Derive the stable semantic palette for one render subject.
Each semantic id supplies a hashed starting colour; allocation visits ids in code-unit order and probes occupied colours forward. An allocated colour therefore depends on the complete claim population when collisions or probing chains occur. Reordering scene.nodes preserves the palette payload, while adding claims can change existing colours through those chains.
This function emits palette metadata. It observes no rendered pixels and does not prove byte-identical GPU segmentation.
Pixels belong to exactly one entry: the drawable that paints them. A
building's logical layers, its spaces, boundaries and openings, paint nothing
of their own and are reached through owner, so a wall pixel resolves to its
element, then to the boundary that wall realizes, then to the room, then to
the building unit. An element that fills an opening is owned by that opening,
which is how a door prop is addressable as a door rather than as an anonymous
panel.
Simulated drawables are addressed the same way. A cloth panel, a planting
cluster and a bound water surface are held by no scene node, so they are
joined by the names their own viewer builders assign; without those names
every one of them would paint the reserved background and a segmentation
consumer would read a curtain, a fern bed and a pond as nothing at all.
Throws when the subject declares more entities than one bounded mask can
address. Silently dropping the excess would make a mask that segments a
different world than the one drawn.
Derive the stable semantic palette for one render subject.
Each semantic id supplies a hashed starting colour; allocation visits ids in code-unit order and probes occupied colours forward. An allocated colour therefore depends on the complete claim population when collisions or probing chains occur. Reordering
scene.nodespreserves the palette payload, while adding claims can change existing colours through those chains.This function emits palette metadata. It observes no rendered pixels and does not prove byte-identical GPU segmentation.
Pixels belong to exactly one entry: the drawable that paints them. A building's logical layers, its spaces, boundaries and openings, paint nothing of their own and are reached through
owner, so a wall pixel resolves to its element, then to the boundary that wall realizes, then to the room, then to the building unit. An element that fills an opening is owned by that opening, which is how a door prop is addressable as a door rather than as an anonymous panel.Simulated drawables are addressed the same way. A cloth panel, a planting cluster and a bound water surface are held by no scene node, so they are joined by the names their own viewer builders assign; without those names every one of them would paint the reserved background and a segmentation consumer would read a curtain, a fern bed and a pond as nothing at all.
Throws when the subject declares more entities than one bounded mask can address. Silently dropping the excess would make a mask that segments a different world than the one drawn.