perf(FragmentsModels): fill tiles in size-sorted sample order to cut visibility runs

[rihokirss]

Part of #278.

generate() fills tiles by walking samples 0..N (file order), while fetchLodLevel decides per sample from its screen size. The two orders are unrelated, so the geometry / wires / invisible cut through a tile lands as dozens of interleaved visibility runs — and every run becomes a geometry.group, i.e. a draw call.

_samplesDimensions already exists and is already sorted big → small (it is the order updateTiles sweeps in, built in processSamplesDimension() during init(), which runs before generate()). Feeding tile construction that same order keeps samples of similar size — and therefore of the same LOD level at any given camera distance — adjacent inside a tile, so the cut is one or two runs.

Tile membership, tile ids, geometry and the per-sample classification are unchanged; only the order of samples inside a tile's index buffer differs.

Measurements

Headless Chromium, Radeon 780M, 1920×1080, postproduction off, 50 FPS cap, medians of 3 runs.

	before	after
draw calls, 1 model, idle / orbit	5 619 / 4 691	4 101 / 3 035
geometry.groups, 1 model	3 670	2 356
FPS, 1 model, orbit	37.1	40.5
draw calls, 8 models, idle / orbit	6 783 / 6 456	4 408 / 3 843
geometry.groups, 8 models	4 527	2 439
render CPU ms, 8 models, orbit	42.2	33.3
FPS, 8 models, orbit	22.7	28.2
long tasks per 5 s idle, 8 models	818 ms	206 ms
settle after orbit, 1 model	685 ms	514 ms

Rendered content is the same; the screenshot diff is ≤0.08 % of pixels and is entirely one façade strip where two coplanar surfaces z-fight — the tile that arrives first wins, and this changes tile order. No difference anywhere else.

This composes with the compacted-index PR (which removes the remaining runs on the main thread): together the 8-model scene runs at 1 817 draw calls and the 50 FPS cap.

---

The companion PR for #278 is #281 (main-thread index compaction). They are independent — either can go in alone — and compose.

🤖 Generated with Claude Code

https://claude.ai/code/session_013rtNQqhSQRM2t6E98DESNE