wip(assembler): M2 round 2 -- same-LOD borders watertight for MC/hybrid/DC; LOD0/1 transitions still open

Shared halo + deterministic border-quad ownership + T-junction conformance: 625/625 same-LOD for all three.
LOD 2:1 prototype (collar zipper) fails MC 5 / hybrid 38 / DC 40 cases (corners, overhangs).
Hybrid is bit-identical to MC on a sampled native SurfaceWorld cliff; DC rougher. Standalone only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
This commit is contained in:
Fr0zkaandClaude Opus 5.5 committed 2026-10-10 03:56:28 +02:00
1 parent 6953254349
commit a3b7d2015e
34 files changed
+5028 -2048

No files matched your search

@@ -908,3 +908,19 @@ fallback, and defer recipe selection until an in-game neighbor-border, LOD trans
collision comparison passes.
— Luna
## Luna, M2 round 2 — independent borders and LOD transition
Hi Claude,
The hybrid and DC now pass the same-LOD independent-neighbor audit: 625/625 each, with zero unmatched probes, open edges, or nonmanifold edges after deterministic ownership and T-junction conformance. MC keeps its 625/625 pass. The sharp cliff crossing also passes the 2:1 LOD fixture for all three methods within the stated 0.50 m physical budget.
The full LOD gate is still failed: MC has 5 bad cases, Hybrid 38, and DC 40. The custom loop zipper/cap does not handle the corner and overhang topologies reliably, and some transition residuals exceed budget. I recommend no production pick and no Unreal change yet; a topology-aware 2:1 transition is still needed. The report records the per-case data and failed-run ledger.
The sampled native SurfaceWorld cliff is bit-identical between MC and Hybrid (2,345 vertices, 4,496 triangles, hash `0x9d58e0603f1bf9c7`). DC differs and has much larger mean field residual. The updated offline contact sheet has the analytic fixtures, native cliff, sharp cliff transition, and LOD seam views; all 25 JPEGs are under 300 KB.
Cost and determinism outputs are retained. All 45 W1/W8/W16 hashes match their W1 references, but the standalone timing is not comparable to the optimized plugin mesher: it times a separate implementation after field-grid sampling and excludes Unreal scheduling, upload, collision cooking, and rendering. The round-3 plan is opt-in and versioned, with real collision cooking and in-game border fixtures after the standalone transition gate passes.
No `Source/` or Core source, owner project file, `.uasset`, or editor was touched. No commit, push, or stash.
— Luna