2,386 words
Status: accepted · 20 Oct 2026. Amends the whole-disk selection of ADR-0015, the long end of the lens in ADR-0017 and the Earthrise clamp in ADR-0018.
Context
The field-of-view slider stopped at 20° because the terrain predicate saturated its patch cap there (ADR-0017). Earthrise at the photograph's framing — Earth a sixth of the frame, and Earth is 1.9° across from Luna — wants 11.4°, so the preset had the right geometry and an Earth too small in it.
The predicate refines while a patch's cell subtends more than cellPixels at
the lens's pixel angle, and ADR-0015 applies
that to the whole visible disk at one scale. At the flight lens that is the
right shape: the frame is most of what a standing eye can see. At a telephoto
it is not. A pixel angle is a claim about what is inside the frame, and an
11.4° frame is a sliver of the disk; refining all of it at that angle spends
the cap on ground behind and beside the camera.
Measured standing at two meters over every survey site of Earth, Luna, the
four zoo bodies, Gliese 908 IV and the far-shore world — sixty eyes — and over a
32-step descent onto each basin, at 1920×1080 display pixels and 16 px a cell,
with the cap at 1,280 and its coarsening ladder (COARSEN_STEP):
| Lens, whole disk | Standing, uncapped | Standing, cut at the loosest rung | Descent peak |
|---|---|---|---|
| 65°, flight | 456–1,061 | 0 of 60 | 704–917 |
| 20° | 932–4,815 | 0 of 60, coarsened on 40 | 1,187–2,610 |
| 11.4° | 2,063–12,330 | 8 of 60, all short underfoot | 2,769–6,389 |
Raising the cap was rejected in ADR-0017 and is rejected again for the same reason at a larger number: a patch is 270 KB of buffers, and 12,330 of them is 3.3 GB for a lens the player has narrowed onto a subject.
Decision
Past the flight lens, the selection refines at the lens only where the frame can reach, and at the flight lens everywhere else; with that, the long end of the lens is 11.4°.
selectTerraintakes aview— the camera's forward in body-fixed axes. With a lens longer thanREFERENCE_LENS, which is the flight lens, a node whose ground can reach the frame's cone refines at the lens and every other node at the reference. The cone is the frame's half-diagonal widened byVIEW_MARGIN_PIXELS, 1,024 of the lens's own pixels. Without a view, or at the flight lens and wider, the selection is the whole-disk one field for field: a property test holds that, and 180 eyes return the same 39,821 patches with and without the change.- The ground a node can hold is bounded by a capsule along its axis, from
(radius − relief)·cos θtoradius + relief, thickened by the column's widest half-width. The relief is a vertical claim — a summit can rise into the frame and a canyon can fall into it — and the capsule keeps it vertical. - A node is in view only if its parent was. The scale then never rises down
the tree, which is what the
MORPH_ENDhandover argument assumes. - Where the two regimes meet, the patch in view takes its morph band at the reference scale. A coarser neighbor outside stopped at the reference's range, a fraction of the in-view patch's range at the lens, so a band taken at the lens would still be sliding at the neighbor's edge. At the reference scale it has arrived on its parent's grid first, exactly as within one regime.
- No patch within a turn of the frame borders coarser ground outside the
view. A patch banded at the reference has a band its same-level neighbors
in view do not share, and two equals sliding their shared odd vertices by
different bands open that edge. The margin does not keep such a patch out
of the frame: one near its own range is 1,024 px across, but one at the
level floor is kept because it may not go deeper — a level-16 patch on Luna
is 41 m wide, over 2,000 px at 100 m through 11.4°. Measured on the live
streamer at 11.4° standing on Luna's map seam, such edges lay in the frame at
4 of 12 headings, open by up to 3.9 cm, 12 px; headlessly they reached the
frame on 2 of 48 standing eyes on Luna and 6 on Mars. So the balance pass
splits a node outside the view wherever an edge neighbor holds a finer node
within
VIEW_TURN_PIXELSof the frame: that patch meets the outside at its own level and keeps the lens's band, and every same-level edge the two regimes still disagree on is at least 564 px out after a full turn of the held selection, over a thousand without one. - The cap bounds the balanced tree. Two regimes grade across the frame's edges, and the 2:1 rings that do it are hundreds of patches the walk's own count never sees: a descent onto Luna at 11.4° whose every walk fitted came back over the cap on 22 of 32 steps. An overrun after the balance pass climbs the ladder like an overflow during the walk.
