wip(assembler): round S14 -- 4-way column cache + memoized MC gradients (output unchanged)

Worker-local XY column cache uses deterministic 4-way sets at LOD3-8: LOD4
misses 8,224 -> 1,437, conflicts 7,407 -> 0 (direct-map control: 6.62 s /
43.5 worker-s vs 4.24 s / 21.7). Exact LOD0 lattice gradients memoized per
tile: MC-exclusive 242 -> 201 ns/sample, two-pass loop unchanged. Hashes
identical, no recipe bump. Primary cold trace 4.24 s full set, ring 0.45 s,
75/75 at 40 m/s; a repeat ran uniformly ~40% slower on every LOD (machine
contention suspected); tick p95 to re-check on an idle machine. CVars
voxel.CoreColumnCacheDirectMapped / voxel.CoreMCGradientMemoization.
Report: Standalone/Research/WorldAssemblerM3/Round18.

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

ARCHITECTURE.md

@@ -694,6 +694,10 @@ driven by `EditorBrush*` props.
   position → seamless across chunk borders (both sides use identical pure samples). NO per-vertex
   `GetDensityAt` (was ~6/vertex, often as costly as the whole grid). `ComputeGradientNormal` is now
   unused. Only NORMALS changed vs the old path; geometry is identical.
+  Round 18 memoizes these exact central-difference `FVector`s in per-worker scratch arrays for the
+  step-1 ShapeWorld path. Validity is reset for each tile; cache hits return the same vector and do
+  not bypass the density grid, halo, or two-pass cell loop. Keep the cache worker-local and exact;
+  do not move gradient probes above the surface-cell test.
 - **Core hybrid sharp mesher (M2)** (`CoreMesherRecipe` / `MesherRecipe`): the property remains
   opt-in and defaults to `MarchingCubesV1`; `bUseVoxelForgeCore` remains separately default-off.
   V1 preserves the round-3 feature-aware dual-contouring path for repeatability. V2 reuses the
@@ -1778,6 +1782,20 @@ shared-page A/B test, not proof that a shared cache will pay after synchronizati
 costs. Round 17 changed no density, mesh, collision, or recipe output; recipe versions remain
 unchanged. See [`Standalone/Research/WorldAssemblerM3/Round17/REPORT.md`](Standalone/Research/WorldAssemblerM3/Round17/REPORT.md).
 
+**Round 18 / S14:** the 65,536-entry worker-local XY column cache keeps the legacy direct-map
+lookup for LOD0–2 (at most 33/28/4 conflict misses in the cold audit) and uses four-way sets for
+LOD3–8. The deterministic coordinate/hash mixer and round-robin replacement cut LOD4 from 8,224
+misses / 7,407 conflict misses to 1,437 / 0, and LOD3 from 1,634 / 1,119 to 574 / 0. Exact key
+validation keeps fill order from affecting samples. `voxel.CoreColumnCacheDirectMapped=1` restores
+the S13 route as a recorded negative control.
+
+The step-1 ShapeWorld gradient memo reduced LOD0 fallback density probes from 11,474 to 3,018 at
+the sampled p95 and gradient-normal cost from 65.10 to 24.16 ns/sample; the two-pass MC invariant
+is unchanged. `voxel.CoreMCGradientMemoization=0` is its recorded negative control. No shared XY
+page was added: cross-worker repeats alone do not establish a net page-cache win, and the S8
+immutable-page lifetime rules would add ownership and memory cost. Round 18 changed no output or
+recipe version. See [`Standalone/Research/WorldAssemblerM3/Round18/REPORT.md`](Standalone/Research/WorldAssemblerM3/Round18/REPORT.md).
+
 ## 9. Multiplayer model (listen-server first, dedicated-friendly)
 
 > **Status: transport is DESIGN ONLY — nothing is networked in-tree yet** (no `Replicated`/`HasAuthority`/RPCs; a