The previous per-step accumulator weighted density by `step_len` — the
RK45 step length. That step size is spatially adaptive (small near the
hole where curvature is high, large in flat regions) and varies between
neighbouring pixels, so `density * step_len` injected a per-pixel
brightness modulation along the radial direction: radial spokes,
densest at the inner disk where steps are smallest.
Fix: a new integrate_disk_segment clips each step's segment to the disk
slab analytically and weights samples by `seg_len` — the world-space
length of the ray-slab intersection, which depends only on the ray's
incidence angle and the local slab height H, never on the RK45 step
size. Slab height uses the step-midpoint r (H/R ratio; r varies <5%
across one step).
Within the clipped segment, N=4 uniform samples replace the single
midpoint. A single sample under-integrates the Gaussian vertical decay
and leaves the disk translucent; four samples resolve the density
profile across the slab thickness while keeping the cost bounded (4 ×
2-octave fbm per in-slab step).
This recovers the pre-regression slab-clipping geometry but with
segment-multi-sampling instead of one midpoint — the slab clip removes
the step-size coupling (spokes), the multi-sample restores the
integration depth (opacity).
Off-tier (disk_quality == 0, flat path) is untouched: it samples a
zero-thickness midplane with no slab, so step-size weighting never
applied there.
xfy
359d2f876f