- The engine owns the camera's pose.
GameEngine.cameraPose()is the pose the frame is drawn from — the script's, the observatory's or the walker's, orflightCameraPoseon the ship arm — and both the streamer andCameraRigread it. The streamer turns its render-space forward into the body's axes through the rotation its patches are placed with. - A held selection is retaken once the view has turned
VIEW_TURN_PIXELS, a quarter of the margin, and never at the flight lens or wider, where the view is not an input to the walk. It is the same distance the boundary's seams are kept from the frame, so no turn a held selection survives brings one into the picture. FOV_MINis 11.4°. It is the photograph's framing, and it is where the cap holds: every eye looking at the horizon fits, and the only cut trees are three of sixty eyes looking straight down beside a cube corner.
Alternatives considered
Raising the cap. Above, and in ADR-0017. It buys the ground behind the camera.
A bounding sphere for the in-view test. The obvious construction, and it does not work: its radius cannot be less than the body's relief, 9.9 km on Earth, so every node within ten kilometers holds the eye inside its sphere and is in view — behind the camera as much as in front. Standing on Earth's basin at 11.4° looking at the horizon that left 3,289 patches; the capsule takes the same eye to 1,756, against 5,504 for the whole disk.
A frustum against the datum sphere alone. Tighter than the capsule and wrong in exactly the pictures a telephoto takes: ground the relief lifts into the frame would refine at the reference, so a summit rising into the frame is drawn nearly three levels coarse until the camera turns toward its base.
A finer or coarser reference. Measured on the same eye, the reference lens barely moves the count: 0.15 to 0.67 of the flight lens's scale gives 1,708 to 1,726, the flight lens 1,756, 2.25 times it 1,969 and 3.4 times it 2,595. The ground outside is set by 2:1 grading across the level gap, not by the reference's own predicate. The flight lens wins on the other count: nothing outside the frame is ever coarser than the flight lens would draw it, so a turn streams the telephoto's increment over ground already resident and never a hole.
Banding at the reference wherever a patch in view borders coarser ground outside, and nothing else. It closes that edge and opens the patch's edges to its equals, in the frame wherever the patch is a floor patch near the eye.
Splitting along the whole boundary rather than within a turn. It removes every disagreement between equals in view, and costs 13% more patches standing at two meters across Earth, Luna and Mars at 11.4°, where the turn-wide rule costs 2% and keeps every remaining one out of a turn's reach.
Widening each patch's in-view test by its own angular size. It brings the floor patches near the eye wholly into view, so none of them borders the outside in the frame, and costs 21% more patches and 20% higher peaks.
No margin. Cheaper — 1,477 patches at no margin, 1,555 at 512 px and 1,756 at 1,024 on Earth's basin at 11.4° — and it puts the ring of demoted patches in the frame's edges, where every one of them draws a level coarse.
Retaking the selection every frame of a turn. Two walks a frame for ground the margin already covers. Measured at 11.4° on the basin, shore and corner of Earth, Luna, Miranda, Gliese 908 IV and the far-shore world, a turn of 256 px asks for at most 33 patches the selection before it did not hold — four frames of the streamer's eight-a-frame build budget, with 768 px of margin left.
