wip(assembler): M3 round 3 -- separated islands + Unreal island route (owner build 18:46)

Islands read as separate (91.89% open sky at mid-height), underside cavity closed,
rim teeth were capture shading. Island manifest import + BuildHost proof passed on the
pre-cache tree; M1 SurfaceWorld and M2 V1/V2 gates passed.

Known open: final XY-column cache compiled in the owner build but is unverified;
island LOD0 p95 ~155 ms misses the 133 ms sprint deadline (legacy ~43 ms); the island
manifest only fits the BuildHost fixture layout (owner's world: Core init fails,
empty world) -> round 4 builds plans inside Unreal.

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

Standalone/Research/WorldAssemblerM3/Round3/REPORT.md

@@ -0,0 +1,83 @@
+# World assembler M3 round 3
+
+Date: 2026-10-10
+
+## Status
+
+The island recipe, manifest import, BuildHost world proof, and matched offline captures are in place. The final profiling pass found a remaining LOD0 sprint budget miss, and a follow-up XY-column cache was then added to `FloatingIslandsInternal.h`. That last source edit has **not** been compiled or exercised: the command runner began failing to start with Windows status `0xC0000142` after a completed profile run. Therefore the verified BuildHost results below apply to the root-parameter-cache version immediately before the XY-column-cache edit, not to the current tree. Treat round 3 as **not fully verified / do not hand off the current binary yet** until that edit is either verified or reverted and rebuilt.
+
+## Scale contract
+
+The owner's legacy voxel size is **12.5 cm** (`Source/VoxelForge/Public/VoxelTypes.h`, `VOXEL_SIZE`). Core operates at **25 cm**, so owner-to-Core lengths multiply by **0.5** and Core-to-owner lengths multiply by **2**. Island radii are 24–48 owner voxels = 3–6 m = 12–24 Core voxels. Site spacing is 112 owner voxels = 14 m = 56 Core voxels. Standalone's old 25 cm metadata was not authoritative; the import validator and audit metadata now pin 12.5/25 cm and the 1:2 ratio. The radius comment is corrected to this contract.
+
+## Shape review and measurements
+
+Round 3's downscaled, matched offline captures show four distinct masses with sky visible between them; the three connectors remain traversable and read as narrow spans. The underside center is a shaded solid buttress/dimple, not a through-hole. The edge capture does not show the former sawtooth rim; the crust edge is smooth at the displayed scale. These are offline CPU renders, not captures from the owner's Unreal viewport.
+
+At the sampled mid-height plane, the open-sky measure is **1,462 / 1,591 = 91.89%** of the cluster XY bounding box. Sampling is at 1 m spacing, Core Z `-964.039`. Island radii are 3–6 m; connector full width is 1.25 m, thickness 1 m, blend 0.75 m, and roughness 0.25 m. The plan has four islands, three connectors, and 28 route segments. The route audit checks 68,816 capsule samples and reports no blocked, unsupported, or over-grade samples; minimum clearance is 99 subvoxel samples.
+
+The mass audit checked 515 bounds / 13,851 samples with no missed island bounds. The thin-pad negative control was caught on all four islands (measured ratio 0.2999 against required 0.70; passing minimum was 0.75). The raised-route negative control was also caught. The underside audit samples the center column at nine points; rerun the audit after resolving the unverified cache edit to retain that numeric output in `results/audit.json`.
+
+## Import and season hashing
+
+Import `Standalone/Research/WorldAssemblerM3/Round3/results/world_assembly_floating_islands.json` through the existing `UVoxelSettings` manifest JSON import (`SourceWorldAssemblyManifestJson` / **Import World Assembly Manifest JSON**). Then assign that settings asset to `AVoxelWorld` and enable **Voxel World → Core → Use VoxelForge Core** (`bUseVoxelForgeCore`). The recipe remains on the Core world-assembly route. Its assembly content hash enters the season hash only when Core is enabled; Core-off keeps the legacy season identity. Water remains fixed for the season and player edits remain unrestricted.
