VoxelForge/Standalone/Research/WorldAssemblerM3/Linux82/REPORT.md

Linux82 — samples near passages

Changes

• Kept the 3D relief definition. The wall relief noise is skipped only when its hard-bounded maximum contribution (including mouth-floor relief and the bridge input) cannot change the final max. Mouth surface and tread noise are evaluated only on paths that use them. This is exact; no 2D relief approximation or world change was made.
• Cached exact active BVH-node masks per tile, passage, segment range, and priority. Repeated samples reuse the mask and skip the per-sample lower_bound; candidate segment ordering and distance tests are unchanged.
• The full-view audit now imports the exported Core plan and resolved layout without a recipe override. Snapshot creation preserves that export provenance so generated recipe overrides keep their own route audit.

Passage census

Mean near-LOD passage share of weighted sampler ticks; L81 has two exact-mode runs and L82 has three.

World	L81 exact	L82 exact	Change
FI	64.56%	58.67%	−5.89 pp
TN	65.35%	58.40%	−6.95 pp

Cold readiness

Three interleaved measured pairs per stratum; medians in seconds.

Stratum	LOD0–2 L81 → L82	Whole view L81 → L82
SW	1.158 → 1.110	2.525 → 2.414
FI	2.864 → 2.505	4.240 → 3.797
TN	2.879 → 2.535	4.042 → 3.632
VS	1.232 → 1.224	2.261 → 2.227

All 960 LOD0–1 tile hashes per stratum matched L81 for density, mesh, payload, and topology; whole signatures matched with zero mismatches. The exact path kept evaluator version 214 and made no world-data change.

TN comparison

Five interleaved TN-only LOD0–2 pairs (max-level=2, default scheduling on both binaries): 90378a5 median 3.078 s; 9cb5e53 median 2.747 s. All five paired deltas favored 9cb5e53 (median −0.281 s). The 3.18→3.51 s slowdown was not reproduced in the requested comparison; this A/B does not isolate an L80-to-L81 pair.

Audit and verification

• Canonical full-view walkability: 21,425/21,425 samples passed; all three resolved passage audits passed. The stale 26-point root route was not used.
• Fast tier: PASS, 298.1 s.
• Linux VFHostEditor build after timing: PASS, 27/27 actions in 70.98 s with -NoUBA -MaxParallelActions=4; log rotation was redirected to /tmp after the default config path proved read-only. Three existing RealtimeMesh deprecation warnings.
• Final source stamp: 99fcbd2fd46e0bb3e250f3ad67e98ade0545e419193f987550356015714c6f1b.

Failed attempts

• The first TN harness expected the newer near-LOD marker, which 90378a5 does not emit. It was changed to use whole-view readiness with only LOD0–2 enabled; all five pairs then completed.
• The first full-view audit used --recipe=surface-world, which rebuilt a recipe snapshot and hit the stale root route (65/5,310). Removing the override imported the canonical export and passed.
• The first fast-tier launch lacked its round-specific binary. A later diagnostic exposed lost export provenance in snapshot creation; that was threaded through, and the final fast tier passed.
• The first UE build could not rotate a log in the read-only user config directory. Retrying with XDG_CONFIG_HOME under /tmp succeeded.

Logs and summaries are under Saved/Linux/l82/. No commit, push, or stash was made. All headless runs used the requested CPU affinity and 6 GiB address-space cap; UBT used the CPU affinity without the cap. No run reached 30 minutes.