Vantage phase 2h: cloud-shaped sun shafts and the shading round the audit rewrote

What shipped

Phase 2h finished the cloud arc that 2g started, in two milestones.

Milestone 1 gave the terminator its money shot. Sun-lit clouds at the limb now render as a warm HDR emitter into the god-ray occlusion buffer, and because the radial blur streaks every bright pixel outward from the sun, each lit cloud edge casts its own light stream. That’s broken-cloud sunset rays, where 2g had uniform shafts dimmed by a scalar. The 2g dim survives, moved onto the occlusion disc, which works because the blur is linear so dimming the source equals dimming the rays. Uniform shafts still fade behind overcast limbs while the cloud-edge rays keep their punch.

Milestone 2 was planned as a tone-curve fix and turned into something else entirely.

Decisions

Inverting the coupling is the good idea of this phase, and it came directly out of 2g’s failure. Black clouds occluding the ray source is geometrically dead, proven at a 0.03% ceiling. But the same geometry that makes occlusion useless makes emission work: the occlusion pass is HalfFloat HDR end to end and radialBlur has no clamp, so putting a warm emitter above 1.0 at the cloud shell means each lit edge streaks outward on its own. Same pass, same blur, opposite sign, completely different result.

Two traps got pinned by research before any GPU time, which is the 2g lesson being applied rather than just recorded. toneMapped must be false on the emitter, or ACES gets applied inside the occlusion pass and crushes the HDR before the blur ever sees it. And color and opacity need splitting, because additive blending contributes rgb times alpha, so density in both channels squares it.

The decision I’m most glad about is procedural: audit before fix, as a hard gate. It rewrote its own plan within an afternoon.

What broke

The reviewer’s diagnosis was wrong, and the process caught it rather than shipping it.

The clamp diagnosis predicted that shading, scatter and rim were all being flattened. The mandated lever audit measured each one instead of assuming, and the table said only cloud shading was dead, at 0.41% of pixels responding. Worse, the prescribed remedy, a soft-knee on the tone curve, measurably moved the headline metric in the wrong direction. Shipping the plan as written would have made the thing it was meant to fix slightly worse, and it would have looked like progress.

The implementer stopped and reported rather than pushing through, and the real defects came out.

First: the sun-side light march accumulated optical depth two hundred times too small, because tau was never scaled by CLOUDVOL_DENSITY. That makes self-shadowing exp(0), which is 1, which means the lever was completely inert. One multiply to fix.

Second: the volumetric rewrite in 2g had silently dropped the old pseudo-normal bump-shading term, the thing that paints sun-angle gradients across cloud bodies. Nobody noticed at the time because the bodies were blown out at the tone-mapped ceiling anyway, so there was nothing to see a gradient on.

The measurement itself also needed fixing. The broad-frame stddev metric is dominated by cloud edges, and it was being inflated by exactly the ceiling pile-up it was supposed to condemn. Kept as-is, it would have scored the correct fix as a regression. The revised gate masks to the exact stark-white body pixels, same mask on both frames.

Numbers

With the units fix, the bump shading restored and the ceiling softened, cloud bodies went from a stark-white pile at the tone-mapped ceiling to real bright-top and dark-flank structure. Within-body variation moved from 10.5 to 34, a 3.2x change, independently recomputed by the reviewer from the raw PNGs rather than taken from the implementer’s harness. Ceiling pile-up dropped from 75% of body pixels to 5.6%.

Cloud-shaped shaft coupling measures 7.4 to 25% of shaft pixels affected with a structure ratio of 2.34x, and clear limbs self-gate to roughly 0%.

The lever audit now shows all ten cloud levers at or above 3.9% of pixels moving at their extremes, most far above. Determinism stayed byte-exact throughout, and the frame still runs about 9x under budget uncapped.

I approved both milestones same-day, which after 2g’s three rounds was a pleasant change.

Next

The cloud arc is closed. Back to the roadmap: phase 3, the diorama system, with particlr integration scoped alongside it.