Context

The streamer drew a 3×3 block of patches at one level around the point under the camera. That is a few kilometers of ground on a body ten thousand kilometers across: the horizon was the datum sphere, every cube-face edge was a hole (five patches of nine over a corner, where three faces meet), and the whole set faded out an octave above the ground because a lone raised tile on a sphere is a sticker rather than a planet.

Worse, the rules that chose the level and the fade were handed distance − radius. For a camera standing on the ground that is groundElevation + height, not height. The terrain rig measured what followed: a summit streamed one level coarser than a basin at the same height above the ground; a level pass at fixed height coarsened and re-refined as the ground rose and fell beneath it; and a summit above radius · 2^(5.5 − maxLevel) was not drawn at any altitude, including zero — two of Miranda's six survey sites were ground that could not be looked at.

The milestone this serves is "one continuous space, orbit to on foot", so the representation has to be one representation at every distance rather than two with a fade between them.

Decision

A quadtree walked once per frame from the six cube faces, with four rules, each of which was chosen against a measurement rather than a preference.

The error is a patch's grid spacing, not its height deviation. Ulrich's screen-space-error predicate wants the mesh's true vertical deviation from the surface. On a planet that number is startlingly small — Earth's relief is two parts in a thousand of a cube face and a patch's 64 quads cut it by another 64 — so a height-error metric alone says a 156 km patch is close enough to stand on. Cesium's shipping tiles carry a geometric error equal to their own sample spacing for the same reason, and that is the number used: the size of the smallest thing a patch can express. Refine while one grid cell subtends more than 16 px.

Refinement stops where the field stops, measured per body. Past some level a patch is a bilinear upsample of its parent. On Mercury a level-9 patch differs from one by 12 cm and levels 10 through 12 by nothing a float can hold — so the old rule's saturation at level 12 was asking for sixteen times the patches of level 10 to produce identical output, at 12.8 ms of worker apiece. surfaceDetailFloor measures the residual from the field itself — 24 golden-angle probes, five samples each, memoized, about five milliseconds, counting a quiet level only when its stencils touched dry ground, because the sea clamp manufactures exact zeros over open water — and lands at level 9 or 10 across the zoo. It lives beside elevationAt because it is a property of those bands: when Phase 2's geology puts detail at scales they do not currently reach, it reports a deeper floor the same day.

Distance is measured to the ground. A node is a cone of directions crossed with the shell [radius − relief, radius + relief] that ground can occupy, and the distance is to the nearest point of that. An eye inside the shell is at zero, which is what standing on the ground means, and the three defects above are gone with the datum they came from.

Levels are morphed and the tree is restricted to 2:1. Each vertex slides toward the position its parent's grid holds for it and arrives there before the parent takes over (CDLOD, Strugar 2009). That is exact rather than approximate: a child covers half its parent's side, 64 quads halved is 32, so every even index of the child lands on a parent grid point, and the field is a pure function of direction. The morph therefore closes a one-level gap perfectly — and a two-level gap not at all, because the finer patch arrives on a grid the coarser one has no vertex on. Unrestricted, that was 30 of 468 patch edges on Miranda with a gap of two or more and a worst case of six, drawn as dashed black arcs along every level ring. The tree is restricted.

Terrain is not drawn where the sphere is the honest picture. A body's relief must cover more than eight pixels first. Past that the mesh and the datum sphere are the same picture, and the sphere already carries a normal map and, on four bodies in Sol, a photograph.

Alternatives considered

Geometry clipmaps. Viewer-centric nested grids, no tree, no per-patch bookkeeping. They fight tile addressing: a clipmap's rings are not regions, so nothing they draw has an address, and addresses are what regions, seeds, scatter and the save format all hang off. Rejected.

Concurrent binary trees (arXiv:2407.02215, KSP2's terrain). The end-game, and the right answer once generation itself lives on the GPU. Rejected now, with a revisit condition: if draw submission ever dominates the frame rather than generation does.

An unrestricted tree, with the morph carrying the transitions. This was the plan of record, and it is wrong. The morph closes one level. Stated honestly the choice is between a wide LOD band, which balances the tree by construction, and a restriction pass — measured, the tree balances itself once a level's band is 3.7 patches wide, and that costs 500 to 1,000 patches for one disk against 300 to 600. The band stays narrow and the pass runs, in packed integer keys with no allocation because the readable version cost 1.8 ms, sixteen times the traversal it was correcting.