A floor at 10° or 8°. At 10° Earthrise frames at its solved angle at every instant of the month, where 11.4° clamps it for about half the orbit by up to 5%. It also doubles the cut trees straight down to six of sixty, one of them away from any cube corner, with overruns to 21%, and puts 117 scatter regions on Luna against a ceiling of 96. At 10° one of 240 horizon eyes cuts, at 8° fourteen, and a descent onto Miranda looking down is cut on 16 of 32 steps.
Tightening the capsule with each node's own elevation range. The lever for the nadir case: a node's column is the body's whole relief tall, so a camera looking straight down refines a disk as wide as a canyon the body's relief could hold. The selection runs before a node's field exists, so the bound it would need is a pyramid of elevation ranges the streamer does not keep. Named here as the next lever rather than built.
Consequences
Measured over the same sixty eyes and eight descents, looking at the horizon (four headings a site) and straight down:
| Lens | Horizon, standing | Cut | Descent peak | Cut | Nadir, cut |
|---|---|---|---|---|---|
| 20° | 687–2,318 | 0 of 240 | 961–1,388 | 0 of 256 | 0 of 60 |
| 11.4° | 950–3,494 | 0 of 240 | 1,314–1,946 | 0 of 256 | 3 of 60 |
| 10° | 1,031–3,792 | 1 of 240 | 1,410–2,132 | 0 of 256 | 6 of 60 |
| 8° | 1,156–4,499 | 14 of 240 | 1,607–2,465 | 0 of 256 | 14 of 60 |
| 11.4° at 8× | 3,319–17,991 | 240 of 240 | 4,958–7,517 | 164 of 256 | — |
The cap holds at the new floor everywhere but beside a cube corner. The
three cut trees at 11.4° are all the corner site looking straight down, over
the cap by at most 12%, and none of them is short of the floor underfoot. A
corner is where regionSpacing's nominal spacing over-refines by up to a level
(ADR-0015), and a long lens multiplies every ring. On a 1600×900 window at DPR
2 the horizon cuts 41 of 240 eyes, 26 of them at corners, by up to 49%, still
none short underfoot; the same window at 20° over the whole disk cut 8 of 60
standing eyes, every one of them short of the floor.
The zoom channel is still uncovered. At 11.4° with the zoom racked to 8× the frame wants several times the cap even inside itself, and the disk goes coarse by a stated amount on nearly every step, as ADR-0017 said it would at 20°.
A turn costs a few frames of detail, never a hole. Ground swinging into the frame was refined at the flight lens and sharpens as the telephoto's increment streams. The memo retakes the walk every 256 px of turn.
Keeping the boundary's seams a turn from the frame costs a little of the cap. The extra splits are 2% of a standing selection, and at the cap that is enough to move 15 of 240 horizon eyes one rung up the coarsening ladder, 184 coarsened where 169 were; an uncapped standing walk costs 1.40 ms where it cost 1.16.
The headless probes ask a different question. ir.descend and
ir.terrainBaseline have no camera to aim, so they refine the whole disk at
whatever lens they are handed; ir.terrain().outside counts the patches the
live streamer refined at the flight lens instead.
Everything the long end implies moved with it. 120.22 mm on the 24 mm
gauge and 5,410 px/rad over 1080 display pixels. The hyperfocal distance with
the zoom racked out is 13.9 km over a 1520 px buffer, so FOCUS_MAX is 30 km
rather than 10; the smallest rock is two pixels at 1,353 m, which is 88
scatter regions on Luna against the ceiling of 96; a 100 px drag at 8× without the lens's own
sensitivity would sweep twenty field-widths.
The billboard threshold follows the lens to 3.08e-5 of angular radius, and the one body it moves in Earthrise's sky is Saturn: seen from Luna at that instant it subtends 4.36e-5, a point at 20° and a billboard at 11.4°. With the zoom racked out the threshold is 3.85e-6 and the body that crosses is Neptune. That is still a plate review rather than a unit test, as the terrain plan says.
Earthrise takes the field it solves. 11.53° at the preset's instant, with Earth 380,396 km away, and the vendored plate is taken at that angle.