#terrain

The method for large runtime-generated terrain in s&box: a persistent cell grid as the single source of truth, chunked greedy meshing (tops + skirts), collision decoupled from render grain, palette-atlas UVs, and dirty-chunk remesh for a live terrain brush.

What "a 20k world" actually means (units vs meters vs Source-1 map folklore, and they are ~1550x apart), why s&box has no VBSP-style coordinate wall, and the real modern ceiling: fp32 position precision that coarsens with distance from origin. Measured numbers, the lore separated from the measurements, and the reproduction method.

A decimated heightfield mesher renders carved cliffs as alternating sawtooth teeth; consistent diagonals and majority-vote corner snapping don't fix the crest. Sort the four corner samples, find the largest gap, and average only the cluster above it.

A stationary vertical sink at a chunk-corner seam defeats eject, mantle, and even hard-recover (the anchor itself can be buried under an overhang). The terminal escape is a top-out to the chunk surface.

A voxel/heightfield world usually carries two surfaces: a fine render mesh and a coarser collision heightfield. Procedural character-facing geometry (climb grips, standoff nodes) must be sampled from the collision surface, or it floats off the real face the character actually touches.

An idempotent ground-snap jitters on curved terrain because the slope curves away within each step: ease the downward snap at a bounded glue rate.

Lowering the sea level in a priority-flood water pass doesn't drain interior lakes: they're perched at their own spill surface. Use a per-basin depth gate and a land floor for below-sea valleys instead.

A Sandbox.Decal projects onto runtime-built chunk meshes (ModelRenderers) and conforms to stepped/curved faces, confirmed live on voxel terrain.

When a continuous-float render surface sits over a quantized-step collision mesh, every quantize boundary becomes an invisible vertical wall the player hits but cannot see.