Context

Most engines that span large scales use a hierarchy of reference frames to rescue floating-point precision: coordinates are kept small by being relative to a nearby parent. Having chosen sectorized coordinates (ADR-0001), precision is already solved everywhere in the universe, which changes what frames are for.

Something still has to answer: what does "3 m above the launch pad" mean while the planet orbits at 30 km/s and rotates underneath at 465 m/s?

Decision

Frames form a tree. Each non-root frame has an anchor that is either

fixed — an absolute UniverseVector, used by star systems, or
dynamic — a pure function of simulation time returning a pose relative to the parent, used by orbits, rotating bodies and landing sites.

Composition applies the transport theorem, so a child inherits the tangential velocity of a rotating parent. A point at rest in a body-fixed surface frame is therefore moving at orbital-plus-rotational speed in universe axes without anything integrating it.

reframe(state, targetFrame, t) re-expresses a state in another frame and provably preserves canonical position and velocity — the numbers an entity carries change, where it is does not. Frame changes are consequently invisible to gameplay, which is what lets an approaching ship be moved from a system frame into a planet frame mid-flight.

Three frames exist per body, and the distinction is load-bearing:

Prefix	Frame	Used by
`b:`	body-centered inertial — translates along the orbit, does not spin	satellites, approaching ships
`bf:`	body-fixed — spins with the body	terrain, anything bolted down
`sf:`	local tangent at one lat/lon, +Y up	meter-scale gameplay, landing

Alternatives considered

Frames as the precision mechanism. The conventional design. Unnecessary here, and it would have forced every entity to belong to a frame close enough to keep its coordinates small — a constraint on gameplay, not just on math.
Absolute positions only, no frames. Would require recomputing an orbiting body's absolute position every tick for everything attached to it, and would lose the velocity composition entirely.

Consequences

Frame-local Vec3 coordinates are precise near their own frame and lossy far from it, because a Vec3 is a double. Expressing a point in a frame four light-years away degrades to meters. There is a test that documents this limit rather than hiding it; it is why canonical state is a UniverseVector and why approaching ships are re-framed.
Ships integrate only in non-rotating frames. Integrating in a rotating frame without Coriolis and centrifugal terms is wrong, and adding them is a lot of subtle code for one case. A landed ship is attached kinematically to a surface frame instead and is not integrated at all.
Surface frames are minted on landing and regenerated from their id on load. The id therefore has to determine the frame completely — angles are quantized to 1e-6 rad and the ground elevation is derived from the quantized direction, not passed in.
Pose resolution is cached per instant, because a tick resolves the same handful of frames for every entity in them and an orbital frame's evaluator runs a Kepler solve.