Vantage phase 1b: a planet from a seed

What shipped

npm run bake turns planet.json, which is a seed plus tunables, into the complete texture set. Domain-warped FBM continents with land-biased ridged mountains. Biome albedo from latitude, altitude and moisture climate belts. cos(lat)-corrected Sobel normals. Procedural continent-scale clouds. Black-Marble-style clustered city lights. And the region and terrain data map, grown by cost-weighted multi-source flood-fill out from seeded capitals.

PlanetDefinition loads all of it from data, so ?planet=test-moon swaps in an entirely different world with zero code changes.

The game now orbits Vantage Prime, not Earth. That was the actual bar for this phase: not “the pipeline runs” but “the fictional planet looks as good as the NASA photography did.”

Decisions

Sampling 3D simplex on the unit sphere rather than 2D noise on the unrolled equirect is the decision everything else rests on. It makes seams and poles non-problems by construction, because longitude +180 and -180 are literally the same point in 3D. The proof that it worked is negative and I like it: there is zero seam special-casing anywhere in the pipeline. No blending band, no pole fixup, no “and then we fade the last few columns.”

The one build-plan amendment is that baked output gets regenerated rather than committed. What’s in git is the script, the seed, and a 512px preview. I changed this on a measurement, not a preference: a 4k set is about 16 MB of PNG that delta-compresses essentially not at all, and that lands in history on every single art iteration, forever. What makes it safe rather than reckless is the byte-determinism, which is verified rather than assumed.

The night lights call came out of review rather than planning, which is the right way round for an art decision.

What broke

Round 1 of review rejected the art outright, and it was correct to.

The night lights splatted as soft blobs, described in review as glowing cotton balls, which is exactly right and exactly the opposite of the Black Marble look I was after. The clouds were scattered puffs. And the coastlines were uniformly crinkled confetti at roughly 50% land, which reads as noise rather than geography.

Every fix was bake-side, which is a good sign about the architecture. A point-scatter rewrite of the lights. Weather-mask frequency halved. Continent frequency lowered with warp raised. Sea level up until land hit 33%. And a connected-component pass that sinks island specks under 30 texels, since a thousand tiny dots is what makes coastlines read as static.

The determinism gate quietly earned its keep through all of this. Every art iteration re-verified byte-identical double bakes, so I never had to wonder whether a visual change was the tweak I made or noise in the pipeline.

Numbers

The 4k bake is 10.6 s end-to-end with workers on the height stage, against a 60 s target. All seven outputs SHA-identical across double bakes, re-verified independently by the reviewer rather than taken from my run.

14 tests green. The region-sampling gate was proven non-vacuous by breaking the flipY reconciliation on purpose and watching three tests fail with the exact expected signature, which is a northern capital resolving to a southern region. That’s the failure I wanted to see, not just any failure.

Land fraction prints at 33.08% against a 28-35% target.

Next

Phase 2 is the ship on its inclined track: orbit camera, time warp, raycast picking that turns a click into lat/lon and then into the region data this phase produced, and the mission-control minimap with its sinusoid ground track.