Summary

Renderer.clear was allocating a fresh { color, depth, stencil, flags } object — and a fresh [r, g, b, a] color array — on every call. The method runs per camera per layer per frame, so this added up to sustained per-frame GC pressure.

Hoist both to module-scope scratch buffers (matching the _temp* naming used elsewhere in this file) and write field values in place. Behaviour is unchanged: the WebGL, WebGPU, and Null clear implementations only read from the options object, they never store it.

Technical details

// before — fresh object + fresh array on every call
device.clear({
    color: [camera._clearColor.r, camera._clearColor.g, camera._clearColor.b, camera._clearColor.a],
    depth: camera._clearDepth,
    stencil: camera._clearStencil,
    flags: flags
});

// after — module-scope scratch reused across calls
const c = camera._clearColor;
_tempClearColor[0] = c.r;
_tempClearColor[1] = c.g;
_tempClearColor[2] = c.b;
_tempClearColor[3] = c.a;
_tempClearOptions.depth = camera._clearDepth;
_tempClearOptions.stencil = camera._clearStencil;
_tempClearOptions.flags = flags;
device.clear(_tempClearOptions);

Verified that all three clear(options) implementations only read fields and do not retain a reference:

• `webgl-graphics-device.js#clear` — reads `options.flags`, `options.color`, `options.depth`, `options.stencil`
• `webgpu-graphics-device.js#clear` — passes options through to `clearRenderer.clear`, which only reads
• `null-graphics-device.js#clear` — empty

Measured impact

Captured via V8 sampling heap profiler (`node:inspector` `HeapProfiler.startSampling`) on a representative scene rendered against `NullGraphicsDevice`:

Public API changes

None. The options shape passed to `GraphicsDevice.clear` is unchanged; only the underlying memory is now reused.

Test plan

• Existing unit tests pass (`npm test`)
• Visually verify a basic scene clears correctly on WebGL2 and WebGPU