WASM Terrain Bench

Four implementations of the same Minecraft world generation pipeline: JavaScript (deepslate, the parity oracle), Rust → WASM, C++ → WASM (emcc) and AssemblyScript → WASM. Same seed, same region, biomes + heightmap + trees + flowers. Four previews side by side show whether the variants agree deterministically, and the table below gives the raw timings.

How the benchmark works

All variants get an identical TerrainInput (seed + center + radius + MC version + features flag).
JS uses deepslate as the parity oracle. Rust/C++/AS are currently v0 heuristics, so their checksum will not match JS until they are replaced by a port of pumpkin-world.
The variant with the shortest time gets a green border.
The L1 cache (LRU, 4096 chunks, ~16 MB) keeps data between runs. The "Clear cache" button resets it.
A worker pool keeps JS off the main React thread, so the UI doesn't freeze at radius 128.
Versions 1.18-1.26: worldgen JSON bundles are lazy-loaded from /data/worldgen-X.json. Missing bundles are greyed out.

Building the WASM artifacts

See apps/web/wasm-terrain/BUILD.md. Short build.sh scripts in the rust/, cpp/ and as/ directories generate files into /public/wasm/terrain/. The bench marks a variant as "build needed" if its artifact is unavailable.