wip(assembler): M1 round 4 -- cap rebuilt (no hiding), sub-voxel audit, routes + mesh-free queries

Luna, M1 round 4 (resumed after the night stop):
- Cliff: sub-voxel audit (29,929 half-voxel XY columns, min run 3.75 voxels, 0 subvoxel runs;
  old taper control exposes 20,963) and a root-line shelf at the two-voxel minimum.
- Horizon slab = aliased LOD8 cap ceiling. The first fix hid the ceiling at LOD8 (challenged:
  enclosed, interior-lit strates); now the LOD8 cap is rebuilt deterministically from the
  native heightfield, skirts preserved; looking up and out shows the cap present.
- Routes: review manifest with 33 stable-ID route points + capsule/support constraints; capsule
  sweep 0 blocked samples / 0 unsupported centres, negative control detected.
- Server path: World::densityAt answers without render meshes. Physics cooking not available
  standalone.
- Hashes (evaluator + 2,226-tile stream) identical at W1/W8/W16. LOD8: 4.446 s for 529 tiles,
  14.25 M field calls, 165 unknown tiles.
- Captures at 60 fps cap + vsync (owner's GPU coil whine). One cliff probe crossed the ~8 GB
  guard (7.69 GiB) and was stopped before capturing.
My review: cap view matches legacy closely, no slab; native cliff is a clean attached ledge
(cleaner than legacy); a few small spikes on the ridge top.

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

@@ -728,3 +728,53 @@ border-capture argument attempt plus its successful retry. No `0xC0000142` occur
 the standalone target, added no source files, and did not commit, push or stash.
 
 — Luna
+
+## Luna, M1 round 4 results (resumed)
+
+Hi Claude,
+
+I continued from the committed `20c6d86` state and settled the LOD8 horizon problem by replacing
+the aliased coarse ceiling geometry, rather than leaving the cap absent. The root was the 256-voxel
+marching-cubes step at LOD8 and a ceiling classifier that read legacy heights while the native mesh
+used `Heightfield2D`. LOD8 now rebuilds the cap from the native plan on a deterministic 128-voxel
+XY grid, clipped to the tile's Z range and `cap >= ground`. It replaces the unmasked coarse ceiling
+triangles while preserving skirt triangles. The new triangles face down into the enclosed strate.
+The matched surface and up/out cap captures show the cap sheet and far horizon present from the
+surface. They are standalone geometry checks; interior light-leak validation remains for the owner
+build. Cap tile keys span x `-7..5`, y `-6..6`, z `-1`; 153 tiles were refined in the 2,226-tile
+determinism run. The final4 cliff overlook pair was recaptured from `(-168, 0, -380)`, yaw `0°`,
+pitch `-34°`; it is matched but still shows owner/native shape differences, so it is not claimed as
+parity. The native frame has no broad fanned-plate trail in this view.
+
+The research manifest `Standalone/Research/WorldAssemblerM1/round4/world_assembly.json` now has 33
+stable-ID route points and explicit capsule, grade, spacing and support constraints. Canonical
+serialization round-trips 7,565 bytes. The actual composed-field capsule sweep sampled 530 centers
+and 143,630 field points with zero blocked samples, unsupported centers or grade failures; its
+unsupported-route negative control detected all 530 centers. The bounds audit passed 298 boxes and
+8,046 samples, with its false-bound control detected. The shelf audit passed 29,929 half-voxel XY
+columns with zero sub-voxel runs and detected the legacy taper control's 20,963 runs.
+
+Evaluator hashes match at W1/W8/W16 (`c5e0371612e927af`). The 2,226-tile density, mesh, payload and
+topology hashes also match across W1/W8/W16. `World::densityAt` compiled and queried 78 exact points
+without generating render meshes. Readiness is separated: local field query and analytic route
+validation are ready; standalone Core does not cook physics collision; render geometry was exercised
+separately in capped viewer runs. The 529-tile LOD8 profile remains expensive: 165 unknown tiles,
+14.25 million field calls and 11.68 million column misses; full stream wall time was 4.446 seconds.
+
+All successful capture and benchmark viewer runs used VSync and a 60 FPS cap and exited afterward.
+The final4 overlook pair exited in 26.4/6.9 seconds at 0.983/1.154 GB peak working sets. An earlier
+overlook pair used a stale executable and was excluded; the corrected pair used `final4`. One
+focused cliff probe crossed the approximate 8 GB memory guard (last external sample 8.257 GB
+decimal / 7.69 GiB); the guard stopped it before capture, so that output is excluded. A black pitch-70
+cap image, one wrapper failure and a mislabeled early grid diagnostic were also excluded or rerun.
+There was no crash or `0xC0000142` retry. The highest successful capture working set was about
+1.52 GiB. I used only the standalone BuildHost-compatible target; no owner project/editor or
+`.uasset` was touched. No new source files, commit, push or stash.
+
+The remaining delivery gates are the manifest/content-hash and `bUseVoxelForgeCore` integration,
+flag-off legacy verification, owner review of the in-game surface/route and enclosed-cap lighting,
+physics collision cooking, and the matched 23.085-second workload with the ~1.57-second aim
+reported separately. See `Standalone/ASSEMBLER_M1_REPORT.md` for the full measurements, captures and
+run ledger.
+
+— Luna