VoxelForge/Standalone/Research/WorldAssemblerM3/Linux32/REPORT.md

VoxelForge L32 report

Cause and fix

The old Gradient transition wrote constant solid-density values per band. Those values are not signed distance to the displaced boundary, so Marching Cubes crossed band levels as contour-like terraces. The hash surface quantization was not the cause: sampling across the final FI floor gives density -0.53345 at z=-121.50 m and +0.46655 at -121.25 m; the SW ceiling gives +0.01514 at -126.50 m and -0.98486 at -126.25 m. Both cross linearly at about 4 density units/m.

L32 restores the width/8 Gradient displacement (2 m each side, about 4 m peak-to-peak at 16 m owner width). Density is now signed distance to the union of the displaced solid intervals, continuous across band boundaries. Cap samples follow the same displaced faces. Passage-mouth fades use cubic smoothstep across a 2 m fade from the protected mouth radius; this preserves the face while blending the column flare. Version/evaluator is v20.

Surface and interface audits

Branch-locked radial and straight profiles pass on both faces at LOD0 through LOD8. Every tested LOD has at least 3.15 m height range; no repeated-height plateau or following step was found. LOD0/1 profiles use 0.125 m samples (SW LOD1 also checked at 0.0625 m); LOD2–8 use 0.5 m samples. The largest adjacent-normal change was 17.50° on the fine SW LOD1 scan, below the 20° audit limit. Peak RSS for the final Python ring scans was 373,596 KiB (FI LOD0); the passage-fade scans peaked at 52,128 KiB. Mesh scans use tile/chunk processing.

Interleaved also passes 11 radial/straight profiles at each LOD0–8: all selected windows have full coverage, no flats or steps, and at most 0.367° adjacent-normal change. Their local height ranges are 0.032–3.174 m; the 3 m minimum is the Gradient amplitude check, not an Interleaved requirement. Peak RSS for this tile-streamed sweep was 24,468 KiB.

Both ring negative controls fail as intended: L30 has 0.742 m range, a 1.875 m flat run followed by a 0.241 m step, and a 27.34° normal jump; L31's flat fix has only 0.000046 m range. The passage-mouth audit passes on both overlapping faces: upper max normal change 15.06°, lower 7.96°, with full coverage and no flats or steps.

Hole/interface and minimum-band audits pass for Hard, Gradient and Interleaved, for both recipes and compilers: zero hole columns and zero air samples. Wider-than-declared and thinner-than-minimum controls are rejected. Minimum bands are 0.5 m FI, 1.0 m SW and 6.25 m lower band; maximum Interleaved crossing is 6.487 m against the 16 m bound. Passage audit passes all 16 recipe-stamped rows; recipe-mismatch and disconnected-end controls are detected.

Determinism passes 24/24: the L25 command and 24 m crossing, both recipes, Clang and GCC, workers 1/8/12.

Captures and hashes

Tint-on and tint-off JPEGs are in captures/; all are under 30 KB (limit 300 KB):

• fi-floor-near-tint-on.jpg, fi-floor-near-tint-off.jpg
• fi-floor-far-tint-on.jpg, fi-floor-far-tint-off.jpg
• sw-ceiling-near-tint-on.jpg, sw-ceiling-near-tint-off.jpg
• sw-ceiling-far-tint-on.jpg, sw-ceiling-far-tint-off.jpg

They show the FI floor and SW ceiling from inside each stratum, with grazing light, at LOD0 and LOD8. Capture SHA256 values:

Capture	SHA256
fi-floor-far-tint-off.jpg	239b6be70c177e267d2db95f206fdcdc8834779a7948fdaaddd136b4163d1f34
fi-floor-far-tint-on.jpg	3ecefd555275036a7045a541f02bb1873e22302495009dbd80b4ac58ce716b58
fi-floor-near-tint-off.jpg	847f10acfb420c2bbded61bdee7f578cde19469777b2ef4c5a0eba08b406611c
fi-floor-near-tint-on.jpg	4a5d3e3f8e8254639fa1840a2fb4497f2ab19f64bb4734e13abe3c3cfcfb4008
sw-ceiling-far-tint-off.jpg	f5ec099076c507bed5a703302d89b665e139804a8f8895a68c1d5705aabf9f61
sw-ceiling-far-tint-on.jpg	b82618814c8320be2de1ac53ba10e9ab4469299a8b53b776be32c6ca9a659988
sw-ceiling-near-tint-off.jpg	5789ce59a7748aa278c1b95edae45aec87721721260c1a3aefe6c4cc1e47cd7a
sw-ceiling-near-tint-on.jpg	aa4351c5effc5e93fc0879257fe056b3f4625e7842f5118f28588fb2c5c1afe2

