VoxelForge/Standalone/Research/OwnerShapeWorld/speed_round/round10/REPORT.md

Round 10 — passage look defects

Result

Two defects are fixed in the final VoxelForgeRound10Final.exe: the broad landing plate and the repeated high-frequency bands on passage shafts. The changes are local to landing overlays and passage-controlled relief. The chamber, tunnel, far, and tunnel-bend captures are pixel-identical to Round 9.

The required cold, full-payload W8 interleave averaged 0.993 s for Round 10 and 1.022 s for VoxelForgeRound9Final.exe. The Round 10 samples were 1.014 / 0.972 s; one sample is 14 ms over the target, while the mean is 7 ms under it. A separate new-only cold repeat measured 1.044 s, so the local run-to-run margin is not robust yet. All runs generated 2,226 tiles with zero cache hits and a 110.90 MiB payload.

Order	Executable	Process wall	Shape stage	Detail band	Far-parent build	Full clipmap	Cache hits
old A	VoxelForgeRound9Final.exe	1.071 s	402.974 ms	210.013 ms	46.082 ms	0.473 s	0 / 2,226
new A	VoxelForgeRound10Final.exe	1.014 s	399.349 ms	206.952 ms	43.985 ms	0.462 s	0 / 2,226
old B	VoxelForgeRound9Final.exe	0.973 s	402.974 ms	211.865 ms	44.336 ms	0.463 s	0 / 2,226
new B	VoxelForgeRound10Final.exe	0.972 s	402.045 ms	210.640 ms	43.764 ms	0.464 s	0 / 2,226
new C (separate repeat)	VoxelForgeRound10Final.exe	1.044 s	399.033 ms	206.829 ms	44.094 ms	0.465 s	0 / 2,226

Primary interleaved means: process 1.022 → 0.993 s, shape stage 403.0 → 400.7 ms, detail band 210.9 → 208.8 ms, far-parent build 45.2 → 43.9 ms, full clipmap 468 → 463 ms. Startup/setup/finalization, derived by subtracting shape and full-clipmap walls from process wall, is approximately 151 → 129 ms on the paired runs; the separate new C remainder is 180 ms, so startup variation dominates the small margin. It is not a separately timed stage. Head-to-head stdout logs: old A, new A, old B, new B. Separate repeat: stdout, runner. The final /MP8 build is here.

Causes and fixes

Passage shaft bands

The bands came from relief displacement on passage capsule surfaces, not from the passage radius profile or segment joins. Radius/slope joins are still zero. The earlier radius and filter probes made no visible change; owner roughness alone and surface value noise alone each retained the bands. Tagging the passage source and suppressing its relief removes them while leaving detail on other cave walls unchanged.

I rejected partial suppression: a 50%-relief probe had shaft high-pass RMS 0.822, and a 20%-relief probe measured 0.386 and still showed residual rings. The final fully passage-controlled wall measures 0.276. The fixed-camera ring check uses a 0.350 limit: it fails Round 9 at 1.331 and passes Round 10 at 0.276 (before, after). The final close-up shows the result.

Landing and floor plate

AddLanding added a shallow positive-rock ellipsoid directly under the floor. It formed the broad brown shelf where the shaft meets the passage floor. The landing carve and native floor remain; the solid ellipsoid was removed. In the fixed passage camera, the warm plate mask drops from 71.2661% to 0.0000% of its ROI. The plate check fails the Round 9 capture and passes Round 10 (before, after). The chamber image is unchanged.

Boundary components / slivers

The final shard audit reports two small components at tile (-3,-14,-3): 56 triangles at 3.079 vox thickness and 20 triangles at 1.347 vox. Neither is a thin sheet. An instrumented cross-tile check found 10/10 and 4/4 exact boundary vertices on the adjacent tile, respectively. They are connected surface pieces split by the per-tile component audit, rather than isolated shards. The final audit still reports zero isolated in-view components and zero thin sheets; its two boundary candidates remain connected across the seam. I made no geometry deletion here because it would cut the continuing surface. If the viewer shows a separate visible seam artifact beyond these audit hits, this check does not claim to resolve it.

Evidence: final shard audit, cross-tile probe. The latter used a temporary audit-only instrumentation build; StandaloneViewer.cpp was restored afterward.

Determinism, audits, and payload

• W1 and W8 match all 44 density, mesh, payload, and topology hash fields: comparison, W1, W8.
• Passage walkability is 5,642/5,642. Shape/look checks pass: GridStair relative join step 0.090 / 0.400, zero thin sheets, detail audit pass, and flat-patch share 6.5%.
• Far survival passes 7,200/7,200 boundary surfaces and all 10,976 connectors across 178,351 sampled sections. Missing surfaces, holes, and unsampled connectors are zero: audit.
• The damaged-shell control still fails as required: 512 triangles removed, 16 missing surfaces and 16 holes, exit code 14: control.
• The direct benchmark and both hash runs generated all 2,226/2,226 tiles, used the complete 110.90 MiB payload, and recorded zero mesh-cache hits.

Captures

The six Round 9 before captures and six Round 10 after captures use the established fixed cameras. Pixel comparison: full results.

View	Camera (x,y,z,yaw,pitch)	Round 9 before	Round 10 after	Changed pixels
Chamber	(-11.75,-123.25,-19.85,0,0)	before	after	0
Tunnel	(-11.75,-123.25,-19.85,90,0)	before	after	0
Passage	(-39.19,6.27,-28.51,63,0)	before	after	932,273 / 1,024,000
Far	(-11.61244,-118.2212,-19.8452,0,0)	before	after	0
Passage-floor close-up	(-39.19,6.27,-25.51,63,-25)	before	after	972,890 / 1,024,000
Tunnel bend	(-198.97,-99.57,-45.58,33.59,23.17)	before	after	0

The remote captures use (500,0,0,yaw,0) at yaw 0/90/180/270. The first comparison is Round 8 vs Round 9; the second is Round 9 vs Round 10.

Yaw	Round 8	Round 9	Changed R8→R9	Round 10	Changed R9→R10
0	image	image	0 / 1,024,000	image	136 / 1,024,000
90	image	image	111 / 1,024,000	image	2 / 1,024,000
180	image	image	32 / 1,024,000	image	0 / 1,024,000
270	image	image	386 / 1,024,000	image	386 / 1,024,000

Per-channel means/maxima are in Round 8 vs Round 9 results and Round 9 vs Round 10 results.

Failed attempts and machine rules

• One capture-wrapper invocation used a nonexistent ...VoxelForge\speed_round working directory and failed before launching a child (Windows error 267). It was rerun from Standalone\Research\OwnerShapeWorld\speed_round; no workload process started on the failed attempt.
• The first six image/hash helper launches through RunMachinePython.ps1 passed the script path to Start-Process without quoting its E:\Projet Unreal path; Python exited 2 with “can't open file E:\Projet.” I quoted the path and reran all six hash, pixel, and baseline/final gate checks successfully (the two baseline gates exit 1 by design). Records: failed launches, constrained reruns.
• The first landing-image check expected 1024×768, while captures are 1280×800. That check failed on both images; the ROI script was corrected and rerun. The recorded final before/after logs above use 1280×800.
• The 20%-relief probe was rejected for visible residual bands, despite passing the earlier looser 0.55 exploratory threshold. The final threshold is 0.35; the final image passes.
• The bad-shell exit 14 and the two baseline image-gate exit 1 results are expected controls. All builds succeeded with /MP8; final benchmark, hash, audit, capture, and image-check processes ran BelowNormal on affinity 0xFF00. No process returned 0xC0000142, and no viewer owned by anyone else was closed.

No commit, push, or stash was made.