Per-node cube-face distortion correction (Zucker & Higashi, JCGT 2018). The gnomonic map is not equal-area — a cell at the middle of a face covers over twice the ground of one at a corner — so measuring each region describes a patch better than its level does. It also breaks the thing the phase is for. The crack-free argument needs a patch and its coarser neighbor to agree on one number, and those two are measured at different points on the face, where the scale differs by up to 22% at level 2: the finer patch ends up 15% short of its neighbor's grid, which is a lit gap. The distortion is smooth in position and the level is not, and no per-node number can be both. The metric is nominal per level; what the distortion costs is over-tessellation near the cube's eight corners rather than a seam.

Keeping the opacity fade. It existed because a 3×3 window is not a planet. With whole-disk coverage there is nothing to fade, and the fade was also hiding a real bug: the surface tier reaches more than eight radii while NEAR_LIMIT is two thousand kilometers, so patches placed at true meters against a compressed sphere are a different object at a different distance. Terrain now rides the body's own render compression.

Switching terrain off for mapped bodies, as the terrain plan's scope carve-out reads. surfaceRadius is one function and the contact test lands a ship on procedural elevation whether or not the body has a photograph — Mars's is ±14.7 km against a sphere drawn 29.4 km under the datum, so a mapped body with no streamed ground is a ship parked fifteen kilometers above a smooth planet. The carve-out is about what may be claimed of a mapped surface, not about whether the ground under the landing gear is drawn.

Consequences

The three datum defects are gone, and every one of the zoo's twenty-four survey sites bottoms out at its own detail floor. The tests that pinned them assert their opposites.

A frame selects a few hundred patches where it drew nine. Measured at 16 px a cell: 236 patches on Earth at two meters, 435 on Miranda, 474 on the zoo's 11,536 km world, and 623 at the worst step of a descent onto it. Selection is 0.11–0.31 ms against a 0.5 ms line. Standing on Miranda's summit the whole frame is 2.04 ms at 63.9 fps with 3.59 M triangles selected.

Terrain's memory is now the notable cost. 203 KB a patch, so 45–126 MB of vertex buffers. Packing the four attributes below float32 is worth about half and frustum-culling the selection about half again; neither is done, and both are named in the roadmap.

A cache smaller than the working set oscillates rather than degrading. The streamer holds two selections — the drawn one and the request one, taken from where the eye is going. At a flat 512 against a working set of six hundred, terrain strobed at every altitude.

The selection's ceiling is not what sizes those caches, and reading it that way is how the second one was set too small. What must be resident is the set the evictor promises to keep, which is the whole pyramid under the request selection — and a quadtree's ancestors are a third again as many as its leaves, so that floors at ~1.33× the selection cap before the starved rung is counted. Geometry sized at the selection plus a slack of 128 sat under that floor: the streamer built four patches a frame and dropped four it had wanted a moment earlier, and every twenty-sixth frame the eviction took a patch the traversal was refining through and the disk snapped from 760 patches at level 7 to four at level 1.

No multiple of the selection cap bounds that keep set, and the cap in force is a measurement rather than a proof. The request set is two independently capped selections — the drawn one and the one taken at the look-ahead eye — so its ceiling is around 2.3× the selection cap in the limit, which is more than the geometry cache holds. The two coincide at a hover and separate as the camera's ground track lengthens, so what moves the keep set is the camera's speed as much as the drawing buffer: measured over Luna, Ganymede and Triton it runs from 957 at a hover over 1600×900 to 1,824 at a 20 km lead over 5120×2880. Twice the cap clears everything measured, by about 11%. Anyone retuning either number wants that sweep again rather than the arithmetic.

The sphere-tier shell wanted a material, and the eight-pixel threshold is where it went. The plan's unconditional level-0–2 shell needs a per-face normal and albedo bake underneath it; without one it is flat tinted ground, and drawn over a mapped body it replaces a measured picture with an invented one. ADR-0020 gave the ground below the threshold a material and left the bake unbuilt, so a mapped body agreed across the switch and a generated one did not. ADR-0026 closes it from the other side: a mapless body's sphere is six faces of the ground material's own reflectance and relief, drawn from the streamer's own patches, so both sides of the threshold are the one graph and neither is a stand-in for the other.

Streaming — the four rules in practice
Rendering — the morph, and why the shader's share is one mix
ADR-0003 — the compression terrain now rides
ADR-0005 — the versioning the detail floor answers to
ADR-0017 — the lens the predicate reads, and every patch count above re-measured through it
ADR-0020 — the material the ground this selects is drawn with
The terrain plan — the milestone this is Phase 1 of

Context

Decision

Alternatives considered

Consequences

Related