Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux119/REPORT.md
T
Fr0zkaandClaude Opus 5.5 707bc61553 core: coarser smooth-route tolerance + shell support bound (L119, evaluator v224)
Chord error 2.5 -> 10 cm, shell work skipped outside the 0-0.75 m band. Sweep pairs -39%, prepare -60%,
near-ready 3-8% faster than L115 but still +13..+22% vs L109. Floor offset 7.88 mm; captures match L115
by eye; fast tier, walk audits, Clang/GCC W1/W8/W12 determinism green. Not for merge yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-30 05:35:59 +01:00

44 lines
3.9 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Linux119 — smooth half-pipe speed
## Result
The route now uses a 10 cm chord-error tolerance (was 2.5 cm); the 8 m spacing cap is unchanged. The evaluator is version 224. Output changed, but Clang/GCC and W1/W8/W12 hashes match.
L119 improves on L115, but misses the L109 near-ready target in all eight stratum/worker cells: **13.3–22.0% slower than L109**. It is 2.8–7.7% faster than L115 at near-ready. Whole-view readiness remains 6.5–12.3% slower than L109.
Five interleaved L80 measurements per cell followed one warmup per version. The timing gate recorded 1-minute load 0.82 and 98.7% CPU idle. Headless runs used CPUs 0–11, the 6 GiB address-space limit, and a 1,740 s timeout.
| Stratum | Workers | Near L109 / L115 / L119 (s; Δ vs L109) | Whole view L109 / L115 / L119 (s; Δ vs L109) |
| --- | ---: | ---: | ---: |
| S0 | 8 | 0.761 / 0.929 / 0.880 (+15.6%) | 1.742 / 1.927 / 1.860 (+6.8%) |
| S0 | 12 | 0.785 / 0.926 / 0.890 (+13.5%) | 1.717 / 1.873 / 1.829 (+6.5%) |
| S1 | 8 | 0.951 / 1.250 / 1.158 (+21.9%) | 1.863 / 2.112 / 2.073 (+11.3%) |
| S1 | 12 | 0.899 / 1.152 / 1.097 (+22.0%) | 1.731 / 2.000 / 1.943 (+12.3%) |
| S2 | 8 | 0.994 / 1.252 / 1.177 (+18.4%) | 1.844 / 2.079 / 2.030 (+10.1%) |
| S2 | 12 | 0.914 / 1.169 / 1.079 (+18.1%) | 1.710 / 1.989 / 1.885 (+10.2%) |
| S3 | 8 | 0.812 / 0.988 / 0.960 (+18.3%) | 1.728 / 1.870 / 1.891 (+9.4%) |
| S3 | 12 | 0.772 / 0.928 / 0.874 (+13.3%) | 1.617 / 1.783 / 1.723 (+6.6%) |
## Designs and work
Two designs were built together. Before building, the fourfold chord tolerance increase predicted about half as many curvature-limited route segments (segment count scales roughly with the inverse square root of the error tolerance). The shell support bound was expected to skip most shell calculations: only flattened distances in the 0–0.75 m shell band can contribute, with a 1-subvoxel guard. L118 had 48.93M W8 sweep pairs and 39.3M sampled smooth-min calls.
| Stratum | W8 LOD0–2 pairs L109 / L115=L118 / L119 (M) | Exact fallbacks L109 / L115=L118 / L119 (M) | Segment visits L109 / L115=L118 / L119 (M) | Prepare ms L118 → L119 | L119 shell-support skips |
| --- | ---: | ---: | ---: | ---: | ---: |
| S0 | 3.36 / 7.56 / 4.55 | 0.241 / 0.413 / 0.652 | 1.19 / 3.41 / 4.68 | 467 → 185 | 3.73M (81.9%) |
| S1 | 6.19 / 16.22 / 9.69 | 0.369 / 0.719 / 1.114 | 1.91 / 7.60 / 9.25 | 919 → 361 | 8.05M (83.1%) |
| S2 | 6.15 / 15.52 / 9.28 | 0.315 / 0.634 / 0.986 | 1.66 / 6.88 / 8.34 | 848 → 345 | 7.69M (82.8%) |
| S3 | 4.09 / 9.63 / 5.88 | 0.074 / 0.131 / 0.203 | 0.48 / 1.41 / 2.19 | 567 → 235 | 4.76M (80.8%) |
L119 reduced pairs 39–40% versus L115/L118, and summed W1 prepare time fell 59.8% (2.801 s to 1.125 s). Exact fallbacks are 55–58% higher than L115/L118 and visits are 21–56% higher. Those extra fallback scans are the likely remaining readiness cost. L119 prepare time is timed directly; L109/L115 prepare time was not separately instrumented.
## Quality and verification
- Floor root offsets: 51 samples at each of 9 stations; maximum 7.884 mm versus L115’s 7.849 mm.
- Fast tier: PASS in 277.9 s. Walkability, floor/mouth profiles, open-air shell, ripple, naturalness, and grade audits passed; no geometry threshold was relaxed.
- Determinism: all 18 Clang/GCC W1/W8/W12 hash runs matched for owner4 surface world and both owner2 canonical recipes. Owner2 v224 baselines were added. Source stamp: `9a592cb548c6768c8c25723f4168f65da9645068cb67dcb317e91f707e7d41df`.
- Captures: 23/23 L115 view names and camera poses rendered with the L119 binary; 19 pass and 4 low-contrast-visible. L115 had 18 pass and 5 low-contrast-visible. [Compare all views side by side](captures/COMPARE.html); [L119 capture index](captures/L119/INDEX.md).
- Linux VFHostEditor: PASS, 27/27 actions in 70.98 s with `-NoUBA -MaxParallelActions=4`. Unreal ran without `prlimit`; three existing RealtimeMesh deprecation warnings remain.
Logs and timing summaries are in `Saved/Linux/l119/`.