Root cause of the radial spokes (the "zebra stripe" artifact): the
volumetric integration used TWO overlapping sampling paths per RK45 step
with inconsistent angle-dependent weights:
(A) in-slab step: accum += density * step_len
where step_len = full RK45 step (may extend far outside the slab)
(B) edge-capture: accum += density * thickness_proj
where thickness_proj = half_thickness / |dir.y|
At a straddling step BOTH fired. step_len (RK45 adaptive) and thickness_proj
vary with the ray's bending geometry differently, so their sum modulated
brightness by angle → radial spokes whose count/position shifted with the
camera angle. This was confirmed decisively: setting the texture to a
completely uniform constant (no noise, no arms) STILL produced spokes,
proving the texture was innocent and the integration was the source.
Fix: replace the two-path design with a single unified path. Clip each
RK45 step's segment to the slab |y|<=half_thickness (parametric clip in
t-space), sample once at the clipped segment midpoint, accumulate × the
in-slab segment length only. This gives consistent, geometry-correct
weights with no double-counting. Verified: uniform-constant texture now
renders a clean smooth ring (no spokes); full texture restores turbulent
filaments without spokes.
Also reverts the temporary uniform-constant debug texture to the full
polar-sampled ridged+density+arms function.
xfy
e00c37a3ae