wip(assembler): M3 round 2 -- islands get real mass: deep crust, separate tapered roots, arched spans

Pinch cause: SmoothStep taper drove the underside radius to zero at one point. Version 2: depth 0.75x
diameter (thin-pad control at 0.30 fails), 3-4 seeded roots, bounded contour warp + roughness, one
off-route side island, 3 m rock spans. Bounds/route/W1-W8-W16 pass. Open: legacy probe metadata says
25 cm but production legacy is 12.5 cm -- round 3 must fix the import scale contract.

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

@@ -1027,3 +1027,39 @@ files are listed there. No `Source/` or `Standalone/Source/` file, owner project
 commit, push, or stash was touched.
 
 — Luna
+
+## Luna, M3 round 2 — FloatingIslands look pass
+
+Hi Claude,
+
+Round 2 replaces the thin profile with a broad crust and three or four deterministic tapered roots
+per island. The previous underside pinch came from tapering the cross-section to zero at its bottom
+plane; version 2 uses a broad underside and independent root lobes. Bounded contour warp and surface
+roughness add broken edges inside the certified detail bounds. One side island sits above and away
+from the three-island route; the route spans are broad, arched and sharp-tag-ready. The main route is
+still a connected backbone, and the native cluster is intentionally denser than the owner's sampled
+field.
+
+The owner settings describe a stochastic site process, not a fixed count: density is 0.5 per hashed
+XY cell, spacing is 112 legacy voxels and radius fields are 24–48 voxels. I used the production
+12.5 cm legacy scale against 25 cm Core, mapping lengths by 0.5. This gives 14 m spacing and 3–6 m
+radii. The frozen probe metadata and a radius comment instead assume 25 cm; Round 3 must reconcile
+that at the snapshot producer before finalizing the Unreal import contract.
+
+The five fresh matched views are in the [contact sheet](../../WorldAssemblerM3/Round2/results/captures/contact_sheet.jpg).
+All images are below 300 KB. Bounds passed 515 boxes with zero misses and the shrunken-bound control
+was detected. All four islands measured at least 0.75 depth/diameter; the 0.30 thin-pad control
+failed. The route passed 254 capsule centers / 69,088 field samples; its raised control failed at all
+254 centers. W1/W8/W16 hashes match. On 32,768 same-physical-location queries, native was 38.06 ms
+vs 64.64 ms legacy, with 1,380 vs 130 solid samples.
+
+The full report, scale caveat, captures, run ledger and Round 3 plan are in
+[`WorldAssemblerM3/Round2/REPORT.md`](../../WorldAssemblerM3/Round2/REPORT.md). It lists one failed
+Core compile, eight failed audit iterations and one corrected renderer failure; the final build and
+audit passed, with no crash or `0xC0000142` retry. The old M1 and SurfaceWorld manifest hashes still
+match; the M2 mesher hash matrix was not replayed. No editor/viewer, project build, `.uasset`,
+`Source/` edit, commit, push or stash was used.
+
+New source file this round: `RenderFloatingIslandRound2.py`.
+
+— Luna