[clark719664]

Why

Lucent kept the GPU at 100% whatever the settings: with the camera still it raised samples per frame until each frame took ~24 ms, so lowering resolution or bounces just let it pack in more samples. Left open in a tab, it never stopped.

What changes

• GPU budget. Work is scheduled against a share of wall-clock time with a token bucket: time earns GPU credit, and each frame only starts the samples it can afford. Per-sample cost comes from EXT_disjoint_timer_query_webgl2 when available, otherwise from a synced re-measure every 2 s.
• Presets. Eco / Balanced / High / Max set budget, frame-rate cap, resolution, bounces and a stop-at sample count. Any edit shows as Custom.
• Stop when converged. At the stop count nothing is traced, and the canvas isn't redrawn unless something visible changes.
• System scan. Times a few samples of the current scene (at half resolution first, so a slow GPU never stalls long), scales the cost to the full canvas, and applies the recommended preset. It runs once on a first visit, waits until the canvas has a real size, and saves the result to localStorage.
• Pause when the window loses focus (on by default except in Max). The HUD and stats show GPU load and cost per sample.

Measured

Windows per-process GPU counters (Claude app's GPU process), RX 7600, Showroom at 1280×720:

Setting	GPU load
Before this PR	84%
Balanced (50% budget)	44%
Eco (25% budget)	26%
Converged (stop count reached)	0.3%

The scan on this machine: 37.5 ms per full-resolution sample → Balanced (75% resolution, 50% budget, stop at 2048 spp).

Checks

• lint, typecheck, test (79, including a scheduler simulation across per-sample costs from 0.5 to 120 ms) and build pass locally.
• Bundle: 48.6 kB gzipped JS.

Merging deploys GitHub Pages.

🤖 Generated with Claude Code