VoxelForge/Standalone/Research/WorldAssemblerM3/Design49

Design49 — flat landing pad at tunnel ends (owner idea, reviewed 2026-11-02)

Owner: "the end of the tunnel should be flat; flatten the arrival with a fade-out that blends to the surrounding density; a cheap stamp applied at arrival, no matter what's there."

A read-only design review (Claude Opus 5.5, medium effort) of wip/l134-cave-planner recommends a clamp, not a lerp. Each tunnel end gets a level disc at the tunnel's end floor P. Out to a blend radius, the native ground is squeezed between two gentle cones (12°) starting at P: a lower one that fills solid, an upper one that carves air. Inside the disc the floor is exactly flat; outside it, the slope is never steeper than max(cone, native), with no overhang or cliff. The full note is in the session transcript; key points:

• Placement: the last step of ApplyResolvedStrateComposition, after the passage composer, on both the native and non-native branches. The fill is masked by the tube-only distance (bore stays open). It replaces the landing-chamber box, shell and floor-shell special cases that caused L152–L156.
• Formula (metres; ×cos α for 1-Lipschitz). d = max(0, r−R), t = tan α.
  • Fill: Lo = P − t·d, F = max(z−Lo, (Lo−Ds)−z, r−(R+B)).
  • Air: Hi = P + t·d, A = min(z−Hi, (Hi+H)−z, (R+B)−r).
  • Result: c = max(min(native, max(F, −tubeDist)), A).
  • Hard min/max inside R; smooth-min width ramps 0 → 0.5 m over the first metre outside it.
  • Why not lerp: native density saturates at about ±4.125 voxels (DetailSupremum), so a lerp makes cliffs and films.
• Parameters: α = 12°, R = max(mouthRadius + 1 m, 3 m), B = min(8 m, |Δh_ring|/t + 1 m), H = 3 m, Ds = 1.25 m.
• Planner:
  • P = h(end), with h′ = 0 over the last R.
  • Sample the edge ring; move P into the feasible band or take the next deterministic site.
  • Rewrite route heights inside R+B to the clamped ground.
  • Content-band check including P−tB−Ds and P+tB+H.
• Bounds: the pad influence box goes into the proof bounds, the fallback overlap test, the tile feature bounds (critical: the composer early-returns on an empty route list) and the proof-miss sampler.
• Audits: new LandingPadAudit (flatness ±2 cm, support, headroom, 36 radial blend transects ≤ 15° with one crossing) plus negative controls (pad off, 2% tilt, B halved, bore mask removed).
• Risks: caves under pads (connectivity), low ceilings, shaft footprints, island edges and their air gap, liminal decks; off-route edges on non-heightfield layers are a metric only.