standalone: per-tile speed at plugin level, streaming like the game, softer lighting

Cause found by profiling: room/SDF graph-cache churn (exact XY bounds + sample order rebuilt it constantly). Region grouping (one LOD3 tile: 1,084 -> 229 builds), taller-tile traversal for LOD6 (36,480 -> 12,160), extent-dependent cache slots, lambdas instead of recursive std::function. Serial ms/tile, 64 nearest tiles per level (before -> after): L0 50.4 -> 50.6, L1 112 -> 110, L2 24 -> 22, L3 223 -> 73, L4 269 -> 134, L5 393 -> 127, L6 283 -> 145. Hashes of all levels unchanged; parity 3,436/3,436 bit-identical. Full clipmap, 16 workers BelowNormal: full detail 66.6 s, all 12,054 tiles 175 s. In-stream L0 tail is high (median 83 ms, mean 198 ms, p95 669 ms) vs 51 ms serial. Streaming: recentre keeps the intersection, generates only missing tiles nearest/finest first; one crossing = exactly 441 new tiles in 6.1 s, no stale results. Lighting: normals already come from the density gradient; the striping was the squared headlamp; broad linear fill + warmer palette. Daytime CPU rule documented in Standalone/README.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
2026-10-03 06:49:41 +02:00
1 parent 3f300b912a
commit 81b7d6d0ff
@@ -28,3 +28,5 @@

 # Keep every design / doc markdown at any depth
 !*.md
+/Standalone/data/*.log
+/Standalone/data/*.err
@@ -54,7 +54,7 @@ $RspLines += @('/link','d3d11.lib','dxgi.lib','d3dcompiler.lib','user32.lib','gd

 Push-Location $BuildDir
 try {
-    $CommandLine = 'call "{0}" >nul && cl.exe @"{1}"' -f $VcVars,$ResponsePath
+    $CommandLine = 'call "{0}" >nul && start /belownormal /wait cl.exe @"{1}"' -f $VcVars,$ResponsePath
    & $env:ComSpec /d /s /c $CommandLine
    if ($LASTEXITCODE -ne 0) { throw "Standalone MSVC build failed with exit code $LASTEXITCODE" }
 } finally {
@@ -8,11 +8,13 @@
 #include <bit>
 #include <cassert>
 #include <chrono>
+#include <charconv>
 #include <cctype>
 #include <cstdarg>
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <functional>
 #include <immintrin.h>
@@ -632,8 +634,73 @@ struct FWriteScopeLock { explicit FWriteScopeLock(FRWLock&L):Lock(L.Mutex){} std
 enum EConsoleVariableFlags { ECVF_Default=0, ECVF_SetByCommandline=1, ECVF_SetByCode=2 };
 class FAutoConsoleVariableRef { public:FAutoConsoleVariableRef(const char*,int32&,const char*,uint32=0){} FAutoConsoleVariableRef(const char*,float&,const char*,uint32=0){} FAutoConsoleVariableRef(const char*,bool&,const char*,uint32=0){} };
 template<class T> class TAutoConsoleVariable { public:TAutoConsoleVariable(const char*,T Default,const char*,uint32=0):Value(Default){}T GetValueOnAnyThread()const{return Value.load();}T GetValueOnGameThread()const{return Value.load();}void Set(T V,uint32=0){Value.store(V);} TAutoConsoleVariable*operator->(){return this;}private:std::atomic<T>Value;};
-struct FCommandLine { static const char* Get(){return "";} };
-struct FParse { static bool Value(const char*,const char*,int32&){return false;} static bool Value(const char*,const char*,float&){return false;} static bool Value(const char*,const char*,FString&){return false;} static bool Param(const char*,const char*){return false;} };
+struct FCommandLine
+{
+    inline static std::string Arguments;
+    static const char* Get(){return Arguments.c_str();}
+    static void Set(std::string_view InArguments){Arguments.assign(InArguments);}
+};
+struct FParse
+{
+    static bool ReadValue(const char* CommandLine,const char* Key,std::string& Out)
+    {
+        if(!CommandLine||!Key||!*Key)return false;
+        const std::string_view Command(CommandLine),Token(Key);
+        std::size_t Position=0;
+        while((Position=Command.find(Token,Position))!=std::string_view::npos)
+        {
+            const bool bTokenStart=Position==0||Command[Position-1]==' '
+                ||Command[Position-1]=='\t'||Command[Position-1]=='-';
