Scattering points with `i * stride % size` lays them on a lattice, not at random

Fact

x = (i * stride) % size is the usual way to place "random-looking" points without an RNG, and it is not random in the way it looks. Consecutive samples differ by exactly stride % size, so the points walk one straight line that wraps — they land on a small family of parallel lines. Doing it independently per axis gives a 2D lattice.

Sand grains placed at x = i*137.51 % 142, y = i*71.31 % 120 step by (-4.49, +71.31) every iteration, so at 900 grains the bunker is covered in diagonal chains rather than grain.

The structure hides at low counts and emerges as density rises, which is exactly backwards from how it gets tested: a handful of points to check the code, then the real count in the render.

Why it matters

This trick is reached for precisely where randomness must not be visible — texture grain, speckle, stars, scattered foliage, dither. It is also the form that reviews as fine: deterministic, no seed to thread, no RNG to import.

The failure appears only in the rendered image, and it reads as a texture or UV problem — repeating diagonals look like a tiling artifact — so the search starts in the sampler and not in the six characters that chose the stride.

How to apply

• Use the R2 low-discrepancy sequence instead. It is the 2D golden ratio, one multiply per axis, and it fills the square without a lattice:

  const x = ((i + 1) * 0.7548776662466927 % 1) * width;
  const y = ((i + 1) * 0.5698402909980532 % 1) * height;

• For grain, prefer low-discrepancy over a plain RNG anyway: uniform random clumps and leaves holes, which reads as blotchiness rather than as texture.
• The tell is the sample count. Raise it tenfold: true scatter only gets denser, a lattice gets more obviously ruled.
• A hash of i works too, but only a decent one — the low bits of i * k are the same lattice by another name.