VoxelForge/Standalone/Research/WorldAssemblerM3/Linux143/REPORT.md

L143 — liminal polish

Captures

Road

Eye-height:

Wide:

Looking up from the cloud floor:

Pillar grid

Eye-height:

Wide:

Kept from L142

These four captures are byte-identical to L142.

Result

The road is one continuous swept deck, and the distant pillar silhouettes remain complete and consistent. The monolith and periodic hills were retained unchanged.

Changes and causes

• The road edge broke where independently capped deck segments overlapped. The road now resolves to one swept-deck stamp with a shared cubic-Hermite centerline and continuous cross-section; the centerline is sampled at 1 m spacing (361 points). The 8 m deck is 376.829 m long and retains 17 support posts. The sweep path is serialized in stamp-vocabulary V5.
• Pillars disappeared or narrowed when the coarse marching-cubes lattice skipped their cross-section. The pillar stamp's interval proof is unsupported, so it conservatively kept tiles but did not protect the mesh from undersampling. At LOD4 the sample spacing is 32 voxels (8 m), already greater than the 5.5 m pillar radius; spacing rises to 64 and 128 voxels at LOD5 and LOD6. The shared meshing path now emits deterministic 24-sided capped-cylinder proxies for regular fillPillar stamps once the sample step exceeds their radius or half-height. The same rule applies outside the liminal fixture. The radial approximation error for this fixture is about 4.7 cm, below one 25 cm voxel.

Measurements

• Road walk audit covered the full 376.829 m route at 1,081 surface stations, with 68,103 capsule samples and 7,096 support samples: 0 blocked, 0 unsupported, 0 grade failures, and 0 surface misses. Measured maximum grade was 8.512° against a 15° limit; deck width was 8 m against a 3 m minimum. road-short-negative-control failed the walk audit as expected at 210.351 m.
• Determinism covered five fixtures—monolith, periodic hills, road, pillar grid, and far-LOD pillar grid—across Clang/GCC and W1/W8/W12 (six compiler/worker cells per fixture). All five reported identical outputs.
• Clang and GCC builds passed on source stamp 1baa3a711f7f96df2e70fc93579f41ad0b0211eb42fcb51dabf522e33d888877. The capture runner and capture-quality checks passed for all requested views.
• Fast tier passed in 292.3 s. Its pinned L130-v227 reach-unit owner hashes match exactly: 12/12 rows for each owner fixture, with W1 and W12 identical. Historical L46/L53 hash comparisons remain diagnostic-only and are not the current owner-hash gate.

Checks and limits

• Passed: road audit and short-road negative control; five-fixture Clang/GCC determinism; Clang and GCC builds; capture quality; fast tier; git diff --check.
• Not run: Unreal build (no Unreal Source/ files changed) and L80 performance timing (this round made no performance claim).
• The coarse proxy is limited to regular fillPillar stamps; irregular pillars and other stamp kinds continue through their existing SDF path. The proxy was exercised at LOD4–6 in the pillar-grid captures.
• Unreal Source/ files changed: none. Changed standalone source files: Standalone/Core/include/VoxelForgeCore/VoxelForgeCore.h, Standalone/Core/src/ResolvedLayoutInternal.h, Standalone/Core/src/ShapeWorldInternal.h, Standalone/Core/src/VoxelForgeCore.cpp, Standalone/Source/StreamBenchmarkHeadless.cpp, Standalone/Source/StreamBenchmarkPipeline.h, and Standalone/Source/WorldAssemblyPrototype.h.

Detailed logs are in Saved/Linux/l143/; the capture and audit transcript is Saved/Linux/l143/logs/capture-runner.log, the walk audit is Saved/Linux/l143/logs/road-walk-audit.log, and the fast-tier summary is Saved/Linux/l143/fast-summary.json.