L74's passage bound omitted the 0.21 m wall relief and 0.20 m floor relief, so its whole-tile
proof was unsound in principle (no tested view hit it). Passage support is now bounded per route
segment by four parameter-subinterval boxes using Catmull-Rom derivative extrema, local max
radius and relief padding. Stratum slicing only at LOD0-2 in FI/VS. Legacy sub-tile block
certificates are audit-only (runtime version was slower). TN regression (-80,-20,-3218) kept.
Exact full-grid vs skip hashes on every tile, 4 views, Clang/GCC, W1/W8/W12; control caught;
80-camera sweep 0 violations. Whole view vs c67b200: SW -1.3%, FI -7.6%, TN -2.3%, VS +0.2%.
Compiles in the Linux UE host (VFHostEditor, 27/27 actions).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
VoxelForge verification
Build stamped Linux executables and run one of the three verification tiers from the repository root. Artifacts and logs go under Saved/Linux/l39 by default.
taskset -c 0-11 prlimit --as=6442450944 -- timeout 1740s env VOXELFORGE_BUILD_ROOT=Saved/Linux/l39/build VOXELFORGE_BIN_DIR=Saved/Linux/l39/bin Standalone/BuildLinux.sh --compiler=clang++-19 --variant=l39-clang --jobs=12
taskset -c 0-11 prlimit --as=6442450944 -- timeout 1740s env VOXELFORGE_BUILD_ROOT=Saved/Linux/l39/build VOXELFORGE_BIN_DIR=Saved/Linux/l39/bin Standalone/BuildLinux.sh --compiler=g++ --variant=l39-gcc --jobs=12
taskset -c 0-11 prlimit --as=6442450944 -- timeout 1740s python3 -B Standalone/Tools/verify/verify.py fast
taskset -c 0-11 prlimit --as=6442450944 -- timeout 1740s python3 -B Standalone/Tools/verify/verify.py medium
taskset -c 0-11 prlimit --as=6442450944 -- timeout 1740s python3 -B Standalone/Tools/verify/verify.py full
fast runs legacy owner regressions against the pinned two-strate fixture at Standalone/data/regression/owner_world_2strate_l61.json, including resolved styles, passage wall/floor ripple and coverage gates, adjacent-profile step, W1/W12 hashes, and LOD0 ring profiles with the L30/L31 controls. It also checks the live four-strate Standalone/data/owner_world.json: generator and resolved-style maps are read from that export, then strict layout, transitions, per-strate deterministic generation, arrival-to-departure routes, and spawn costs are audited. The tier also exercises stale-source, style-mislabel, black-capture, historical ripple, and L38 step controls.
medium adds the Clang style-override matrix, transitions, route/proof audits, LOD0–8 rings, and pinned two-strate owner captures, shape, near/overview crease, and fragment checks. Run medium before full. Full reuses that evidence only when the source stamp, verification-tool hashes, binaries, inputs, logs, and audit-script hashes still match; otherwise it reruns the Clang medium work. It adds the owner/style/transition/proof matrix with GCC, the 30-cell determinism matrix, and five interleaved L36/L39 cost pairs per recipe. The final captures stay in the L39 output directory.
BuildLinux.sh hashes a sorted manifest of Core inputs, standalone sources and generated wrappers, build files, and optionally the Unreal adapter when VOXELFORGE_INCLUDE_UNREAL_ADAPTER=1. The manifest hash is compiled into the headless and offscreen binaries and printed before any other runtime log line. Each tier probes every available built executable at startup and re-probes it at completion, rejecting stale stamps or binary changes. Summaries store tool hashes, binary and JSON SHA-256 values, resolved recipe/style/layout identity, commands, elapsed time, and audit-script hashes.
Medium and full stop at their first failed gate. Fast defers a four-strate stage failure until after it has collected the pinned two-strate regression evidence. Every process is limited to CPUs 0–11 and 6 GiB of address space; individual build and audit processes time out at 30 minutes. Summary files are fast-summary.json, medium-summary.json, and full-summary.json.