VoxelForge/Standalone/Research/WorldAssemblerM3/Linux5/REPORT.md

World Assembler M3 — Linux L5

Result: no octave tail is safely removable under the declared bounds. The exact far meshes already exceed the 1 px field-error and LOD-swap budgets, and the cutoff flag is 2.2% slower in the W12 view. The candidate therefore generated the exact same meshes; recipe version stays 1.

Build and machine

Built VoxelForgeCoreHeadless-linux5-l5-final7 and its capture target on assembler/world at 7770074 with Clang 19 -ffp-contract=off, then GCC 14 default contraction, both with 12 jobs. W1/W12 hashes match across both compilers; the L3 FP-contract preamble and runtime self-test remain active. Runs used CPUs 0–11, normal priority, and the 6 GiB process limit. Maximum process HWM was 898,692 KiB (GCC compile); the largest runtime process was 454,828 KiB. CPU governor stayed powersave. /dev/dri/renderD128 was absent, so EGL used Mesa llvmpipe with LP_NUM_THREADS=12. No packages or system settings were changed.

Capture and gate

LinuxOffscreenCapture.cpp reads the .vfmesh export from the headless streaming pipeline, creates a surfaceless EGL/GLES3 pbuffer, and writes 2560×1440 PNGs with libpng. ffmpeg made 1280×720 review JPEGs; the largest is 73,923 bytes. The shader combines surface normals and 80 m height bands. Only the small JPEGs were opened for review.

The gate samples a 65×65 XY grid on the closest native-surface tile at each LOD, excluding skirts. It reports two comparisons: displayed mesh against exact Heightfield2D ground/cap roots, and against a regenerated exact same-LOD mesh. The second is the zero-delta regression control; the first measures the actual field error. Swap error compares the coarse tile with its fine children at those same field roots. Full missing-surface, 2 m floating-surface, and missing fine-ring injections were rejected at every LOD3–8; so were both S13 controls.

This is a sampled heightfield gate, not a full camera-ray rasterizer. It can miss cracks narrower than one grid cell (1 m at LOD3, growing to 32 m at LOD8), and its direct-field row does not model tunnel walls or other non-heightfield faces. The same-LOD mesh comparison covers extra and missing mesh intersections at sampled points; a denser or full-screen geometry pass would be needed to certify sub-cell cracks and non-heightfield surfaces.

Closest display distances and budgets

The code sweeps clipmap desired sets across X-boundary phases and measures tile AABB distance. The result matches the clipmap ring formula 3 × 32 chunks × 2^(LOD−1) × 0.25 m; LOD0 is the player tile. The pixel angle is 90° / 2560 = 0.000613592 rad.

LOD	Closest display distance	1 px at that distance	Active rule
0	0 m	0 cm	exact, 2 cm
1	24.001 m	1.473 cm	exact, 2 cm
2	48.001 m	2.945 cm	exact, 2 cm
3	96.001 m	5.891 cm	1 px
4	192.001 m	11.781 cm	1 px
5	384.001 m	23.562 cm	1 px
6	768.001 m	47.124 cm	1 px
7	1,536.001 m	94.248 cm	1 px
8	3,072.001 m	188.496 cm	1 px

Exact-field error and swap pop

Baseline and cutoff values are identical. The exact same-LOD regression control is 0.0000 px p50/p99/max at all LOD3–8, with no missing or floating samples. Against the continuous Heightfield2D roots, however, the current exact meshes already exceed 1 px at every LOD. The owner’s expected near-zero control holds for exact-versus-exact mesh output, not for the true-field comparison.

LOD	Exact field p50 / p99 / max px (baseline → cutoff)	Swap p50 / p99 / max px	Swap cracks / floating
3 → 2	0.9035 / 4.5839 / 7.3277 → same	0.6714 / 4.1861 / 7.3277	0 / 0
4 → 3	1.1170 / 5.6056 / 8.3307 → same	0.8269 / 5.1958 / 8.2105	0 / 0
5 → 4	1.5667 / 45.3010 / 126.2667 → same	0.0000 / 20.7227 / 63.2266	0 / 0
6 → 5	0.5870 / 21.2445 / 82.3615 → same	0.0000 / 21.2445 / 82.3615	0 / 0
7 → 6	0.3393 / 16.4341 / 47.9760 → same	0.0000 / 16.4341 / 47.9760	0 / 0
8 → 7	0.9616 / 24.9479 / 50.0934 → same	0.9616 / 24.9479 / 50.0934	0 / 0

The baseline gate rejects on the field row and every swap row. The cutoff run returns the same rejection. Fault injections also rejected: missing-surface samples were 4,225–13,642; floating-surface samples 4,225–13,641; missing fine-ring cracks 4,225–8,450.