+            const std::size_t ValueStart=Position+Token.size();
+            if(bTokenStart&&ValueStart<=Command.size())
+            {
+                std::size_t ValueEnd=Command.find_first_of(" \t\r\n",ValueStart);
+                if(ValueEnd==std::string_view::npos)ValueEnd=Command.size();
+                if(ValueEnd>ValueStart)
+                {Out.assign(Command.substr(ValueStart,ValueEnd-ValueStart));return true;}
+            }
+            Position+=Token.size();
+        }
+        return false;
+    }
+    static bool Value(const char* CommandLine,const char* Key,int32& Out)
+    {
+        std::string Text;
+        if(!ReadValue(CommandLine,Key,Text))return false;
+        const auto Result=std::from_chars(Text.data(),Text.data()+Text.size(),Out);
+        return Result.ec==std::errc{}&&Result.ptr==Text.data()+Text.size();
+    }
+    static bool Value(const char* CommandLine,const char* Key,float& Out)
+    {
+        std::string Text;
+        if(!ReadValue(CommandLine,Key,Text))return false;
+        char* End=nullptr;Out=std::strtof(Text.c_str(),&End);
+        return End==Text.c_str()+Text.size();
+    }
+    static bool Value(const char* CommandLine,const char* Key,FString& Out)
+    {
+        std::string Text;
+        if(!ReadValue(CommandLine,Key,Text))return false;
+        Out=FString(Text);return true;
+    }
+    static bool Param(const char* CommandLine,const char* Name)
+    {
+        if(!CommandLine||!Name||!*Name)return false;
+        const std::string_view Command(CommandLine),Token(Name);
+        std::size_t Position=0;
+        while((Position=Command.find(Token,Position))!=std::string_view::npos)
+        {
+            const bool bTokenStart=Position==0||Command[Position-1]==' '
+                ||Command[Position-1]=='\t'||Command[Position-1]=='-';
+            const std::size_t End=Position+Token.size();
+            const bool bTokenEnd=End==Command.size()||Command[End]==' '
+                ||Command[End]=='\t'||Command[End]=='\r'||Command[End]=='\n';
+            if(bTokenStart&&bTokenEnd)return true;
+            Position=End;
+        }
+        return false;
+    }
+};
 struct FCrc { static uint32 MemCrc32(const void*Data,int32 Length,uint32 Crc=0){const auto*P=static_cast<const uint8*>(Data);uint32 V=~Crc;for(int32 I=0;I<Length;++I){V^=P[I];for(int J=0;J<8;++J)V=(V>>1)^(0xEDB88320u&static_cast<uint32>(-static_cast<int32>(V&1)));}return ~V;} static uint32 StrCrc32(const char*S,uint32 Crc=0){return MemCrc32(S,static_cast<int32>(std::strlen(S)),Crc);} };
 template<class T> using TAtomic = std::atomic<T>;
 class FThreadSafeCounter { public:int32 Increment(){return ++Value;}int32 Decrement(){return --Value;}int32 GetValue()const{return Value.load();}void Reset(){Value.store(0);}private:std::atomic<int32>Value{0};};
@@ -1,8 +1,8 @@
 # VoxelForge standalone world

 This C++20 program loads the owner snapshot from `data/owner_world.json`, runs the copied
-VoxelForge density pipeline and marching-cubes mesher, builds all clipmap levels with six worker
-threads, and displays them through Direct3D 11. Unreal is not linked or needed at run time.
+VoxelForge density pipeline and marching-cubes mesher, streams the clipmap with six worker threads
+by default, and displays it through Direct3D 11. Unreal is not linked or needed at run time.

 ## Build and launch

@@ -14,7 +14,18 @@ From the plugin root, run:
 ```

 The build uses the configured VS 2022 `vcvars64.bat`, MSVC `/std:c++20 /O2 /arch:AVX2 /fp:precise`.
-The executable finds the JSON beside itself under `Standalone\data\`.
+`Build.ps1` starts MSVC at BelowNormal priority and does not enable `/MP`.
+Pass another executable name to build beside any viewer that may be open:
+`.\Standalone\Build.ps1 -OutputName VoxelForgeStandalone_test.exe`. The executable finds the
+JSON beside itself under `Standalone\data\`.
