Extract the AABB family, b3Plane, and the wide position validators

[rowan-claude]

Moves the remaining fixed-point-typed aggregates out of box3d. Additive only — 367 insertions, no existing line touched.

What moved

narrow (fixed_vec.h) — b3Plane + b3IsValidPlane; b3AABB with b3MakeAABB, Contains, Area, Center, Extents, Union, Inflate, Overlaps, Transform, b3ClosestPointToAABB, b3IsValidAABB.

wide (fixed_wide.h) — b3AABBWide and the same operations; b3WorldTransformWide; b3IsValidWideCoord, b3IsValidPosWide, b3IsValidWorldTransformWide, b3IsValidAABBWide; b3OffsetAABBWide.

Both widths are unconditional, and that is the point

In box3d these two implementations sit in opposite branches of BOX3D_LUDICROUS_MODE. That flag is off by default, so the wide half does not compile in an ordinary build — it is dark code. Here both are unconditional and both run on every CI job.

One deliberate behaviour change

Everything else is a faithful port. This is not, so it gets its own heading.

box3d's wide b3AABB_Extents narrows each bound to b3Fixed and then subtracts. Past Q48.16 range both bounds saturate to INT64_MAX, their difference is zero, and an ordinary box reports zero extents — at exactly the distances ludicrous mode exists to serve. b3AABB_Transform consumes Extents, so transformed distant boxes collapse with it.

This version differences in 128-bit and then narrows. box3d's own wide b3AABB_Area already does it that way round, so this restores consistency inside that file rather than inventing a convention. For any box whose bounds both fit local range the two forms agree bit-for-bit — checked directly by the correspondence cases — so the fix is invisible to every build that was already correct.

I found this by writing the test, not by reading the code. The first draft of this commit ported Extents faithfully and the test failed.

What did not move

• b3IsBoundedAABB / b3IsSaneAABB — both read B3_HUGE, which resolves through b3GetLengthUnitsPerMeter(), a mutable box3d global. They encode world-scale policy; a pure math library should not hold an opinion about how big a world is, nor acquire a runtime global to do it.
• b3SegmentDistance, b3LineDistance, b3PointToSegmentDistance, b3Steiner — physics and geometry that happen to be written in fixed point.

b3ClosestPointToAABB did move, correcting my earlier classification of it as a geometric query: it is b3Clamp against the box, an operation of the type itself.

Testing

test/aabb_test.c calls every symbol this PR adds. A header-only helper that nothing references is invisible to a passing build — which is exactly how b3WideMin/b3WideMax shipped unreachable in #1.

Proven able to fail in both directions:

Injected fault	Result
Restore box3d's narrow-first Extents	fails the two distant-box checks (266, 274)
Break narrow b3AABB_Union	fails its direct assertions and the narrow/wide correspondence at line 223