LOD	S13 octave-drop p99 / max px	Stride-2 column interpolation p99 / max px
3	960.2333 / 998.8654	9.3917 / 13.9093
4	3.7405 / 5.1103	9.2115 / 15.3435
5	105.7114 / 129.3089	125.1435 / 191.6424
6	218.7777 / 749.9535	82.1098 / 148.9485
7	95.1867 / 129.1796	40.6881 / 80.6012
8	38.1461 / 282.3288	28.8473 / 278.6236

Both controls were rejected at every LOD3–8. The octave-drop case produced missing samples at LOD3, 5–8 and floating samples at LOD5–8; stride-2 interpolation produced missing/floating samples at LOD5, 7 and 8.

Cutoff contract and detail cost

The version-2 shared visual-detail contract declares stable octave terms with wavelength, octave count, displacement bound, and noise bound. A suffix is eligible only when its wavelengths are undersampled and its full displacement bound fits within one sixty-fourth of the tile’s pixel budget. Skipped zero-mean terms retain the original normalization denominator. Recipes with no terms keep the exact path; FloatingIslandFieldV1 declares none, and its baseline/cutoff LOD hashes match. This is separate from content recipe version 1, which stays unchanged because output did not change.

SurfaceWorld declares 17 terms per LOD. No term or octave tail qualifies at LOD3–8. Sample steps are 16, 32, 64, 128, 256, and 512 voxels respectively. The bounds explain the result: even the smallest declared maximum displacement is 256 voxels, against a 0.2356-voxel 1 px budget at LOD3. The cutoff skips zero octaves, so eligible octave time is 0 ms.

Octave	Terms	Wavelength range (voxels)	Max displacement range (voxels)
0	17	25–10,000	1,024–640,000
1	17	12.5–5,000	512–320,000
2	15	6.25–1,250	256–160,000
3	4	41.664–250	3,968–24,800

W12 profile from the exact baseline; octave milliseconds are summed across evaluated calls/workers. They are all measured detail work, not removable work under the bound above.

LOD	Worker sum s	Octave 0 ms	Octave 1 ms	Octave 2 ms	Octave 3 ms	Detail total ms (% worker sum)	Eligible ms
3	1.660	71.585	69.911	68.535	15.407	225.437 (13.58%)	0
4	3.321	135.512	131.224	128.645	26.506	421.887 (12.70%)	0
5	2.283	82.131	80.450	77.545	16.262	256.388 (11.23%)	0
6	3.587	153.409	148.854	144.175	29.984	476.422 (13.28%)	0
7	1.447	35.301	34.493	33.138	7.069	110.001 (7.60%)	0
8	2.685	58.763	57.751	56.277	11.848	184.638 (6.88%)	0

The cutoff is recipe-generic, display-mesh-only, and activated only for LOD3+. Exact density, detail, collision, edit, and walkability paths remain outside it. LOD0–2 keep the 2 cm path. ARCHITECTURE §8.10 invariants remain: lazy thread-local caches, two-pass MC, SSE noise, and clipmap streaming. The cutoff state is scoped to mesh generation under the async-safety rules.

W12 speed, interleaved with L4

Three L4/candidate pairs, all W12, measured on the L1 camera. The candidate is the final build with --visual-detail-cutoff.

Measure (median)	L4 binary	Candidate	Change
Whole-view ready	2.183 s	2.232 s	+2.2%
Stream wall	1.895 s	1.963 s	+3.6%
Far worker sum, LOD3–8	12.247 s	12.458 s	+1.7%

LOD	L4 worker sum s	Candidate worker sum s
0	3.799	4.094
1	5.304	5.573
2	1.281	1.340
3	1.102	1.138
4	2.723	2.719
5	1.898	1.976
6	2.875	2.944
7	1.243	1.282
8	2.355	2.400

There is no speed benefit because nothing qualifies for skipping; the small regression is cutoff setup and bookkeeping.

Hashes and captures

Baseline and cutoff hashes match for all LOD0–8. The old L4 binary’s aggregate hashes also match: whole view c10697a8f48380c7 / e75c9119a1498cd5 / 49a9a8d1835a6c8c / edcb23288b5e8724 and full detail 0ee0579bd15cbbe4 / 32e5d4964d610fee / 1932b96ee4b1d910 / 69743c6bad3447dc. The recipe stays version 1 because output did not change.

