Vantage phase 2g: volumetric clouds, and the f32 landmine three green gates missed

What shipped

The planet’s cloud layer went from a 2.5D shaded shell to a true volumetric raymarch, across two review-gated milestones: M1 for the bake, march and wind, M2 for live ground shadows and shaft coupling.

Nubis-style density shaping (coverage remap, detail erosion, powder) now runs per march step, with a short sun-side light march for self-shadowing, two-lobe Henyey-Greenstein phase for the silver linings, and an angle-aware extinction blend so a thin shell reads as actual volume overhead rather than only at the limb.

The noise is offline-baked tileable 3D Perlin-Worley loaded into Data3DTextures, which was the call that kept the WebGL2 fallback alive. Wind is real advection now, not the old sinusoid that could only pulse opacity in place.

Decisions

The backend constraint drove the architecture. Compute shaders would have been the natural way to generate 3D noise, and they’re WebGPU-only, which would mean two code paths forever. Data3DTexture with texture3D() sampling works on both backends in r185, and I verified wrapR is honored on both by reading the backend source rather than assuming. The bake is 2 to 3 seconds and drops into the existing hash-gated deterministic bake with no special handling.

The shaft coupling decision is the one where I had to accept a disappointing answer. The physically motivated version, where clouds occlude the ray source, is geometrically incapable of doing anything. The occlusion buffer is a black background with a small bright disc on it, so black clouds only matter where they overlap the disc, and the cloud shell’s roughly 1% limb ring can never cover a source that extends into open sky. The measured ceiling was 0.03% of shaft pixels. The rescue attempt, enlarging the source disc, changed the approved look by 11 to 18% and delivered zero cloud benefit, so it got reverted. I picked the CPU density dim from the presented options instead: it samples weather at the grazing-limb point with shader-identical shaping, cuts shafts by about 70% at a cloudy limb, leaves a clear limb untouched, and is warp-safe by construction.

Also worth writing down: perf gates get decided by triangulation now, never by a single paced number. This phase hit every failure mode at once.

What broke

This is the real story of the phase, and it’s the review loop working exactly as designed.

The first delivery passed every machine gate. Tests green, both backends booting, perf fine, determinism A/Bs clean. And my eye still said the clouds were sitting on the horizon layer. That’s an unfalsifiable-sounding complaint, the kind that usually gets answered with “it looks fine to me.”

Instead the reviewer built a pixel instrument: settled single-render captures, then radial bucketing of a density-lever A/B. That turned my taste note into a number. Cloud response existed only in a hard-edged annulus between r/R 0.6 and 0.9. The inner 60% of the disc was binary-dead. So “clouds on the horizon layer” wasn’t taste at all, it was a measured bug, and I’d been describing it accurately without knowing why.

The cause was one line. mix(1e9, tEnterInner, hit) in the ray-shell bounds, which the plan itself had suggested. In f32 that cancels catastrophically: the ulp at 1e9 is 64, so 1e9 + (t - 1e9) evaluates to 0, and the march length went to zero for every near-nadir ray. Three green gate suites could not see it, because nothing about it is a type error, a failed assertion, or a non-deterministic result. It’s just quietly wrong arithmetic in a place nobody was measuring.

That’s now a standing rule: in shader graphs, select between real values. Never mask with large sentinels.

Numbers

The density-lever A/B with radial bucketing is the measurement that mattered. Before the fix, response was confined to the 0.6-0.9 annulus with the inner 60% of the disc dead. After, it’s across the disc.

Shaft coupling measures 9.3% of frame pixels at a genuinely cloudy limb and exactly 0% at a clear one, which is the self-gating behavior I wanted rather than a constant tint.

Frame cost is about 1.5 ms real, established by triangulation. The paced reading said a dead-flat 57 fps, and the tell was that it stayed 57 from 1080p all the way down to 360p, so it was the headless pacer again. Uncapped runs gave 855 and 654 fps. Separately, a 120-versus-38 fps mystery turned out to be Marvel Rivals holding 82% of the GPU in the background, and some runs degraded mid-measurement on a cold shader cache. Any one of those numbers on its own would have failed a passing build, and two of them nearly did.

Next

Phase 2h finishes the cloud arc: the deferred idea from this phase, where sun-lit clouds render into the ray source as a bright emitter instead of trying to occlude it, plus a shading-fidelity round on the cloud bodies themselves.