+
+Verified artifact hashes:
+
+- Plan: `61fd58f65c5fe157`
+- Recipe: `ef1340d32b072ba3`
+- Nested floating-island manifest: `54160b4d25d0279f`
+- World assembly content: `94b563f1ce622def`
+- W1/W8/W16 standalone determinism: `5764b702482740e8`
+
+## BuildHost proof recorded before the final unverified edit
+
+BuildHost runs used `Tools/BuildVoxelForgeCoreHost.ps1` and the test scripts with a 900 second timeout, BelowNormal priority, and at most four parallel build actions. No owner project/editor was opened or built. Measured process peaks stayed below 4 GB. The passing island run recorded:
+
+- 2,553 standalone tile hashes matched; Core reported 512 LOD0 tiles, 448 LOD1 tiles, 432 LOD transitions, 573 local-coverage checks, and 3,438 neighbor-face checks.
+- Aggregate hashes: density `e46d4280f73f03f3`, mesh `a7c92c6ce22ba08b3`, payload `040e800d58e1c2f1`, topology `da689102032f94da`.
+- Coverage completed; collision was cooked and capsule contact passed on both an island top and a span.
+- Height/anchor agreement passed for the island and span; content anchors reported 14 trees and no floaters. Corrected strate and atmosphere strate both reported 1; water was enabled at local `-49,376 cm` and unchanged.
+- Blueprint edit passed: carve/fill on an island with legacy radius 6 (=75 cm) and strength 10; the 90 cm outside sample had zero density delta.
+- Core-off passed with legacy density unchanged; the Core-only hash was omitted while Core was off.
+- SurfaceWorld M1 regression test passed with the original worker hash `0x8cc18e2981c50c4e`; HybridV1 and HybridV2 M2 recipe fixtures passed.
+
+Logs are in `results/m1_surfaceworld_buildhost.*`, `results/m2_hybrid_v1_buildhost.*`, `results/m2_hybrid_v2_buildhost.*`, and `results/m3_floating_islands_buildhost.*`. These pre-profile logs do not prove the final XY-column-cache edit.
+
+## Streaming profile and remaining work
+
+The final instrumented BuildHost commandlet completed and its internal validator printed PASS. The outer PowerShell wrapper then exited 1 because its console-title channel failed (`0xE9`); this is recorded as a wrapper failure despite the commandlet pass marker. The timing profile showed density/meshing dominates; classifier and column work were negligible.
+
+The same measured per-tile p95 is compared with the available time at each movement speed: at 600 cm/s the deadlines are 1,333.333 ms for LOD0 and 2,666.667 ms for LOD1; at 6,000 cm/s they are 133.333 ms and 266.667 ms, respectively. Core and legacy timings from the instrumented run were:
+
+| LOD | Tile kind | Core p95 | Legacy p95 | Walk deadline / result | Sprint deadline / result |
+|---|---|---:|---:|---|---|
+| 0 | Air | 0.017 ms | 38.842 ms | 1,333.333 ms / pass | 133.333 ms / pass |
+| 0 | Island | 157.671 ms | 43.702 ms | 1,333.333 ms / pass | 133.333 ms / 24.338 ms late |
+| 0 | Mixed/span | 153.082 ms | 42.146 ms | 1,333.333 ms / pass | 133.333 ms / 19.749 ms late |
+| 1 | Air | 0.009 ms | 46.549 ms | 2,666.667 ms / pass | 266.667 ms / pass |
+| 1 | Mixed/span | 151.685 ms | 42.701 ms | 2,666.667 ms / pass | 266.667 ms / pass |
+
+LOD1 mixed was 151.685 ms against its 266.667 ms sprint deadline; LOD1 air was 0.009 ms. The root-parameter cache preserved hashes but did not clear the LOD0 sprint miss. The XY-column cache was added afterward to target the measured mesher/density hot path; it has not been compiled or measured. Resolve that before claiming the streaming goal complete.
+
+## Run ledger and incidents
+
+- SurfaceWorld M1, HybridV1 M2, HybridV2 M2, and the island BuildHost proof all passed on the verified source state described above.
