Fr0zkaandClaude Opus 5.5
7022d02e6e
Exact one-owner slabs (33,813 samples audited, controls fire), floor/seams/determinism green, routes already
satisfy R > S (S up to 7.9 m). But near-ready 8-21% behind 4041d4b and UE 30 m/s near 40.2 -> 31.5%.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2.4 KiB
L128 slab ownership
Base 4041d4b; branch wip/l128-slab-ownership; core version 226. Source stamp: 8682a9a088f0f02357b3b9722cda9b18a5ebce9886a13cba2d114bc0655d8c70.
Each route body uses half-open frame-plane slabs, so a boundary sample belongs to one segment. The required support is S = r + max(4.0 + 0.21 + 0.125, 0.75 + 0.21) + 0.05 m: bore radius plus the larger sweep/relief/corner or shell/relief reach, with a 5 cm audit margin. The largest body support is 7.885 m; the 4.5 m mouth profile reaches 8.885 m and is assigned to transition slabs. Local uniqueness requires curvature radius R > S (κS < 1). For equal adjacent chord lengths L and turn θ, this is L / (2 sin(θ/2)) > S. Nonlocal mutually-normal centerline pairs must be farther apart than the sum of their supports. Intersections between different passages remain a deterministic field union.
All three owner-world routes pass the body curvature and self-separation checks; no route control points changed. Minimum body curvature radii are 11.565 m (0–1), 14.042 m (1–2), and 8.869 m (2–3), giving at least 0.984 m margin over body support. Route 0–1's 8 tight mouth joints and 5 nonlocal mouth pairs use transitions. The prepared/off-grid query paths share ownership. The audit tested 33,813 support samples: zero missing or duplicate owners, and both negative controls were detected.
The slab projection supplies s to the continuous floor profile h(s); round-distance does not choose floor height. Across 459 floor roots at nine stations, the largest adjacent root step was 1.765 mm and the largest absolute station offset was 7.884 mm. Seam audit: 4,900 shared samples, zero mismatches. Fast suite passed (264 checks, 5 diagnostic); walkability passed. Clang and GCC W1/W8/W12 hashes matched across all three datasets. The 23 L125 captures are in captures/L128/ (19 pass; 4 low-contrast with geometry visible).
Performance bar failed. Five interleaved runs per group across all four strates and W8/W12 put LOD0–2 readiness 14.8–28.0% behind L109 and 8.0–20.6% behind 4041d4b. The five-run 30 m/s UE median also failed versus 4041d4b: near coverage 31.46% vs 40.22%, moving far holes 5.66% vs 3.46%, p99 tick 12.42 ms vs 10.80 ms, and worst tick 14.35 ms vs 13.27 ms. L128 is not ready.
Linux UE host build succeeded (27/27 actions, -NoUBA -MaxParallelActions=4); Unreal ran without prlimit. No commit made. Detailed logs are under Saved/Linux/l128/.