VoxelForge/Standalone/Research/WorldAssemblerM3/Linux66/REPORT.md

VoxelForge L66 — Linux report

Findings and fixes

• The saved SW passage range is ReachMax=5000 owner voxels, despite the C++ type default of 96; why the export carries 5000 is unresolved. At 0.125 m per owner voxel, the sampled lower reach was 318.539 m. With FI content top at −504.5 m, the old lower mouth resolved to −823.039 m, 79.539 m below FI content bottom. The planner selected the config from the upper SW stratum and did not cap that reach to FI’s band. This was not a metre migration error; L64’s prose value of −822.06 m was not reproduced by the saved layout. Core now caps each sampled mouth reach to its own usable content interval. The new 0–1 lower mouth is −737.000 m, 6.5 m above FI content bottom, and strict layout validation passes.
• The Unreal adapter’s old SW search tried 648 fixed 64 m candidates; all were rejected on grade. The adapter now tries a denser bounded search with shorter lengths for a single-mouthed SW route. In the four-strate HEAD spawn, 35 candidates were eligible and the selected route passed with 0 blocked, 0 unsupported and 0 grade failures. The 33-point SW arrival-to-departure audit also passes with 0 unsupported points. L14 still cannot run this SW route.
• The passage interfaces remain the main failure. Their unsupported samples are concentrated in the composed passage floor even though the underlying native fields report 5/5 support probes at the corresponding segments. A four-strate grounded-tread experiment did not reduce these counts and introduced blocked samples in controls, so it was reverted. The 0–1 rock column also falls short of its relief threshold. At 1–2 the sampled floor jumps upward by 0.2498 m at 109.280 m along the passage; the source of that one-Core-voxel offset is not resolved.
• Internal-chain floor/support blending reduced 2–3 floor ripple from 3.57 cm to 0.10 cm and 1–2 from 4.78 cm to 2.90 cm. The 1–2 floor and wall ripple gates still fail. No full quality pass is claimed.

Four-strate audit

Live input SHA-256: f42f96adca58a4ebd680b8553b2969d6cd9ca8a73f875f042bfbc7b8c649fdca. The generator and style maps were derived from that export: SurfaceWorld, FloatingIslands, TunnelNetwork, VerticalShafts; passage styles {0:1, 1:1, 2:1, 3:1}.

Audit	Before	After L66
Strict layout	Rejected 0–1: lower mouth −823.039 m, below FI	Accepted; clamped reach 318.539 → 232.500 m; lower mouth −737.000 m
SW route	24/33 points unsupported; Unreal spawn sweep failed	33 points, 0 unsupported, 0 grade failures; mouth binding passes
Transitions	Not accepted with the rejected layout	3/3 pass
Per-strate generation	Not audited	Each generator produced 27 tiles; W1/W8 each produced 12 hashes and matched
Passage chain	0/3 walkable	0/3 walkable; detailed results below

Interface	Unsupported / blocked / grade failures	Wall ripple	Floor ripple	Other result
0–1 SW→FI	59 / 0 / 0	1.49 cm, pass	0.08 cm, pass	Column relief 0.172 m < 0.20 m; route support 46/46, wall margin 1.75 m
1–2 FI→TN	3 / 0 / 0	6.78 cm, fail	2.90 cm, fail	0.2498 m floor offset begins at passage distance 109.280 m
2–3 TN→VS	2 / 0 / 0	2.28 cm, pass	0.10 cm, pass	Ripple passes; unsupported support probes remain

The fast tier checks arrival-to-departure for each stratum through its adjacent interfaces. Those four gates fail because the interfaces above are not yet walkable. The standalone route audit for strate 0 (SW) passes.

Per-strate W1 cost (whole_view_ready, worker-seconds): SW 1.015 s / 0.839 s; FI 1.424 s / 1.254 s; TN 0.558 s / 0.388 s; VS 0.208 s / 0.043 s. Each run generated 27 tiles.

Unreal trace — L14 vs HEAD

L14 could not run the four-strate plan, and its compact 16-chunk SurfaceWorld route also failed the fixed-length search. Both versions therefore used the same transient 16-chunk FlatPlain trace plan copied from DA_Strate1; DA_Settings and DA_Strate1–DA_Strate4 hashes matched between hosts. Three interleaved 45 s traces completed per version. No Unreal build ran during the timing trials.

Measure	L14	HEAD
Cells per tile	16³ = 4,096	16³ = 4,096
Desired tiles LOD0–8	512/448/320/208/144/96/96/96/306	same
Unique generated keys LOD0–8	1264/896/464/284/156/96/96/96/306	1272/896/464/284/156/96/96/96/306
Generated records LOD0–8 over trace	1355/928/464/284/156/96/96/96/306	1363/928/464/284/156/96/96/96/306
Worker-seconds by LOD0–8	4.962/2.985/1.078/0.751/0.441/0.323/0.322/0.316/1.063	5.168/3.007/1.138/0.786/0.454/0.262/0.326/0.320/1.118
Total worker-seconds	12.200	12.561
LOD0 worker p50 / p95	3.324 / 7.351 ms	3.604 / 7.313 ms
generateTile / mesh copy / RMC build	11.980 / 0.036 / 0.179 s	12.358 / 0.032 / 0.167 s
Initial walk gate	9/9 ready at 0.244 s	9/9 ready at 0.249 s
LOD0 ring	512 tiles ready at 0.827 s	512 tiles ready at 0.823 s
Full set	2226/2226 at 3.326 s	2226/2226 at 3.309 s
Median trace wall time	45.037 s	45.027 s

Core generateTile accounts for 98.2% of measured L14 phase time and 98.4% for HEAD. Mesh copy is 0.29% / 0.25%; RMC build is 1.46% / 1.33%. Each version passed 12 ready gates through movement; the full-set trace had 125 collision tiles desired and 78 ready at the recorded completion point.

Verification tiers

• Standalone/data/regression/owner_world_2strate_l61.json pins the L64 two-strate export; SHA-256 7b095f9bcc6b90133d5a463f99f0b571133b2c549ce1ecf2d64fe9175e41e183. All 60 fast-tier runs that consume owner data used this fixture. Legacy gates, including styles, passage audits, W1/W12 determinism, L52 hash comparisons, rings and controls, passed. L46 hash comparisons remain non-gating diagnostics and differ; L53 diagnostics and gated L52 hashes match.
• verify.py reads the live four-strate generator/style expectations from owner_world.json; legacy helper scripts receive VOXELFORGE_OWNER_DATA so they use the pinned fixture. The fast summary is FAIL only because the new-plan stage records nine failed interface/column/ripple gates and then fails its final stage gate. Generation, W1/W8 determinism, strict layout and transition checks pass.
• Standalone/Tools/verify/README.md documents the two input sets and deferred fast-tier result.

Unreal source files for the owner's Windows build

• Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp
• Source/VoxelForgeEditor/Private/VoxelForgeExploreCommandlet.cpp

Both Linux L14 and HEAD host builds completed successfully. These are the two Source/ files for the owner's Windows build.

Failed runs

• An initial L14 trace used an output path outside the plugin Saved directory; the commandlet rejected it.
• The L14 compact SurfaceWorld attempt found no eligible 64 m candidate. The common FlatPlain plan was used for the paired timing traces.
• The first fast-tier rerun exposed hard-coded live-owner paths in the ring and motif fixture helpers. Both were switched to the pinned input; the subsequent run passed all legacy gates.
• The final fast tier remains red on the four-strate gates listed above.
• The grounded-tread support experiment compiled, but left unsupported counts at 59/3/2 and added blocked control samples; its source change was reverted.