VoxelForge/Standalone/Research/WorldAssemblerM3/Research26/first-pass.md

VoxelForge R26 — lattice generator family

Legacy inventory

Family	Generator and output	Per-sample work and bounds
Maze	BuildMazeStack composes solid rock, FLatticeCorridorSource, three-octave SDF roughness, and carving. The deterministic 3D lattice has origin-directed parent edges plus optional loops. Parameters include cell size, corridor radius, roughness, and boundary seal.	Hot path: floor to the containing cell and test up to 24 cached corridor capsules; roughness noise runs near the corridor surface. A cache miss rebuilds up to 24 edge decisions. Its box propagation scans up to 64³ cells and returns a geometric lower bound with an unbounded upper bound; oversized queries become unknown. No feature BVH. Parameters · source and bounds
Vertical shafts	BuildVerticalShaftStack composes rock, hash-placed full-height shafts and connector capsules, roughness and carving, periodic partial ledges, connector air/floor modifiers, and optional structural posts. Parameters include shaft spacing/radii, connector radius, ledge spacing/depth, roughness, and seal.	Hot path checks the local 3×3 shaft neighborhood and its connector capsules, then applies roughness, ledge, and connector operations. A cell-cache rebuild does the broader hash-grid and parent-link work. Box propagation scans up to 64² XY cells and also returns a lower bound with an unbounded upper bound; it has no BVH. Parameters · field · ledge and connector modifiers

These are code-level cost descriptions, not legacy per-sample benchmarks; I found no comparable timing for either generator.

Native design and shaft walkability

Add a native lattice recipe with immutable, stable-ID graph records resolved from (seed, strate, recipe version, cell, feature ID). Put the required spine between the actual resolved arrival and departure anchors first. Use a finite graph search to choose its cells and edges; A* is a standard option for this graph-search problem. Add branches and loops only after that spine exists. Attach branches to it, and give any navigable dead end a supported landing or a clear return path.

Represent each room, corridor, and shaft as one filleted void feature with a stable ID. Reuse L54’s analytic SDF construction: cap each void against its floor with smooth-max, then union voids; L54 used a 2.4 m smooth width and measured a 0.702 m composed fillet. Keep route floors out of roughness displacement. L54 field and fillet L54 construction details

Store dimensions, floor profiles, blend widths, and padded bounds in the native resolved layout, following the existing micrometre-based cavern records. Build a BVH over bounds that include blend, floor support, and any bounded roughness. Use exact primitive intervals where available, propagate sound intervals through the composed field, and return unknown when a bound cannot be proven. Reuse the cave candidate lookup, field-classification, and tile-shape-free path; the BVH culls absent features, while interval proofs decide whether a tile is all air or solid. Cavern records · BVH · interval classification · isTileShapeFree

The shaft route must not depend on falling. The historical note reports arrival-to-departure failures on 15 of 16 seeds. The code’s diagnosis explains the structural gap: a large connected origin-spine component still does not guarantee that shaft networks at the measured passage mouths join. A full-height air column is not standing support; periodic one-sided ledges and sparse horizontal links do not establish a continuous route. Connectivity diagnosis Landing-point constraints

For a walkable descent without custom gravity, make the shaft a view well around a continuous ramp: a helix or broad switchbacks between landings. Enforce |rise| / horizontal run ≤ tan(15°) on every route segment. For example, a 4 m rise needs at least 14.9 m of horizontal run; a 4 m-radius helix rises 4 m per 25.1 m turn, about 9°. Verify the final composed field with a capsule flood fill from the actual mouths, plus local grade, support, and headroom checks—the method L54 used for caves. L54 route audit

Cost and VR read

Performance class: medium; unmeasured. Use the measured L54 cave whole-view time (0.727 s) and L56 motif time (0.524 s) as comparison points, not forecasts. Caves · Motifs Capsule and cylinder fields are simpler per active feature than L54’s curved edges, but a dense lattice, many ramps, and broad mixed regions can outweigh that. Sparse BVH candidates and tight intervals are the main cost controls.

For an initial scale, try a maze pitch near 16 m, 3.5–5 m clear corridors, and a shaft spacing near 20 m; those pitch values match the legacy defaults’ documented Core scale. Repeat a consistent grid rhythm, then vary a few junctions and landmark bays so it reads as a maze rather than a featureless corridor set. Pair local junction markers with an occasional global landmark; a recent VR wayfinding study reports that landmark usefulness varies with navigation strategy. Wayfinding study

For shafts, expose the central void and repeated landing tiers, with visible handrails and human-scale references. Treat vertigo as an intended effect to assess, not a reason to drop the design: a 42-person VR height study found increased distress and physiological responses on an exposed platform up to 40 m, strongest among participants with greater height intolerance. Study Keep the walking route clear and forward-readable; Meta’s guidance recommends gentle slopes, distant walls, and spatial sound cues for orientation. Meta locomotion guidance

Slices and verification

1. Maze slice: a small multi-level lattice with the real arrival-to-departure spine, one optional branch, and one loop. Check deterministic layouts, interval soundness against brute-force samples, BVH candidates, tile-shape coverage, and final-field capsule reachability with grade ≤15°. Capture an eye-height junction view and a high oblique cutaway showing the route and repeated cells.
2. Shaft slice: two or three shaft cells joined by one helix or switchback descent. Run the same checks, then repeat the mouth-to-mouth walkability audit across the historical 16-seed matrix. Capture the ramp at eye height, a view across the shaft to its tiers, and a cutaway showing the complete route.
3. Density pass: add branches and additional levels only after both slices pass; repeat interval, determinism, route, and whole-view timing checks at the denser preset. Frame captures to show route support as well as scale.

Question for the owner: should the first playable slice be the maze or the shaft descent?