VoxelForge/Standalone/Research/WorldAssemblerM3/Linux48/REPORT.md

Linux48 report

Whole-surface audit

AuditWholeSurfaceL48.py scans every LOD0 triangle in the route corridor plus 3 m and the chamber box plus 3 m. It welds shared edges, compares geometric face normals, and scores area-weighted normal variation, high-angle edge density, maximum jumps and sliver area in 1 m cells. Connected hot cells become located clusters. A separate floor pass bins the full floor patch at 0.25 m along the route and 0.5 m across it, removes a local moving-mean trend, and checks ripple and adjacent steps. Thresholds use ground-facing native patches from the same export. JSON results and red-triangle XY maps are emitted.

The detector has no general quadric residual or periodic-spectrum metric yet; terraces outside its floor patch rely on the normal-cluster measures. The control results are:

Mesh	Result	Evidence
L45 ramp	Expected fail	One normal cluster, 0.669 m²
L46 entrance	Expected fail	84.3 mm floor ripple; 159.1 mm largest step
L47 wall-foot	Expected fail	One 5-cell cluster, 5.243 m², 42 severe edges; floor profile 34.5/57.6 mm ripple/step
Plain native	Pass	No clusters; floor profile not present
L47 smooth floor	Pass	0.59 mm ripple; 1.15 mm largest step
L48 candidate	Fail	Two clusters, 7.095 m² total, 65.5 severe edges, 303 flagged triangles; floor profile 34.5/57.6 mm

The L47 cluster is at approximately x=[-5,0], y=[22,25], z=[-505,-504] m. At 0.188 m above the floor its wall-band p95 dihedral is 54.6° with a 160.4° maximum; the selected floor rails did not reach this band. The complete control output is in Saved/Linux/l48/whole-surface-audit.json.

Teeth measurement and attempted fix

The L47 measurements localize the teeth to the floor-cap/bore-wall junction. The post-stamp cap has a hard radial cutoff there, and the floor-adjacent re-carve also uses a non-SDF weighted blend. A re-carve fade alone left the teeth in the capture. The final trial replaced both cavern cap gates with a rounded max of the sloped-floor distance and radial bore distance, restoring the L47 re-carve weight. This reduced the apparent tooth height in the capture but did not clear the whole-surface gate; the final mesh has more flagged area and a second cluster. The hard cap edge is implicated, but the cause is not proven and this is not a validated fix.

Hashes and cost

Clang and GCC builds used source stamp 05ea28750b2ee89aa61655c386c6d74a6841c4d071374cda03c73a89b5efb578.

Demo-off whole-view hashes match L47 exactly at W1 and W12:

• floating-islands: density 598e9daaa1744ccb, mesh 3114e0f4f6dbf734, payload c49c4e48bff74089, topology cadd4edd10d4d8ef.
• surface-world: density 7dc75e2c09302258, mesh 0cb34c40ceec4ce3, payload 9d23e8c46b515cf1, topology 2962c9464991eedd.

Fast took 202.7 s. Medium took 637.6 s; the whole-surface stage took 161.8 s and failed on L48. Full and its five-pair cost comparison were not run because medium did not pass. All build and verification processes used TMPDIR=Saved/Linux/l48/tmp; no command wrote to /tmp.

Tiers and captures

Tier	Result
Fast	Pass, 202.7 s
Medium	Fail, 637.6 s: L48 whole-surface gate
Full	Not run: medium prerequisite failed

The five captures are ground-entrance.jpg, entrance-looking-in.jpg, overview-entrance-and-chamber.jpg, inside-to-dais.jpg, and entrance-floor-low.jpg in captures/. Flag maps: flagged-regions-L47-wall-foot-control.svg and flagged-regions-L48-final.svg (red flagged triangles, gray ROI, blue route).

Failed runs

• The first fast invocation had no built binary; after the Clang/GCC builds, fast passed.
• Bypassing the post-stamp floor cap failed entrance rails. Broad SmoothMin/cap-mask trials either closed the mouth or failed the built-shape check; one planar mouth patch measured 5.575 m².
• Radial re-carve fades at 0.5 m and 1 m preserved the rails but left visible teeth and failed the whole-surface audit. Changing the support inset had no mesh effect.
• The final rounded-cap trial passed the legacy rails and shape checks, but failed medium with the clusters above. The 15-trial budget is exhausted. No full tier, commit, push or stash was performed.

Orchestrator review (Claude)

Split. The whole-surface audit is the result of this round, and it's committed here as a tool, not yet wired into the tiers. All three historical defects fail it (L45 ramp, L46 entrance, L47 wall foot) and plain native terrain passes. It also shows that L47's floor near the walls carries 34.5 mm ripple and a 57.6 mm step, which the three rails never sampled. The Core change (rounded cap) did not pass and is parked with the tier wiring on wip/l48-surface-audit. Wiring the gate now would turn medium red on HEAD, because the flagged demo still has the teeth. assembler/world stays on L47 Core. L49 builds the floor and bore as one SDF, with a filleted floor/wall corner, before stamps, instead of patching with a post-pass cap.