1,654 words

Status: accepted · 6 Sep 2026. The preview gesture is superseded by ADR-0036; the camera descent remains in effect.

Context

The planetarium has two camera arms and one way between them. ADR-0018 put the orbit arm above 1.5 radii and the surface arm below it, and stand is how the camera crosses: it cuts. That is right, and its own docstring says why — the surface arm is the instrument a plate is captured through, and an ease means every capture waits an unspecified number of frames for a filter to settle before the picture is the picture.

What a cut cannot be is a thing to watch. Arriving on a world is the most photographable act the mode has and it happened between two frames.

The other half is that there was no way to choose where. surveySites finds six places on any body — the highest ground, the lowest, the shoreline, the steepest, the addressing corner, the pole — because a seeded world has no place names and typing a latitude into a sphere lands on the same undifferentiated mid-slope every time. Six derived sites is a good answer to "somewhere interesting" and no answer at all to "there, that spot, the one I can see".

Decision

A figure dragged from the chrome onto a drawn world, and a ballistic entry that the camera flies to where it lands.

The trajectory is the conic, and its periapsis is under the ground

packages/rendering/src/entryArc.ts is a Keplerian conic about the body's center through the release point and the touchdown, with its apoapsis at the release. Its periapsis lies below the surface, which is what makes it an entry rather than an orbit, and drawing that continuation is what says so.

The record holds 1 − e, not e. A drop from far out onto ground nearly under the eye has e within 1e-13 of one, and the polar form's denominator 1 − e·cos ψ written as a subtraction loses every digit it has there. Written (1 − e) + 2e·sin²(ψ/2) it is a sum of two positive terms and loses none. A touchdown directly under the release has e exactly 1: the ellipse has collapsed to a radial line, and that is the correct answer, drawn as one.

The schedule is logarithmic in height, and the radius is not monotonic

The band from orbit to eye height is six decades. A linear descent spends 99.9% of its time above the altitude where terrain is drawn and then arrives in a frame — the same argument the surface arm's own scrub is built on. Measured on Earth from 30,000 km, the halfway frame is at 1.7 km rather than at 15,000 km.

The camera's radius rises where the ground rises under it: the conic is solved against the ground below the touchdown, the track crosses ground that is higher, and the stance's height clamp lifts the eye over a ridge rather than flying it through one. Bounded by the body's own relief — 9.9 km on Earth, against the 12 m a test track actually meets.

The camera watches the ground, then turns to the star

For the first two thirds the aim is the touchdown point, which is the picture: the ground coming up. Over the last third the heading swings to the star's bearing and the pitch rises to the horizon, so the frame the descent ends on is the world it landed on, lit from where the light is.

The heading is also eased toward the star as the aim steepens. The bearing of a point nearly under the eye swings with every meter of sideways travel, and that is loudest exactly where the descent is fastest. Landing on the night side faces the star's bearing with the star below the horizon, which is the honest answer rather than a special case.

The aid is scene geometry, in the body's own frame

It was an SVG in the HUD first and no amount of styling fixes that: an overlay is flat. It cannot be occluded by the limb it crosses, its dashes do not shorten with distance, and a ring it puts on the ground is a circle rather than the ellipse a circle on a sphere is.

And it is placed the way a terrain patch is, not the way a point is. A body is not drawn where its metric position says — render compression (ADR-0003) pulls it nearer and shrinks it so its angular size survives, and placement.scale is the radius it comes out at. A point put through its own compression lands at a different depth from the sphere it is meant to lie on: the ground ring sank inside the planet and vanished. So the observatory answers in body radii, in body-fixed axes, and the drawer hangs the aid off the placement the body was drawn with. One unit is the drawn surface by construction, and the buffers are camera-independent.

The figure is held at a fifth of the viewer's altitude, above the aim

At the viewer's full orbital radius the mark is 3.3 body-radii out at Earth, which leaves a 65° frame exactly when the aim reaches a limb — where a horizon comes from and where anybody composing a picture aims. A fifth puts it a half-radius clear of the ground: outside the disk near a limb, so the fall is drawn side-on, and still in frame. The cost is stated rather than hidden: aim at the middle of the disk and that vertical points at the camera, so the fall foreshortens.

Both rings are sized by angle rather than in meters, because the gesture spans six decades and no fixed size survives that; the ground ring follows the drawn terrain and floats a twelfth of its own radius over it, because a fixed lift is a handful of float32 bits at a planetary radius.

A ray that misses answers with the limb

The near limb is the one part of a sphere a pointer cannot land on from outside — the ray grazes it at a tangent, and a pixel either way is the difference between a hit and the sky. Refusing there makes the last few degrees of every drop unreachable. A miss answers with the point on the surface nearest the ray, which is continuous across the limb: the mark slides onto the edge and stays there rather than blinking out.

Nothing canonical is written

The drop is Observatory.drop, it rides the surface arm, and standing is true from its first frame — which is what keeps the orbit writers refused for the whole descent, so a wheel notch mid-drop cannot rewrite the state ascend returns to. ir.ascend() abandons one in flight at the framing it left. observatory.test.ts compares world.stateHash() across a drop.

Alternatives considered

Ease stand instead of adding a verb. One arrival, one behavior, less code. Rejected on the reason stand cuts in the first place: it is the plate rig's entry point, and pnpm presets:plates would then capture an unspecified frame of a filter settling. Two verbs, two intents.

A click rather than a drag. The camera already spends every click it has — a click in the sky focuses whatever it hits — so this would cost the mode its primary verb, or hide the gesture behind a modifier nobody would find. Picking the figure up is an unambiguous statement of intent and nothing is decided until it is let go.

Land where you point, with the figure on the same ray. The obvious reading, and it draws nothing: the release and the touchdown are then on one line through the eye, and a line pointing away from the camera foreshortens to a point. The figure has to be somewhere the camera can see it beside where it lands, which is what holding it on the local vertical does.

An overlay aid, with the arc projected to screen. Shipped first and replaced. It cannot be occluded, and it made the ring a circle on a plane rather than an ellipse on a sphere. The measurements that killed it are in the commit; the geometry that replaced it is above.

Consequences

Good.

Arriving on a world is something to watch, and the frame it ends on is the picture the gesture promised.
Any spot on a drawn body is reachable, including the limb, which is where a horizon comes from.
The aid is in the scene, so it is occluded, perspective-correct, and free during a frame where the hand held still.
ir.drop(lat, lon) is the console's own verb, so a capture script composes the same descent a hand does.

Costs, honestly.

The aid draws the figure's fall, and the camera flies its own arc. They end in the same place and take different paths there — the camera leaves from where it is, the figure from where it is held. The ring is a promise about the destination, not about the camera's route.
Aiming at the middle of a disk foreshortens the fall to almost nothing. It is geometry rather than a bug, and the mark stays legible because both rings are angular.
The drop is wall-clock timed rather than tied to the tick, like every other presentation ease. Pausing the simulation does not pause a descent, which is the same choice TRAVEL_TAU already makes and for the same reason.
The gesture has no keyboard equivalent. ir.drop and the Ground section's site buttons are the two ways in without a pointer, and neither lets a keyboard choose an arbitrary spot.

ADR-0018 — the two arms this crosses between
ADR-0003 — the compression the aid is placed through
ADR-0021 — the drawn radius the stance and the ring stand on
ADR-0011 — the camera precedence this stays inside
Planetarium — the mode's own account