VoxelForge/Standalone/Research/WorldAssemblerM3/Linux114/REPORT.md

Linux 114 — floor profile and speed

Evaluator version: 222. Worktree remains on wip/l111-halfpipe-spline; no commit was made.

Floor

The L109, L112 and L113 50% rock-station captures all show transverse ribs. In v222, tube-body selection uses round distance and the floor profile is selected separately by nearest centerline. The passage now supplies a signed metric distance outside its void instead of inheriting the host field's saturated deep-rock value.

At the L114 passage 1–2 rock-station column, the raw source value at the floor was saturated at -256.160522 m; the composed floor value was -0.007221 m. The vertical profile is nearly linear:

Height from planned floor (m)	-0.50	-0.25	-0.10	0.00	+0.10	+0.25	+0.50
Density (m)	-0.5064	-0.2568	-0.1070	-0.00722	+0.0926	+0.2423	+0.4919

Along the roughly 10° rock slope, the density at the planned floor changes sharply across one band:

Route distance (m)	57.495	57.795	57.895	57.995	58.095	58.295	58.495
Density at planned floor (m)	+0.00724	+0.00726	+0.00724	-0.07260	-0.05748	-0.02614	+0.00667
Zero-crossing offset (m)	-0.00722	-0.00720	-0.00721	+0.07234	+0.05730	+0.02607	-0.00665

Across 51 samples, the maximum offset is 5.9 cm on passage 0–1 and 7.3 cm on passage 1–2. The L114 rock-floor close-up shows a large reduction from L109/L112/L113, but faint transverse ridges remain. The floor is improved, not fully smooth yet. Full probes: Saved/Linux/l114/audit-owner-clang.log.

Speed

Timing was deferred until load fell from 5.46 to 0.69; the runner's gate measured 0.64 immediately before the batch. It ran five interleaved L80 samples per cell across all four strates. Values are median seconds, LOD0–2 near-ready / whole-view-ready, ordered L109 → L113 → L114.

Strate	W8 near / whole	W12 near / whole
SW	0.778 → 0.851 → 0.849 / 1.808 → 1.874 → 1.861	0.802 → 0.877 → 0.936 / 1.751 → 1.872 → 1.912
FI	1.177 → 1.271 → 1.380 / 2.201 → 2.255 → 2.450	0.946 → 1.122 → 1.102 / 1.816 → 2.012 → 1.993
TN	1.112 → 1.201 → 1.211 / 1.977 → 2.109 → 2.166	0.959 → 1.090 → 1.074 / 1.786 → 1.925 → 1.918
VS	1.019 → 1.149 → 1.156 / 2.065 → 2.308 → 2.263	0.950 → 1.074 → 1.098 / 1.963 → 2.131 → 2.166

L114 is 9–17% slower than L109 at near-ready and 3–11% at whole-view. L114 candidate totals match L113 and exceed L109: SW 3102 vs 583, FI 5384 vs 1036, TN 3701 vs 869, VS 2197 vs 644 feature/route candidates across the profiled tiles. The L77 W8 census estimates median extra worker CPU of 0.71 s at LOD0–2 and 0.88 s whole-view across the four cameras. The largest added buckets are broad other work (+0.72 / +0.60 s) and passage work (+0.21 / +0.40 s); composition-bands work is flat or lower. L77 values are weighted sample estimates, not direct phase timers.

Checks and captures

• Walkability and the 15° forward-grade plan cap pass; the fast audit reports maxima of 10.48° (SW) and 10.01° (FI).
• Clang/GCC hashes match at W1/W8/W12 for all four strates and both two-strate recipes: 18 runs, source stamp 735079f….
• Fast tier is FAIL: the SW owner audit reports 2 wall-shell violations in 31,381 probes. Both samples return the same positive host density (+35.742 m, +35.578 m) where the shell audit expects solid. This remains unresolved after the two floor geometry designs.
• Linux VFHostEditor build succeeded in 80.6 s with -NoUBA -MaxParallelActions=4; Unreal was not run under prlimit.
• The capture index lists 30 programmatically verified views: 14 open-air outside/interior close-ups and one sloped rock-floor close-up for each of L113 and L114. Four low-contrast open-air images are flagged as geometry-visible. See the L114 rock floor and the matching L109, L112 and L113 station captures.

Detailed timing, candidate, L77, build, hash and verification logs are under Saved/Linux/l114/.