Goals

Move the RAD splat renderer/demo toward Spark 2.x Coit Tower parity while keeping this PR reviewable as the renderer-side follow-up to the SPZ/RAD loader work.

Use Spark as the reference experience for Coit Tower: same dataset, Spark-like FoV/LoD/foveation budgets, and an explicit side-by-side capture so the remaining visual/performance gap is easy to discuss. @ibgreen feedback requested on the renderer architecture and remaining parity gap.

Actual changes

• Adds the RAD splat renderer integration, render-page compaction, LoD traversal, and deck-layers test coverage.
• Adds worker-backed RAD chunk decoding so page loads do not block the UI thread as heavily.
• Adds active-row WebGPU projection/sort paths for paged LoD frontiers instead of rebuilding/sorting full resident pages every frame.
• Tunes the Gaussian splats Coit defaults toward Spark: FoV 75, LoD 1.5 budget, Spark-style foveation controls, sqrt(8) Gaussian support, larger residency budget, lower request pressure while moving, and settled/interacting request cadence.
• Adds tile-sort overflow telemetry and frontier diagnostics for selected parent rows, leaf rows, max child count, expanded opacity, decoded scale, page count, draw count, requests, evictions, and upload time.
• Preserves RAD color range decoding and premultiplied alpha blending in the render path.
• Adds the current Spark-vs-loaders Coit capture as a PR review asset.

Current Coit comparison

Current loaders.gl capture metrics from the screenshot run:

• 3,749,995 / 50,937,127 splats
• 60,609 ms
• 264 chunks | 58 pages | 3,749,995 draw | 0 req | 18 evict | 0 ovf | 591 ms

This is intentionally not called complete parity yet. The tower is more resolved and the tile overflow path is controlled at this page size, but the background/city still does not match Spark and the coarse LoD appearance is still visibly wrong. The next useful review target is whether the remaining gap is mostly LoD frontier selection, parent splat covariance/projection, color/alpha evaluation, or sort/blend behavior.

Reproduction Steps

1. Install dependencies from the repo root:

yarn

2. Start the Gaussian splats example:

yarn --cwd examples/website/gaussian-splats start --host 127.0.0.1 --port 5177

3. Open http://127.0.0.1:5177 in Chrome with WebGPU enabled.

4. Load the Coit Tower RAD URL:

https://storage.googleapis.com/forge-dev-public/asundqui/rad/260217/coit-40m-sh1-lod.rad

5. Use the Spark-like demo settings:

FoV: 75
LoD: 1.5
Scale: 1.00
Behind: 0.20
Cone: 0.40

6. Let the camera settle for about 60 seconds, then compare the overlay metrics and visual result against the Spark 2.x reference capture above.

Validation

• yarn lint fix was attempted on this Windows checkout, but the existing wrapper exits 1 with no diagnostics through the git hook/lint path.
• Direct Biome format/lint over the touched RAD/splat files completed cleanly.
• yarn test-headless modules/deck-layers/test/splat-layer.spec.ts
• yarn test-headless modules/splats/test/rad-source-loader.spec.ts
• yarn --cwd examples/website/gaussian-splats build

The Gaussian splats build passes with existing Vite warnings about browser externalization for worker_threads/Node stream shims and large chunks.