+
+### Daytime resource limit
+
+Run tests and benchmarks at Idle or BelowNormal priority and use no more than 8 of this PC's 16
+logical CPUs. The default viewer uses six workers. Use `Build\LaunchBelowNormal.ps1` for any test or
+benchmark process; it starts the child at BelowNormal and verifies the priority. Builds use no more
+than four compiler processes (the current build is serial) and retain `/fp:precise`. The all-core
+16-worker runs documented below were the owner's one-night exception.

 ## Controls

@@ -23,8 +34,9 @@ The executable finds the JSON beside itself under `Standalone\data\`.
 - F toggles wireframe.
 - 1 through 7 toggle clipmap levels 0 through 6 individually; 0 toggles every level.
 - N/P changes the seed. T switches between the owner's base tunnels and the research-shaped round,
-  winding tunnel option. Either action rebuilds the clipmap. Moving three chunks from its center
-  also recenters and rebuilds it.
+  winding tunnel option; `--round-tunnels` starts in that variant. Either action rebuilds the clipmap.
+  Crossing a chunk recenters streaming:
+  tiles still desired stay resident, only new tiles are generated, and leaving tiles are dropped.
 - Esc exits.

 The console reports density parity at startup, each level's generation time, and the total time.
@@ -50,10 +62,11 @@ and seed. The export also records their geometry in JSON for reference.
 - Standalone `VOXEL_SIZE` is 25 cm; the plugin's checked-out value remains 12.5 cm. No other voxel
  conversion constant needed changing. Mesh vertices stay in centimetres in the copied mesher and
  the viewer converts them to metres at upload.
- The compatibility `TMap` and `TSet` are vector-backed containers, so lookup is linear and their
-  iteration follows insertion order rather than Unreal hash-table order. Owner strate slots and
-  passages are read from ordered JSON arrays. Perlin noise and deterministic `DetSinCos` are copied
-  unchanged; MSVC uses `/fp:precise` without fast-math/FMA contraction.
+- The compatibility `TMap` uses an open-addressed hash index with stable pair storage;
+  `TSet` uses that index for lookup. Iteration follows insertion order rather than Unreal hash-table
+  order. Owner strate slots and passages are read from ordered JSON arrays. Perlin noise and
+  deterministic `DetSinCos` are copied unchanged; MSVC uses `/fp:precise` without fast-math/FMA
+  contraction.
 - Unreal reflection and UObject construction are replaced by explicit JSON-to-holder mapping.
  Density-relevant settings, resolved strate layouts, passages, biome parameters, and terrain-op
  assets are loaded. Runtime actor/material assets are not instantiated; the viewer uses one lit
@@ -98,3 +111,136 @@ comparing later speed changes.
 | 5 | 23.809 | 294 | 199,503 | 2.02 M |
 | 6 | 296.304 | 4,704 | 4,032,866 | 32.26 M |
 | **Total** | **515.117** | **12,054** | **6,605,189** | **253.86 M** |
+
+## Profiling result and speed change
+
+The compatibility layer's old `TMap`/`TSet` linear lookups and heap-only vector growth were
+replaced with an open-addressed hash index, stable pair storage, small inline storage, and a
+thread-local pooled allocator. That removes known container and allocation mismatches with the
+Unreal path, but did not improve the earlier LOD0 timing (72.1 s before, 74.1 s after); profiling
+showed those were not the dominant cost. The same audit also found two recursive room-graph walks
+wrapped in `std::function`, which could allocate and dispatch indirectly on every graph build. They
+now use self-recursive generic lambdas with the same depth-first visit order, removing that wrapper
+overhead from the generation hot path. In matched profile runs, RoomGraphBuild fell from 4.531 to
+4.417 s for 64 LOD5 tiles and from 5.210 to 5.038 s for 64 LOD6 tiles; graph build counts and
+density hashes were unchanged.
+
+Per-stage timers and cache counters found the real hotspot in `GetDensityAt`: the room/SDF graph
+cache is keyed by exact XY bounds, and the original X-fast lattice walk kept rebuilding it. On one
+LOD3 tile, cache-region grouping cut graph builds from 1,084 to 229 while writing samples to their
+original lattice indices. A 64-tile LOD6 profile attributed 13.189 s of 19.874 s to 36,480 room
+graph builds. For tiles wider than 16 chunks, keeping each XY column's Z run contiguous reduced
+that to 12,160 builds and 5.210 s of 10.972 s, with the same lattice hash. The final recursive-
+lambda profile measured 5.038 s in room-graph builds and 10.600 s total for the same 64 LOD6 tiles.
