Function sortedIndexNode

Double-buffered draw-slot → storage-slot index — the TSL twin of GLSL_SORTED_INDEX (depth-sorting spec §2.1 tier 3).

A permutation must swap ATOMICALLY: a half-applied ordering is not a reordering but a corrupt permutation (elements drawn twice / not at all). Each new ordering therefore streams into the INACTIVE attribute across frames, and uSortedIndexSlot flips only once that buffer holds the whole permutation.

BOTH attributes are referenced unconditionally, which is required, not merely tidy: the WebGPU backend only uploads graph-referenced attributes, so an unreferenced back buffer would never receive its chunked uploads — and RenderObject dereferences a referenced attribute before its undefined guard, so every geometry must carry both. attachElementStorage allocates them as two DISTINCT buffers together — never aliased, never added later, since the WebGPU vertex layout is cached from the attribute set at first draw and a set that grows afterwards renders the scene black.

uSortedIndexSlot must stay a RUNTIME uniform: a compile-time flag would rebuild the graph on every swap. (The lines material treats uIsOrtho as compile-time — deliberately NOT the pattern here.)

BRANCHLESS on purpose. .select() emits an if/else STATEMENT, and the pick factories consume this index at varying(float(...)) — evaluated OUTSIDE their Fn() body, where a statement cannot legally land. That produced a shader which built and code-generated fine but rendered zero pixels (the same vacuous-black failure mode the gsplat harness comment in tests/e2e/harnesses/tsl-harness/gsplats.ts warns about). A pure arithmetic mix is position-independent, so the one helper is safe both inside and outside an Fn(). The multiplier is a uniform, so there is no per-vertex divergence to save by branching anyway.

Returns an INT node (not uint): the downstream int(...) wraps and float(...) casts accept it unchanged.