Vantage phase 2: riding the orbit, and the sign flip the textbook hides

What shipped

Vantage plays like an orbit now.

The ship rides a 55-degree inclined track with a smoothed follow-cam composing the vast arc. , and . step time warp through 1x, 10x, 100x and 1000x, with the auto-return-to-1x stub wired for later. Clicking the globe resolves region and terrain through the exact body-fixed transform chain. The minimap draws the wrap-split sinusoid ground track with a ship marker and predicted pass windows. And b toggles thin cyan region borders drawn in-shader straight from the region-ID texture, default off, because the naked planet is the hero.

Decisions

The research round caught two traps before either became a bug, and they’re both the kind that would have been miserable to find later.

The first: the textbook ground-track longitude formula is east-west mirrored under our longitude convention. We use lon = atan2(-z, x), and the standard form assumes the other handedness. What makes this nasty is that the wrong version looks completely fine. The track sweeps across the globe at the right inclination with the right period, and your eye has no way to tell it’s reflected. It only dies when you check it against the region maps, where the ship is suddenly over the wrong continent. That’s a class of error I want caught by data every time, because eye-tests structurally cannot catch it.

The second: f32 shader time decays. At simSeconds = 1e5 the resolution has degraded to 3.9ms, which means after a long 1000x warp session the star twinkle would visibly freeze. Game logic stays f64 and the shaders get wrapped phase uniforms instead, cloud phase mod 1.0 and twinkle phase mod 1e3. Resetting the clock on warp exit was the obvious alternative and it hitches.

Picking went analytic for a similar precision reason. At glancing angles the 128x64 sphere’s facets deviate from the true sphere by more than a 4k texel, which is enough to resolve the wrong region near the limb. The closed form is both exact and cheaper.

What broke

Review rejected the first vast-arc framing, and the diagnosis was more interesting than the bug.

At follow-cam range, the additive atmosphere fresnel spans a wide grazing band. Stack that with bloom over backlit clouds and the limb blows out into a wall of white. The obvious fixes are global: turn down bloom, or turn down the lit clouds. Both were measured, and both dulled the standard-distance hero shot, which runs straight into the one hard constraint on this project.

So the fix became a camera-distance ramp instead: a smoothstep that fades atmosphere scale from 1.0 to 0.5 and lit-cloud peak from 1.0 to 0.92 as the camera closes in. The far framing came out byte-identical, which is the proof that matters. The near framing reads as a controlled bright rim rather than a blown one.

Numbers

Ground track verified to 1.1e-13 degrees against an independent matrix construction, against a 1e-9 target. Picking to 3e-14 degrees across 36 synthetic rays at three spin angles, with every region and terrain ID correct.

22 tests green. Reviewer-measured around 775 fps sustained uncapped at 1440p on the real GPU with a median frame of 1.3 ms, so the 60 fps gate holds with an order of magnitude of headroom.

Next

Phase 3 is the diorama system: a separate staged scene per surface tile, seeded generation, orbit-to-diorama transitions with GameState hash integrity, and the first autonomous agents, which are vehicles on splines kicking up dust.