+Finer tiles retain the plane-first traversal, which is faster for their shorter cache runs.
+
+Cache-slot experiments found a second scale-specific case: 32-chunk LOD5 tiles improve with two
+room graph cache slots (26,926 to 9,700 builds; profiled RoomGraphBuild 10.151 to 4.475 s across
+64 tiles). Two slots regress 64-chunk LOD6 tiles (12,160 to 23,104 builds), so the generator uses
+two only for 32-chunk tile contexts and one otherwise. The selected-tile density hash is unchanged.
+
+The controlled before/after run below meshes the 64 nearest tiles for each LOD serially, without
+profiling or hash collection. It uses the same owner seed and camera-centered desired set. The
+selected-tile lattice hashes from a separate final run are identical at every LOD.
+
+| LOD | 64-tile density hash |
+|---:|---:|
+| 0 | `0x3d31a090650399d3` |
+| 1 | `0x6150c607c7a5c76f` |
+| 2 | `0x9f1d560bc14e67f1` |
+| 3 | `0xdd3e76020ce69cd7` |
+| 4 | `0x59a209a5db034b46` |
+| 5 | `0x1fdeb60160e4216f` |
+| 6 | `0xbac69461f812493d` |
+
+| LOD | Before ms/tile | After ms/tile | After p50 | After p95 |
+|---:|---:|---:|---:|---:|
+| 0 | 50.4 | 50.551 | 50.521 | 63.148 |
+| 1 | 112.3 | 110.175 | 33.030 | 407.767 |
+| 2 | 23.9 | 22.430 | 12.015 | 51.078 |
+| 3 | 223.0 | 73.482 | 69.614 | 94.331 |
+| 4 | 269.0 | 133.512 | 98.003 | 214.854 |
+| 5 | 393.1 | 127.325 | 116.386 | 173.401 |
+| 6 | 282.8 | 145.457 | 140.952 | 183.234 |
+
+Run mode: BelowNormal priority, serial headless benchmark, 64 nearest tiles per LOD, profiling and
+hash collection off; final executable `VoxelForgeStandalone_night20.exe`. The before timings are
+from the earlier 64-tile baseline. The user's prior plugin reference is about 95 ms worker time for
+a 25 cm LOD0 tile. The final nearest-tile LOD0 average is 50.6 ms. In the complete streamed run,
+LOD0 median was 82.9 ms, but its mean was 197.6 ms and p95 was 668.8 ms. The plugin was not timed
+on the same tile coordinates, so the rough reference comparison is not apples-to-apples and the
+streamed tail does not meet that target yet.
+
+With 16 BelowNormal workers, the final full run completed all 12,054 tiles in 175.056 s, down from
+the historical 515.117 s six-worker run. It reached full detail after 66.565 s. Worker time per
+level in the final run was:
+
+| LOD | Worker sum | Tiles | Avg/tile | p50 | p95 |
+|---:|---:|---:|---:|---:|---:|
+| 0 | 542.284 s | 2,744 | 197.625 ms | 82.946 ms | 668.796 ms |
+| 1 | 501.518 s | 2,009 | 249.635 ms | 89.990 ms | 742.258 ms |
+| 2 | 68.712 s | 1,274 | 53.934 ms | 55.633 ms | 120.102 ms |
+| 3 | 77.055 s | 588 | 131.047 ms | 122.697 ms | 212.157 ms |
+| 4 | 103.552 s | 441 | 234.812 ms | 192.137 ms | 418.249 ms |
+| 5 | 66.689 s | 294 | 226.834 ms | 223.139 ms | 328.030 ms |
+| 6 | 1,439.125 s | 4,704 | 305.936 ms | 303.062 ms | 391.440 ms |
+
+The worker sums are concurrent wall-time measurements and therefore add to more than the 175.056 s
+clipmap wall time. The historical and final full runs used different worker counts; the serial
+64-tile table is the controlled per-tile comparison. The full run also passed density parity at
+3,436/3,436 samples with zero mismatches.
