[mbeczynski]

Measurement (simulator, A/B old main 147f911 vs new main a6f231d)

Interleaved A/B/A/B, 5 regions × 2, live 45 s races with nitro:

• Frame time / hitches: unchanged — p95 16.8–16.9 ms in both (simulator is vsync-capped at 60 fps), hitches > 25 ms: 320 vs 320 in total, CPU work per frame p95 0.36–0.42 vs 0.38–0.45 ms (noise).
• Race loading: +0.75 s (~9%) in every pair. Track build 0.68 → 1.29 s: every scenery material loaded the Metal library and built its CustomMaterial from scratch (45 hazy materials + road, walls, ground, shopfronts).

Changes

• SceneShaders: one Metal library, one template per surface shader; each material is a copy (a copy shares the compiled shader). Later races in the same session compile nothing.
• The five scenery shaders compile at the start of the loading screen, while textures load in the background.
• tools/scenery-shots.sh: absolute output path (simctl launch --stdout with a relative path fails with "No such process"); rotate only portrait screenshots.

Result (city + forest, 4 runs each build)

	new main	old main	delta
before this PR	9.32 s	8.57 s	+0.75 s
templates only	9.25 s	8.54 s	+0.71 s
templates + warm-up (this PR)	9.31 s	8.78 s	+0.53 s

Track build itself is now 0.25 s (old main 0.68 s); the warm-up shares the main thread with car loading, so the first race gains ~0.2 s net. The remaining ~0.5 s comes from the larger texture set (asphalt, new impostors) and first-time shader compilation.

Tests

• swift test: 61 tests, 0 failures; Debug / Family / Release builds succeeded
• Fixed-frame screenshots (city, snow): no visual change

🤖 Generated with Claude Code

https://claude.ai/code/session_018YYRkMvxEa4GxWLW2SrVyF