RT vegetation: a dense casting meadow needs more builds per frame than the budget allows, so the TLAS is withheld near it #1578

What happens

In the IntegratedRenderer benchmark meadow (Scenes/Benchmark/IntegratedRenderer.olo, camera inside the grass), the two casting grass layers plan 1,681 to 1,954 ray-traced vegetation groups within their 150 to 170 m view distance. The per-frame vegetation budget allows VegetationPolicy::UpdatesPerFrame = 1024 acceleration-structure builds. A group is one build per part, whatever its plant count.

At the live editor's 0.1 s frame step, every group is past its proxy deadline (0.05 s), so every group refreshes every frame. The casting plan from #1354 then puts 657 to 930 groups OUT and withholds the TLAS from every ray-traced effect. It is explicit, counted and stable (pressure: frameWork, one fallback engagement, raster shadow maps instead), but it is withheld for as long as the camera is anywhere near the meadow.

Measured live on Vulkan with olo_rt_scene_stats, Release editor, RTX 4090, 2026-10-09:

The complete lower-cost tier (cards) does not help here: a card group is still one build, so the build count, not memory, is what binds.

The decision this needs

Which of these, or what mix:

1. Coarser casting groups beyond the mesh distance, e.g. card groups sized by area rather than by CardPlantsPerGroup = 256 plants, so a far field costs fewer builds.
2. A larger UpdatesPerFrame for casting layers, measured against the backend's actual build time.
3. A longer proxy deadline for far casting groups, i.e. a looser shadow error bound at distance than the 0.25 m MaximumWorldDisplacementError.

Each one changes either a declared error bound or the budget the backend is sized for, so it is a design call rather than a bug fix.

Found while working on #1354.
Not fixed in #1579 because: needs a decision — the fix changes a declared error bound or the backend's per-frame build budget, and both options trade shadow fidelity or frame time.

capability: 3
craft: 3
stability: 2
decay: 2
effort: 3
confidence: 0.6
learning: 3
fun: 2
kano: performance
blocked_by: []
blocks: []

Rebuild validation in #1603

Not fixed in #1603 because: needs a decision — changing the vegetation frame-work budget, grouping or declared shadow error/representation-quality bounds changes frame cost or fidelity. The staged rebuild preserves the complete-representation policy.

The committed FoliageMeadowToWoodland scene has 491,773 plants. Six Vulkan RT-enabled near/overview/far CSM/VSM cells preserve census through replacement but withhold the shadow TLAS under the default budget. At Step8 there are 239 planned casting groups and 173 refused; admitted castingDemandVertices is 992031, not the complete-scene cost. Unlike the original integrated-meadow example above, the rebuild fixture also reaches the vertex-work cap, so raising the build-count cap alone would not resolve every case.

The preserved synchronous control has the same 173 refusals, admitted demand, withheld shadow TLAS and canonical census before/after rebuild. This is an existing budget-policy gap, not a staged-publication regression or successful RT shadow execution. Raster fallback remains active. Summary: https://github.com/drsnuggles8/OloEngineBase/blob/feature/foliage-rebuild-stalls-1582-code/docs/analysis/foliage-rebuild-stalls-1582.md . Raw proof stays in ignored local .v-build/pr-evidence-retained/ (final-live/rt-control and live-matrix/vk-rt).