Which geometry kinds take the partial-extend tolerance, in one place.
deriveNodeViewState widens a node's slice tolerance across dimensions the
node only PARTIALLY extends through. Lines opt out: their segment bounds
already encode the non-displayed extent, so applying it again double-counts
during clipping. Every other kind opts in — a mesh has no per-element bounds
at all, so there is nothing to double-count.
This is the data half of GEOMETRY_DESCRIPTORS, split out into a
module with no runtime imports so leaf handlers and refinement loops can read
one boolean without pulling in every node loader and loader factory. The
descriptor still exposes the flag
(GEOMETRY_DESCRIPTORS[kind].applyPartialExtendTolerance) and now sources it
from here, so there remains exactly one definition;
tests/unit/data/scene-loader/partial-extend-tolerance.test.ts pins the two
together, and fails on any NEW hardcoded site.
Twelve call sites used to hardcode the boolean, against a table documented as
the single source of truth that only two of them read. Hardcoding is not a
compile error and the wrong value is not a crash — it is a clipping result
that is quietly too generous or too tight, which no test would name.
Which geometry kinds take the partial-extend tolerance, in one place.
deriveNodeViewStatewidens a node's slice tolerance across dimensions the node only PARTIALLY extends through. Lines opt out: their segment bounds already encode the non-displayed extent, so applying it again double-counts during clipping. Every other kind opts in — a mesh has no per-element bounds at all, so there is nothing to double-count.This is the data half of
GEOMETRY_DESCRIPTORS, split out into a module with no runtime imports so leaf handlers and refinement loops can read one boolean without pulling in every node loader and loader factory. The descriptor still exposes the flag (GEOMETRY_DESCRIPTORS[kind].applyPartialExtendTolerance) and now sources it from here, so there remains exactly one definition;tests/unit/data/scene-loader/partial-extend-tolerance.test.tspins the two together, and fails on any NEW hardcoded site.Twelve call sites used to hardcode the boolean, against a table documented as the single source of truth that only two of them read. Hardcoding is not a compile error and the wrong value is not a crash — it is a clipping result that is quietly too generous or too tight, which no test would name.