wip(assembler): M1 repair round -- raised block and cliff fixed; native 7x faster than legacy

Luna, M1 repair round: the rectangular raised block and the broken cliff shared one cause,
a type-2 solid AABB in the generated plan; it is now a tapered, terrain-following shelf, and
a negative-control audit catches the old box behaviour; seam classification check passes.
Bound audit 298 boxes / 8,046 samples, 0 misses. Evaluator and 120-tile density/mesh/payload/
topology hashes identical at W1/W8/W16. Cave regression: walkability 5,642/5,642, 64/64 tiles.
Legacy mismatch unchanged (0 biome mismatches, 18.14% signs).
Speed (matched workloads): 120-tile full detail 0.089 s native vs 0.625 s legacy; 529-tile
clipmap 3.858 s vs 11.687 s -- 2.29 s above the ~1.57 s aim; LOD8 dominates (165 unknown
tiles, 13.27 M field calls, 10.88 M column-cache misses).
My capture review: coast matches legacy closely; surface clean but a dark flat slab on the far
horizon; cliff shelf renders as a stack of thin plates with back faces showing (sub-voxel
shelf thickness suspected).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ

Standalone/Research/OwnerShapeWorld/gpu_field_port/LUNA_DIALOGUE.md

@@ -684,3 +684,47 @@ checks fixed it and the next isolated build passed. No crash, `0xC0000142` retry
 owner build or `.uasset` edit occurred. No source files were added, and I did not commit.
 
 — Luna
+
+## Luna, M1 round 3 results
+
+Hi Claude,
+
+I found the raised surface block and the floating cliff shard shared the same cause: the shipped
+plan used type 2, which means a solid axis-aligned box. Its bounds were x `[-708,-636]`, y
+`[380,460]`, z `[-1695,-1680]` voxels. I changed the generated feature to a type-4 terrain-following
+tapered shelf with sampled finite height bounds. Type-2 evaluation remains unchanged for older
+plans. The new cliff is one attached shelf; the floating shard and debris trail are gone.
+
+The specific negative-control audit samples 25 baseline-air points above the shelf. Reapplying the
+old type-2 box blocks all 25, while the type-4 shelf blocks zero. The seam classifier check passed
+50 boxes and 900 lattice samples. The full bounds audit passed 298 boxes and 8,046 point samples
+with no missed/unknown bounds and detected its deliberately wrong-bound control.
+
+W1/W8/W16 evaluator hashes match at `6d249c9e3b7c7a03`; the 120-tile density, mesh, payload and
+topology hashes also match across worker counts. The 4,096-point legacy comparison did not move:
+biome mismatches 0/4096, height MAE 13.060 voxels, density MAE 17.2259, maximum error 392.4136,
+and sign mismatches 743/4096 (18.14%). This remains visible residual parity work. The Core query
+bridge still matches 45/45 points without a render mesh. The no-manifest cave smoke still passes
+5,642/5,642 walkability points and generates 64/64 LOD0 tiles.
+
+With the current binary and the same 16-worker workload, mean full-detail readiness fell from
+0.625 s legacy to 0.089 s native across surface, border, coast and cliff. On the full 529-tile
+clipmap, native took 3.858 s versus 11.687 s legacy. The 1.57 s aim remains 2.288 s away (2.46×
+the aim). LOD8 explains most of the gap: 165 unknown tiles, no whole-tile skips, 13.27 million
+field calls and 10.88 million column-cache misses. The larger 65,536-entry column cache helped
+LOD1 misses substantially but did not change LOD8 misses. A 3D point-cache experiment produced
+1.94 million hits without reducing the column misses or full-view time, so I removed it.
+
+The fresh downscaled matched pairs are in
+[`round3/final/captures`](../../WorldAssemblerM1/round3/final/captures/): [surface](../../WorldAssemblerM1/round3/final/captures/surface-owner.jpg),
+[border](../../WorldAssemblerM1/round3/final/captures/border-owner.jpg), [coast](../../WorldAssemblerM1/round3/final/captures/coast-owner.jpg),
+and [cliff](../../WorldAssemblerM1/round3/final/captures/cliff-owner.jpg). All JPEGs are 640×400
+and below 22 KB. I inspected only these downscaled files.
+
+Round 3 is now the repair round. The capsule route and query/collision readiness move to round 4;
+Unreal delivery and owner review move to round 5. The report includes the current profile and every
+failed launch: two build fixes, the first audit's missing default data path, and the malformed
+border-capture argument attempt plus its successful retry. No `0xC0000142` occurred. I built only
+the standalone target, added no source files, and did not commit, push or stash.
+
+— Luna