract(NaN), fract(±∞)	0.99999994 (largest value below 1)	NaN
quaternion(from: v, to: -v)	180° about a deterministic axis	NaN
matrix3x3/4x4(from:) with a non-unit quaternion	x²−y²−z²+w² diagonal	1−2(y²+z²) (different result)

• clamp and fract now use Float/Double.maximum/minimum (IEEE maxNum/minNum), as simd does.
• quaternion(from:to:) is now a faithful port of simd's algorithm from the SDK header. Previously it was only the "reduced" formula, which Apple documents as unstable beyond 90°.
• normalize returns the identity when the squared length is zero.
• One deliberate exception: for quaternions whose squared length is subnormal (Float components around 1e−20), simd's fast reciprocal square root returns inf/nan. Copying that artifact makes no sense, so those inputs are left out of the tests.

Staying correct

• Scripts/check-simd-differential-coverage.py (make check-simd-coverage) fails if a function with a simd branch has no test. Each test lists what it covers with // covers: name(Types) annotations, and the script also flags annotations that no longer match a function. It runs in the macOS CI job. I checked that removing one annotation makes it fail and name the function.
• The macOS CI job already runs the tests with --disable-default-traits (from Speed up the portable 4x4 inverse, adjugate and quaternion multiply #76). make test-portable does the same locally.

Verified

• All tests pass: 311 with simd and 379 without, including the 68 differential tests. Both check-inlinable configurations pass, and lint is clean.
• Benchmarks are unchanged (Mat4x4f.fromQuaternion 7 ns portable), and allocations are still 0.

https://claude.ai/code/session_011WSUtoLptjea3s3iC5wTNC