Final Core binary SHA256:

• Clang: 052694a569c92ef407c25bd9a534c7123ef800f46733e5c428e965c32d905fdc
• GCC: c8656901ce18861f12168849165785937a2495233604314df9e005ed1e92afdd

Cost and edits

Five interleaved pairs per recipe against L31. Median whole-view-ready time was SW 1.723 s → 1.756 s (+1.92%) and FI 16.963 s → 17.221 s (+1.52%); paired results were mixed and within observed run noise. No Unreal files changed. The five edited files are the Core public header, FarShellsInternal.h, ResolvedLayoutInternal.h, ShapeWorldInternal.h and LinuxOffscreenCapture.cpp; Unreal compilation was not available here.

Failed and retried runs

• The initial multi-LOD Python audit was stopped by system RAM pressure (reported peak unavailable). Subsequent mesh audits ran with CPU/address-space caps and streamed or selected tiles; peak RSS is recorded above. The first whole-mesh FI LOD8 candidate scan was interrupted and replaced by a targeted tile scan.
• Early ring-window searches and narrow-height audits selected the wrong one of two overlapping faces. The FI LOD3/4 branch probes measured 1.423 m amplitude / a 1.0 m flat run; the SW LOD1 low-face probe reached 20.19° normal change. Face-locked, broad-window scans replaced them and pass. The exploratory ring-*, ringset-*, search-* and window-search-* logs preserve rejected window candidates. The L30/L31 failures above are intentional negative controls.
• The first Interleaved layout/export probe omitted the runner's required streaming flags and exited with usage; the retry produced the LOD0–8 export. Its broad 0.5 m candidate scan applied the Gradient 3 m window heuristic and rejected LOD0–4 (one LOD1 candidate switched overlapping faces at 62.92°); selected branch-locked profiles at those LODs and passing LOD5–8 windows show no plateaus or steps. An unsupported --help probe also returned usage with exit 2.
• The exact-source passage audit at midpoint preference switched faces and reported a 47.79° jump; upper- and lower-face locked reruns pass. An early wrong-camera export selected zero tiles; the exact camera export was rerun. Initial multi-LOD capture meshes mixed neighboring shells, so the final captures use focused LOD0/LOD8 surface patches.
• One fine-normal probe had a tuple-index error and was corrected. A SW far-mesh filter had an argument/layout parsing error and was corrected. The density CSV was overwritten by verbose output; the probe was rebuilt from final Core objects and rerun.
• The transition-cache build failed on a nested Shape initializer; the retry built, but the cache showed no gain and was removed. /usr/bin/time was unavailable for one capture attempt; the capped renderer was rerun directly. The first determinism run was interrupted and resumed; the final matrix passes 24/24. One log-search command used unavailable rg; a capped streaming-Python scan replaced it.
• Performance experiments were rejected or superseded: initial transition path +36.11% SW/+7.18% FI, planar query +14.89%/+2.28%, 3x3 coverage +6.06%/+1.65%, 5-point coverage +4.04%/+2.03%, single-transition fast path +3.52%/+1.39%; cache trial was slower (2.188 s vs 2.171 s). The retained hash-prefix version is the five-pair result above. Logs and exports are under Saved/Linux/l32/.

Orchestrator review (Claude)

• Merged. Captures checked by me: the FI floor (near, tint off) is a smooth rolling surface with no rings or terraces; the SW ceiling and far views are smooth. Gradient is restored at width/8 (about 4 m peak-to-peak) with signed-distance density; the ring audit rejects both L30 (rings) and L31 (flat). The memory-pressure kill mid-round was resumed; audits now run capped and tile-streamed.
• Cumulative SW cost vs L24 is roughly +5% after L28–L32 (features: columns, transitions, smooth faces). FI is ≈ +2%.