Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux127/REPORT.md
T
Fr0zkaandClaude Opus 5.5 4041d4bb4b research: smooth-tunnel merge verification (L127)
Builds, fast tier, walk 21425/21425, seams bit-identical, floor 7.884 mm, 18/18 hashes = v225, captures 23/23.
Headless near-ready vs pre-merge: SW +11..14%, other six cells faster. UE start->LOD0-2 2.23-2.73 -> 2.22-2.51 s.
30 m/s UE: near coverage 43.4 -> 40.2%, far holes 3.14 -> 3.46%, p99 11.25 -> 10.80 ms, worst 13.97 -> 13.27 ms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-30 16:10:56 +01:00

1.9 KiB
Raw Blame History

Linux127 — smooth-tunnel merge

Target: aeb60bb on merge/smooth-tunnels.

Build, fast-tier, walkability, seam, floor, determinism, and capture checks pass. The UE moving comparison shows a repeatable coverage drop and is the item to review before landing.

  • Clang and GCC headless builds pass. UE builds pass: c743 baseline 157/157 actions; merge 29/29, using -NoUBA -MaxParallelActions=4.
  • Fast tier passes (259 checks; 5 diagnostic checks). Walkability is 21,425/21,425. Four seam pairs (4,900 shared samples) are bit-identical. Floor probe maximum is 7.884 mm.
  • All 18 owner4/owner2 Clang/GCC W1/W8/W12 hash runs match the L125 v225 baselines.
  • L80 timing protocol on a quiet machine (load 0.69, CPU idle 99.4%): five interleaved measurements per cell, W8/W12 across all four strates. Merge vs L125 is within -4.05% to +2.83% near-ready and -4.58% to +0.95% whole-view. Vs c743, S0 is slower (+10.90% to +14.33% near, +1.26% to +3.52% whole); the other six cells are faster.
  • UE startup reaches full LOD0–2 in all four strates: c743 2.232–2.726 s; merge 2.221–2.514 s. Far-apply CVar reads back as 1 and produces applies in every start and moving run.
  • Five 30 m/s movement runs per build: median near coverage 43.36% → 40.22%; far holes 3.14% → 3.46%; tick p99 11.25 → 10.80 ms; worst tick 13.97 → 13.27 ms. Both builds reach 100% near coverage and zero far holes after stopping. The lane works; moving coverage remains lower on the merge.
  • 23/23 L125 poses rendered and verified from the merge binary; 19 pass pixel quality, 4 show visible low-contrast geometry. See captures/INDEX.md and captures/COMPARE.html.

No source fix was made. Logs and summaries are in Saved/Linux/l127. An early probe wrapper wrote to the L125 probe output; the L125 binary regenerated it, with restoration checksums recorded in Saved/Linux/l127/restore-l125-probe.log.