VoxelForge/Standalone/Research/WorldAssemblerM3/Linux34/REPORT.md

VoxelForge L34 report

Verdict and fix

The 0.125 m floor and wall profiles found no 0.5 m flat run followed by a step of at least 0.1 m in any of the 24 style, recipe, fixture, and compiler runs. The L23-style negative-control profile (2.0 m flat run, then a 0.25 m step) was detected. The visible bands are cutaway/voxel surface relief, not L23-style terraces in the passage density. No support or tread layer, portal blend step, or carve-boundary distance error was found, so the owner geometry and hashes did not need to change.

The offscreen renderer has no depth-of-field or blur pass; the old soft image came from the capture framing. The reusable capture script uses 80° FOV, 1.6 m eye height, outside mouth cameras 5.15–6.27 m from the route mouth, LOD0 meshes, and tile skirts off. Inside views use clip planes with cap triangles. Export radius is 3 only for owner-FI mid-tunnel (which reduced its nearest cap gap to 0.058 m); other views use radius 2.

Audits and captures

• Clang and GCC builds passed. Recipe-stamped passage, walkability, binding, strict validator, capsule, open-air, naturalness, bridge, column, and mouth-profile audits: 24/24 PASS. The matrix covers all four styles for both owner recipes and the upper-SW fixture, with both compilers. L24 and L26 negative controls also triggered.
• Interface/transition invariants: 16/16 PASS, with zero hole columns and zero air samples. L13/L9: 4/4 PASS. Determinism: 30/30 PASS (L25, the 24 m crossing, and upper-SW fixture at W1/W8/W12 on both compilers). Validator controls rejected 10/10 invalid cases.
• Standard set: 12 sharp 1600×900 JPEGs, all 23.5–45.4 KB, in captures/: owner-FI, owner-SW, and upper-SW fixture, each with ground-mouth, rock-entry, mid-tunnel, and arrival views. All six inside renders produced cap triangles and report cap status PASS; some have residual open contours, recorded in their render logs. SHA-256 values are in captures/SHA256SUMS.txt.
• Owner W12 hashes match L33. Recipe/evaluator versions remain 2/20; no Unreal files changed.

Recipe	Density	Mesh	Payload	Topology
SurfaceWorld	953eb1df1705f277	d8006cb43f44aa5b	9217bff766970de5	83e3aa6f3a063e60
FloatingIslands	806537cdd98f8afb	a4b06dd4d260943d	763a99aa04a73eae	baa37d80d3620d87

• Cost: five interleaved L33/L34 pairs per recipe. Median whole-view-ready delta was +0.06% for SurfaceWorld and +0.52% for FloatingIslands, within run noise. All 20 paired runs passed.

Failed and retried runs

• The first fixture determinism runner used a nonexistent L34 fixture path: its 24 owner rows passed, six fixture rows failed; correcting the path to the L33 fixture directory made all six pass.
• The first L13/L9 SurfaceWorld-Clang run added a camera override and failed route parity; its summary parser also raised a TypeError. Removing the override and fixing the parser produced the final 4/4 pass.
• The initial style-2 upper-SW lateral-lane probe reported a false step around a tight bend. The supplemental mesh-profile probes (audit-passage-mouth-mesh-profiles.log, r2–r7) also had incomplete coverage on several capture exports. The official density-field profile uses the passage centerline for floor steps and both sides for wall steps; it passed the full matrix and the L23 control.
• Capture renderer exit-3 retries were recorded in run-captures.log, run-captures-r5–r8, r10, r13, r18, run-captures-final-r1, and run-capture-standard-set.log. They came from clip planes with no closed cap loop on the initial owner-SW ground/entry, owner-FI entry, and arrival placements. A three-value entry clip was rejected; a four-value outward clip had a 5.1 m open gap, and an inside entry clip obscured the passage. Exterior views now use no clip. Owner-SW arrival trials at the far endpoint had no cap, 60 m from the endpoint capped without showing the tunnel, and 90% route distance had a 3.884 m cap gap; the final view uses the exact route camera at 80%. The upper-SW fixture arrival at 60% failed with a 40.786 m gap; 45% passes. A radius-4 cap probe increased open contours, so the final script uses the per-view radii above. The final full capture run passed all 12 views.
• Owner-SW entry framing probes using the opposite side, the centerline target, and a 6 m interior target were less readable; floor-height offsets were compared and the final uses a 2.5 m portal target with no clip or focus crop. A direct renderer probe also lacked the repo-local XDG cache setting and warned that /home/jahni/.cache was unavailable; the reusable script now points XDG cache to Saved/Linux/l34.
• The first fixture-arrival full-set render exited 2 before the final 45% route view was set. The later 60% cap probe failed as noted above; the final 45% render passes. No final run failed.
• No run exceeded 30 minutes or remained stuck for 15 minutes. All builds and audits used the requested CPU, memory, and job limits. Nothing was committed, pushed, or stashed.

Orchestrator review (Claude)

• Merged as tooling: the standard 12-view capture set is sharp and comparable, and the mouth-profile audit with its L23 control is a good addition. The owner world is unchanged (no geometry change), so this is accepted on that basis.
• Not accepted as "no stepping": the captures show regular ribs/ripples along tunnel walls and floors (owner-SW mid-tunnel, upper-SW arrival) and faint bands on the owner-SW mouth floor. They are periodic and smaller than the 0.5 m-flat + 0.1 m-step detector, so the detector can't see them. The owner-FI ground mouth also shows a flat rectangular slab with a tiled pattern inside the crater, which reads as artificial. Both go to L35.