+- Earlier island BuildHost attempts, in order, failed as follows: (1) desired tiles were 2,227 instead of 2,553 because the route player was absent; (2) the GameMode spawned an origin pawn; (3) possession caused 2,557 desired tiles because dynamic collision support was active; (4) the support tile key did not match; (5) disabling the collision gate restored 2,553 tiles but exposed a hash mismatch at `(-2,0,-31)`; (6) `CreateLocalPlayer(true)` hung and was manually stopped after roughly 3–4 minutes; (7) the base UPlayer fixture lacked controller registration/root transform and failed preflight after 5:10; (8) explicit `AddController`, root setup, and `SetActorLocation` fixed the fixture and the island proof passed. The manual stop was not a crash.
+- Earlier compile failures included an invalid `DuplicateObject` overload, private helper access, an unresolved `GetEffectiveWorldSeed` export, and integer narrowing. They were corrected; subsequent builds passed.
+- One initial standalone build used the wrong `-Workers` switch and failed immediately; `-CompileWorkers 4` succeeded.
+- M1 threshold experiments before restoring the original chunk mapping failed at LOD8 edge bounds: threshold 4.0 measured 52,075.827 > 51,200 at tile `(2,0,-1)`; 4.125 measured 53,960.427 > 52,800; 4.25 measured 54,806.919 > 54,400. Threshold 4.5 passed that edge check but left no full-span LOD0 candidate. Restoring SurfaceWorld's original 32-voxel chunk mapping, using the separate 16-Core-voxel island mapping, and restoring threshold 4.0 made M1 pass.
+- The last commandlet itself passed, but the wrapper returned 1 on console error `0xE9`. Subsequent PowerShell and `cmd.exe` launches returned `0xC0000142`; the requested retry was made, then additional shell inspection launches also failed. No further build was attempted. No new `UECC-*` folder was reported by the latest BuildHost crash check; only historical folders dated 2026-09-17 and 2026-10-08 were present.
+- No `0xC0000142` BuildHost retry was observed in the completed runs. Peak recorded BuildHost test process memory was about 3.98 GB.
+
+## Owner hand-off
+
+After the final cache edit is verified, import the manifest above into `UVoxelSettings`, assign the settings asset to the world, and turn on `bUseVoxelForgeCore`. To find the cluster, start at the actor-local route point `(-380, 631.8, -23,837) cm` and follow the route toward the middle island (roughly +X / -Y). The center island is near actor-local `(600, -306, -23,872) cm`; the span contact is near `(113, 161, -23,860) cm`.
+
+For an underside view, use a free camera near `(600, -306, -24,750) cm`, looking upward toward `(600, -306, -24,300) cm`. For an overview, start near `(680, -5,000, -21,500) cm` and look toward `(678, -556, -24,100) cm`. These are actor-local suggestions; apply the actor transform in the level. The contact sheet and matched captures are under this Round3 `results/` directory and are downscaled offline renders; BuildHost validation uses null RHI.
+
+Likely Unreal compile spots are `VoxelSettings.h/.cpp`, `VoxelContentManager.h/.cpp`, `VoxelWorld.cpp`, `VoxelForgeCoreAdapter.cpp`, and `VoxelForgeExploreCommandlet.cpp`; canonical Core changes are in `Standalone/Core/include/VoxelForgeCore/VoxelForgeCore.h`, `Standalone/Core/src/ShapeWorldInternal.h`, `FloatingIslandsInternal.h`, and `WorldAssemblerInternal.h`.
+
+## Source and artifact inventory
+
+Round 3 adds these research/support sources: `Standalone/Research/WorldAssemblerM3/Round3/FloatingIslandsAudit.cpp`, `BuildFloatingIslandsAudit.ps1`, and `RenderFloatingIslandRound3.py`. Existing source edits include `Standalone/Source/OwnerWorldLoader.h`, `Standalone/Core/src/FloatingIslandsInternal.h`, `Standalone/Core/src/ShapeWorldInternal.h`, and `Source/VoxelForgeEditor/Private/VoxelForgeExploreCommandlet.cpp`. There is no new Unreal runtime source file for the manifest path. BuildHost scripts and `.log`/capture/JSON outputs are evidence/support artifacts.