[piexelated]

Summary

CNG fractures warped the second axis by reading the same fracture noise at swapped coordinates, so every fractured style mirrored itself across x = z. CellGenerator's shuffle had the same swap. Each axis now gets its own reseeded fracture, read at unpermuted coordinates.

 2D
-  fz = z + fracture(z, x)
+  fz = z + fracture1(x, z)
 3D
-  fy = y + fracture(y, x)
-  fz = z + fracture(z, x, y)
+  fy = y + fracture1(x, y)
+  fz = z + fracture2(x, y, z)

• CNG.fractureAxisSeed(seed, axis) derives the seed. Axis 0 keeps the fracture's own seed, so the x warp is unchanged. Axes 1 and 2 use a splitmix64 mix, because seed + k would collide with RNG.nextParallelRNG.
• fracture.reseededForAxis(axis) returns a cached copy rebuilt from that seed. Custom generators can't be rebuilt from a seed, so they get a seed-derived offset instead.
• A reseeded fracture is an ordinary node with its own seed, so the native and GPU kernels and the Rust crate are untouched, and zoom stays exact.

Evidence

Same seeds and coordinates. Each badge is the share of pixels equal to their mirror across x = z. The grid's 65.5% on this branch is chance agreement, since most of its pixels are background.

• Before: CNGFractureMirrorTest fails on master, 780 of 780 sample pairs mirrored.
  After: it passes.
• NoiseProgramParityTest passes unchanged, and 79 of the 207 parity shapes stay native-backed, the same as master.
• All 42 util.noise test classes pass on linux-x86_64, 164 tests.
• GPU tests not run, this machine has no GPU.

NoiseProgramBenchmark, median of 8 alternating master/branch pairs. The noise floor is ±4%.

• Deep fracture chains are 1.2 to 1.6× faster on every 3D path, because each level needs at most 4 reads instead of up to 9.
• 2D and single-level fractures are flat.
• NOWHERE 3D scalar is about 10% slower and 2D scalar on the double chains about 7% slower.

Merge Danger

Door: one-way

Fractured terrain changes for existing worlds, structure carve blobs and CellGenerator output included. There is no compatibility flag.

Blast Radius: worldgen

VolmitSoftware/Iris#1264 moves Iris to CNG.fractureAxisSeed and CNG.reseededForAxis. It does not compile against VolmLib master, so merge this PR first. The Mac pin in NoiseSeedCohesionTest is still stale there. A CNG subclass that changes sampling must override reseededForAxis, since its reseeded copy is a plain CNG.