Fix GetArbitraryAxis: the 1/64 threshold is integer division and is always zero

[redbluevn]

The problem

Matrix4.GetArbitraryAxis(XYZ zaxis) implements the arbitrary axis algorithm, which picks its reference axis from the world Y when the normal is close to the world Z, and from the world Z otherwise:

if (Math.Abs(zaxis.X) < (1 / 64) && Math.Abs(zaxis.Y) < (1 / 64))

1 / 64 is integer division between two ints, so the threshold is 0. The condition can never hold, and the Wy branch is dead code.

Why that is not harmless

Every normal then takes the Wz branch. That is correct for a genuinely tilted normal. It is badly wrong for a normal that means +Z but carries the rounding dust real files record — this one is taken verbatim from an architectural drawing:

(-3.707652143685886E-13, 8.818399612355268E-14, 1)

Wz × N is then on the order of 1e-13, and normalising it returns a direction decided entirely by that dust. The frame that comes back has nothing to do with the entity. The zaxis.Equals(XYZ.AxisZ) fast path does not catch it either, since the vector is not exactly +Z.

Downstream in ACadSharp this positions block references: measured across eighteen drawings, 169 inserts whose normal was +Z to within a rounding error were being placed by a garbage rotation — enough to leave the drawings' computed extents tens of times too large. A block reference with Normal = (-3.7e-13, 8.8e-14, 1), Rotation = 1.2, Scale = (1,1,1), whose block spans 0..113 × 0..426, was reported at (-9741106, -1745685) instead of near its insertion point (3952258, -9072802).

The fix

1.0 / 64.0.

Tests

There were no tests for Matrix4 at all, and none for GetArbitraryAxis, which is why this survived. The new Matrix4Tests pins:

• identity for +Z, and the X/Z mirror for -Z;
• the regression case — a near-+Z normal must leave a point essentially where it was (this test fails against the current implementation and passes with the fix);
• that either side of the threshold still yields a frame whose Z is the normal.

dotnet test on CSMath.Tests: 248 passed / 0 failed.