Ashfall Station

The concept is the contract

Seven sheets came before any authoring — key art, an orthographic turnaround of the rover, three zone sheets, a prop sheet and a material callout board. Twenty-three named elements became inventory rows, every row a real Blender asset. The pack is also where the honest notes live: the key art paints the rover dust-brown while the turnaround paints it orange, so the turnaround was declared authoritative and the key art kept only for mood.

Key art — the fissured plain, the station on its hill, the volcano still working

The same station live in the engine: baked cabin, radome, guyed weather mast, the rover on its pad

Nine rounds of gray boxes

The whole drive ran and was won as flat-gray primitives before any Blender time, and the blockout is where this piece earned its physics. Nine defect rounds, each traced to a root cause: a mirrored wheel axle that made the left and right sides fight each other and held a 20 kN rover to walking pace; a crater rim the rover simply drove over at 16 m/s and fell out of the world; and a stop at the first sensor that pitched the rover onto its roof, where a six-wheeler has no traction and can never recover. The fixes — a real rim wall, a low centre of mass, and a self-righting assist — all ship inside the finished game.

Blockout: 55 obstacle colliders, telegraphed bomb rings, the HUD bearing indicator

The finished drive across the same terrain — the blockout's collision geometry, unchanged

One mesh, drawn and collided

The ash plain is a single Blender-displaced grid, and the engine builds its physics trimesh from that mesh's own vertices at load rather than from a second copy of the height function. Collider-visual parity is true by construction: the shipped selftest measures the worst vertex against the authoring formula and reports 0.00 cm. The rover's wheels get the same treatment — their rendered tyre radius is measured rotation-invariantly (an axis-aligned box measures tread phase on a lugged wheel, not radius) and fitted to the collider radius to 0.00 cm.

The turnaround sheet — the contract for the hero asset

Audit round 4: a 19.2k-tri detail sculpt baked onto the 5.2k-tri shipped hull

Rapier vs cannon-es, measured

The piece exists to answer one question, so the answer was measured rather than asserted: the same scene shape in both engines — this piece's own terrain on the same 110 × 110 grid, the same deterministic spawns, the same 1/60 fixed step, 600 steps with 60 discarded as warmup. The cannon-es heightfield column is included deliberately, because comparing only against its known-weak Trimesh would have been a strawman. Two bugs in the benchmark harness itself were found and fixed before the numbers were trusted; both are written up in the ledger.

Contact quality was a wash — Rapier is not more accurate here, it is one to two orders of magnitude faster for the same result. The surprise was tunnelling: dropping a 0.55 m ball straight into the same terrain, Rapier held at every speed up to 1208 m/s with continuous collision detection switched off, while cannon-es punched through at 72 m/s and again at 403 m/s even with four substeps. On this piece CCD is cheap insurance; Rapier's speculative contacts are what actually do the work. The shipped selftest asserts the requirement — no tunnelling with CCD on — and merely logs the CCD-off result, because an assert that "CCD off must tunnel" would have encoded a wrong assumption.

Measured, honestly

Real engine features, earned