0 Fork 0

Files

T

Fr0zkaandClaude Opus 5.5 36d954cf2f passages: rounded half-pipe crest (L113, v220); floor bands diagnosed, speed not recovered

Open-air walls: 0.75 m rounded shell, relief fades over the top 0.25 m - the needle spikes are
gone. Every owner plan passes the 15 deg forward-grade check (max ~10.7 deg) and all walk audits.
Floor bands: median 1.25 m interval, ~0.21 m drop on a ~10.4 deg floor = one 0.25 m voxel row
projected along the grade (not control points, not slab thickness) - not yet fixed. Exact
segment-to-tile filter did not reduce candidates; removed. Timings this round are unreliable
(a gitea process used ~43% CPU throughout). Fast tier PASS; Clang/GCC deterministic; UE builds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1

2026-10-29 21:57:13 +01:00

2.9 KiB

Raw Blame History

Linux 113 — half-pipe crest, floor bands, and readiness

Version v220 is on wip/l111-halfpipe-spline at the existing 16554cd base. No commit was made.

Shape and walk audits

The open-air wall uses a 0.75 m rounded shell, three 0.25 m voxels thick. Wall relief fades across the top 0.25 m instead of ending at the crest. The captures show the needle-like spikes are gone; some low scalloping remains at the voxel scale.

All three owner passage plans pass the 15° forward-grade check. Their maximum sampled grades are 10.391°, 10.687°, and 10.289°. Passage, mouth, floor, and shell audits pass.

The floor bands remain visible. On the exported mesh, upward-facing floor samples along the long open-air run show a median 1.25 m interval and 0.21 m vertical drop (most drops are 0.17–0.25 m). The planned grade there is about 10.43°; a 0.25 m voxel row projects to about 1.36 m along that grade. Control points are spaced a median 9.18 m apart, and a temporary 1 m sampler control reproduced the same audit misses. The measured bands align with voxel rows; they do not match the control-point interval or the 0.75 m slab thickness. They were not removed in this round.

Captures

Capture index contains 50 L112 before images and 50 L113 after images. All 100 pass the pixel and geometry checks. The set includes mouths, stations, open-air views, and 14 after 4 m wall close-ups; their measured camera distance is 4.00 m within 0.02 m.

Readiness

Five interleaved runs per cell, all four strates. Each value is median LOD0–2 / whole-view seconds in L109 / L112 / L113 order.

Strate	W8 near	W8 whole	W12 near	W12 whole
SW	1.038 / 1.027 / 1.511	2.291 / 2.235 / 2.947	1.304 / 1.623 / 1.629	2.476 / 2.886 / 2.819
FI	1.081 / 1.122 / 1.238	2.123 / 2.081 / 2.267	1.080 / 1.187 / 1.187	2.048 / 2.168 / 2.132
TN	1.497 / 1.379 / 1.312	2.538 / 2.403 / 2.352	1.152 / 1.415 / 1.279	2.126 / 2.385 / 2.286
VS	1.444 / 1.617 / 1.136	2.676 / 3.117 / 2.639	0.818 / 0.896 / 0.925	1.753 / 1.824 / 1.853

FI/TN W12 near readiness is 1.187 / 1.279 s for L113 versus 1.080 / 1.152 s for the matched L109 rerun. The speed target was not met.

Strate	L109 p95/max	L112 p95/max	L113 p95/max candidates per tile
SW	3 / 21	18 / 185	14 / 137
FI	4 / 47	22 / 420	17 / 297
TN	4 / 23	22 / 116	18 / 92
VS	4 / 16	18 / 82	12 / 58

The second speed design, an exact segment-to-tile distance filter, barely changed L113 candidate counts and did not recover near readiness. It was removed; there was no third design.

Verification

Fast tier: PASS, 264 checks.
Clang/GCC determinism: all 18 W1/W8/W12 hash runs match within each owner configuration.
Linux VFHostEditor: succeeded after timing with -NoUBA -MaxParallelActions=4; no prlimit was used.
No commit, push, or stash.

2.9 KiB Raw Blame History Unescape Escape