VoxelForge/Standalone/Research/WorldAssemblerM3/Linux125/REPORT.md

Linux125 — passage slab stamping

Result

No-go for the L109 speed bar. The v225 stamp candidate keeps the L119 10 cm curve tolerance and smooth half-pipe look. It is 4.72–10.19% slower than L109 for near-ready LOD0–2, though 1.96–14.23% faster than L119. Whole-view readiness ranges from 0.47% faster to 8.04% slower than L109.

The BVH trial was rejected: preparation was 5–6× slower. The final design stamps conservative route slabs with an outside sentinel and evaluates segment geometry over tight expanded bounds. The separate 25 cm tolerance trial was also rejected: it remained 11–23% slower than L109 and its maximum floor-root offset was 8.601 mm.

Count first

The tile cache already stores prepared segment contributions. At LOD0, each stratum still evaluates 3.11–8.49M segment/sample pairs, while the exact one-sample band is 5.16–7.41% of grid samples and pair overlap is 67–73%. Stamping cuts exact fallback work, but preparation remains dominated by segment/sample work.

Stratum	LOD0 pairs (M)	Exact band	Pair overlap L119 → stamp	Prepare ms L119 → stamp	Exact fallbacks (k) L119 → stamp	Max tile cache (MiB)	Process RSS MiB L119 → stamp
S0	3.11	5.16%	67.3 → 67.2%	246 → 272	426 → 73	3.93	724 → 723
S1	8.49	6.53%	72.6 → 72.0%	620 → 700	912 → 172	7.85	580 → 583
S2	8.30	7.41%	72.6 → 72.0%	606 → 621	892 → 180	7.85	643 → 641
S3	5.27	5.35%	67.7 → 67.6%	443 → 463	198 → 73	3.93	657 → 667

Across strata, mean preparation time by LOD was LOD0 479→514 ms, LOD1 121→130 ms, LOD2 3.8→4.1 ms. Exact-band fractions were 6.18%, 4.14%, and 6.16%; pair overlap was about 71%, 67%, and 68% respectively. Per-caller LOD0 exact fallbacks show the main gain: DensityGrid fell from 479k calls per run on average to zero. Remaining callers account for about 124k calls per run, led by skirt refinement (75k), midpoint correction (19k), and skirt extrusion (18k). Off-grid misses remain on the canonical exact path: 96–235k per stratum at LOD0.

Timing

Five interleaved L80 measurements per cell, one warmup per version, all four strata. The final quiet-machine gate was 1-minute load 0.63 and 98.8% CPU idle. Values are medians in seconds; near-ready is LOD0–2.

Stratum	Workers	Near L109 / L119 / v225	Whole view L109 / L119 / v225
S0	8	0.745 / 0.850 / 0.780	1.694 / 1.762 / 1.686
S0	12	0.767 / 0.878 / 0.807	1.680 / 1.783 / 1.714
S1	8	0.921 / 1.144 / 0.981	1.771 / 2.017 / 1.855
S1	12	0.887 / 1.082 / 0.962	1.711 / 1.912 / 1.794
S2	8	0.973 / 1.165 / 1.025	1.811 / 2.020 / 1.861
S2	12	0.905 / 1.070 / 0.965	1.689 / 1.862 / 1.758
S3	8	0.809 / 0.943 / 0.879	1.667 / 1.871 / 1.801
S3	12	0.760 / 0.855 / 0.838	1.587 / 1.687 / 1.674

Quality and verification

• Floor probe: 9 stations × 51 samples; maximum root offset 7.884 mm, matching the L119 probe. Route centerline openness passed for 21,425/21,425 samples; maximum grade was 10.695°.
• Fast tier and walk audits: PASS, 264 checks across 80 runs.
• Seams: the audit exits nonzero on any shared-sample mismatch. Four adjacent LOD0 tile pairs passed, with 1,225 shared samples per pair (4,900 total), bit-identical and max delta 0.
• Determinism: all 18 Clang/GCC W1/W8/W12 runs matched across owner4 surface-world, owner2 surface-world, and owner2 floating-islands. Output changed; evaluator version 225 and v225 canonical baselines were added. Source stamp: c1ac9c910e8bb26eba5e25f0396a3b68845627b27bf52cf3f951a37c873783f3.
• Captures: 23/23 L119 views rendered at the same poses; 19 pass and 4 low-contrast-visible. Compare captures side by side.
• Linux VFHostEditor: PASS, 31/31 actions, 75.24 s total with -NoUBA -MaxParallelActions=4. Unreal ran without prlimit.

Logs, counts, probes, hashes, and timings are under Saved/Linux/l125/.