LOD	Density	Mesh	Payload	Topology
0	769f2a706d585a72	33eeeb39fa443816	645621a1d968983f	c2457ed905036c9d
1	69cc5a5c0a2c60af	b48335f5410c57d5	d15c5480162ab8ba	feca1bd025823604
2	85deed5683a77202	53469895987aeea2	e55f8ae7e1532033	d5feb12a0e85c09b
3	6518caefb3be7e7f	ec513c90b853a614	f62fa1137ff17f82	ccf1dcf681965816
4	62d4be96021992f4	f1e4c6efcc2615a9	9dc5f9c88be02f33	fe6f55125a07c3bf
5	5e4a8c38b8cea6fc	08ee56998aaf21e5	541eea8558f26cc8	6539374a8b3336d8
6	e036e95b9006945b	624a33f941a99b02	81b73e14f68bdf86	4de4c63ce1f80a56
7	8ee358af56a30d8d	46ab3c6184048b7c	0a6164b7e68e80c5	0c26b85bb03bc831
8	c3458e1cbdb9c471	46c03d0994d75192	8f89fc5bde12d230	d16dc0cf62031898

Nine direct-Core query probes also match baseline/cutoff and W1/W12 Clang/GCC: density and unclamped e4d4a8eb95b35933, detail efcba20c0c70c415, collision e604823a249029bf.

Before / after pair	View
before_l1.jpg / after_l1.jpg	Exact L1 camera 0,-35,-487.94,0,-22
before_horizon.jpg / after_horizon.jpg	Same position, pitch −2°
before_boundary_lod2_3.jpg / after_boundary_lod2_3.jpg	20 m above local heightfield, aimed at the 96 m LOD3 entry band

The boundary camera is 0,-35,-480.77246,0,-12; the local heightfield is at −500.77246 m, placing the ground intersection about 96 m away.

The six review JPEGs are byte-identical within each pair. PNGs are in Saved/Linux/captures/; JPEGs are in Linux5/captures/.

Failed and retried runs

• The first headless invocation omitted required streaming flags; the next used the default binary-relative data path and could not open the owner JSON. Both were retried with all flags and the explicit Standalone/data/owner_world.json path.
• The first process-monitor wrapper errored after its Clang build had completed; the run was repeated with MeasureProcess.py to record process HWM and duration.
• The first capture named a nonexistent mesh path and exited 2; the exported .vfmesh path was corrected.
• The first gate build failed on a duplicate Distances local; the corrected build passed. A first stride-2 control changed the MC lattice instead of reproducing S13’s every-second-column interpolation; the final control uses bilinear column interpolation.
• Initial comparison against only the continuous field mixed surface interpolation with unrelated tunnel intersections. The final instrument reports exact heightfield roots separately from the exact-mesh regression and ring checks.
• Early W12 cutoff candidates took 2.70–2.91 s. The zero-eligible path and distance setup were moved off the expensive profiling/phase-sweep path; the final interleaved candidate median is 2.232 s. It still does not improve L4.
• One final GCC W12 command omitted --stream-pipeline and exited 2; the corrected rerun passed. The initial −50° boundary capture mostly showed near ground; the final pair uses the camera listed above.
• The final baseline and cutoff gate commands exit 6 by design: both correctly reject the exact-field error and swap pops. No process exceeded 30 minutes or stalled for 15 minutes.

Core/API changes are in VoxelForgeCore.h, WorldAssemblerInternal.h, FloatingIslandsInternal.h, and VoxelForgeCore.cpp. Standalone changes are in StreamingWorld.h, StreamBenchmarkPipeline.h, StreamBenchmarkHeadless.cpp, LinuxScreenSpaceGate.h, LinuxOffscreenCapture.cpp, and BuildLinux.sh; process measurements are in MeasureProcess.py.

For the owner: Far terrain will look exactly as it does now because no octave tail met the bound. Existing far LOD swaps exceed the 1 px target, so the mesh sampling and transition error need attention before octave skipping can be used.

Orchestrator review (Claude)

• Not merged into assembler/world: kept on side branch wip/l5-far-cutoff. With the cutoff flag OFF, the L5 build is slower than L4. My interleaved W12 runs: L4 2.054/2.097/2.179 s vs L5 final7 2.127/2.193/2.208 s (medians 2.097 → 2.193, +4.6%), same world. Cost and no benefit.
• Instrument caveat: the gate prints metric=vertical_height_error. On steep slopes and cliffs, a small horizontal offset becomes a large vertical error, so vertical error overstates screen displacement. p50 is ~1 px but p99/max reach 20–126 px, the pattern such an artifact produces. The claims "today's far meshes exceed 1 px" and "LOD swaps pop by up to 126 px" are therefore NOT yet established. They need a screen-space measure: error along the surface normal or along camera rays, projected at the display distance.
• The cutoff criterion (full displacement supremum ≤ 1/64 px) cannot pass for any SurfaceWorld octave (smallest bound 256 voxels vs 0.24), so the null result was decided by the criterion, not measured.
• The capture path (EGL/GLES pbuffer; llvmpipe, since the sandbox hides /dev/dri) works and is worth keeping.