+
+## Streaming and uploads
+
+The viewer rebuilds the desired tile set at each chunk crossing, keeps the intersection resident,
+cancels queued jobs that left the set, and enqueues only missing tiles. A heap orders by camera
+distance, then prefers finer LOD at equal distance. Workers generate CPU meshes concurrently. D3D11
+uploads are limited to four batches or 2 ms per frame. A leaving or changed batch keeps its prior GPU
+buffer until its replacement is uploaded, avoiding a hole while the batch is rebuilt.
+
+The final 16-worker run reached full detail 66.565 s after spawn (4,753 LOD0–1 desired tiles; 5,173
+resident through LOD1) and the full clipmap at 175.056 s. In a controlled one-chunk crossing on the
+same executable, 4,856 tiles stayed resident, 441 left, and only 441 new jobs were submitted. Those
+jobs covered 18.91 M density samples, used 88.573 worker-seconds, and finished in 6.108 s wall time;
+all 441 were generated with no stale results or queued cancellations. The upload budget remained
+four batches / 2 ms per frame.
+
+For a deterministic movement check, `--test-one-chunk-crossing` moves once after full-detail
+readiness and closes the test viewer after that crossing's jobs and bounded uploads drain.
+
+## Lighting comparison
+
+The copied plugin mesher already computes central-difference density-gradient normals on the sampled
+density lattice and interpolates them along marching-cubes edges. The standalone passes those
+normals through to D3D11, so replacing face normals was unnecessary. Profiling measured this existing
+gradient stage at 2.575 ms across 64 LOD0 benchmark tiles (about 72 ms over all 2,744 streamed LOD0
+tiles); this round adds no density samples or meshing work for normals.
+
+The visible striping was amplified by the viewer's squared, high-intensity camera-facing light. The
+shader now uses a broad linear camera fill at similar average strength. Geometry, density, and LOD
+are unchanged; the shading path uses fewer multiplications. The D3D11 pixel-shader adjustment has
+no CPU mesh-generation cost.
+
+Both captures use the default seed, default camera, 1280x800 view, and are saved from the D3D11 back
+buffer after all full-detail tiles upload:
+
+| Before | After |
+|---|---|
+| ![Before lighting](data/night_lighting_before.png) | ![After lighting](data/night_lighting_after.png) |
+
+## Tunnel palette pass
+
+As a separate, render-only visual pass, the neutral blue-gray rock constants were changed to warm
+limestone and ochre, with a softer cyan camera fill. This follows the color balance of the approved
+cave reference while leaving its mesh, 25 cm resolution, and detail untouched. The shader changes
+only existing constant values and adds no texture reads or operations; CPU density and meshing cost is
+zero. The matched 16-worker captures generated the same 7,521,175 vertices and 41,506,623 indices;
+full-clipmap wall times were 172.772 s before and 171.676 s after, within run variance.
+
+Both images show seed 0 with the round/winding tunnel field, at 1280x800 after the full clipmap
+uploaded. Camera: `(-11.75, -123.25, -19.85)` m, yaw 90 degrees, pitch 0. Capture mode used 16
+BelowNormal workers:
+
+| Before palette pass | After palette pass |
+|---|---|
+| ![Before tunnel palette](data/night18_tunnel_view_before.png) | ![After tunnel palette](data/night19_tunnel_view_after.png) |
@@ -1,30 +1,74 @@
 # Shared C++ library plan

+This round keeps the standalone copy in place. The next extraction should make one implementation
+serve both hosts; it should not turn the compatibility layer into a second public API.
+
 ## Library contents

-Move the deterministic CPU world pipeline into a portable `VoxelForgeCore` library: plain voxel
-and world types, math/noise, density operations and caches, cave and passage fields, strate
-composition, marching cubes, and desired clipmap tile selection. Keep renderer, file loading,
-asset/reflection access, gameplay actors, and task scheduling in the host applications.
+Move the deterministic CPU pipeline into `VoxelForgeCore`:
+
+- Plain voxel, tile, vector, bounds, and resolved-world input types.
+- Deterministic math, noise, density operations, cave and passage fields, strate composition, and
+  the room/SDF caches they use.
+- Marching-cubes tables and meshing, including the density-lattice gradient normals.
+- Clipmap desired-set calculation and the shared density evaluation entry points.
+
+Keep JSON and Unreal asset loading, material/actor creation, D3D11 and Unreal rendering, collision,
+worker scheduling, cancellation policy, and per-frame mesh upload budgets in the host applications.
+The cache-region traversal found in this round belongs with the density/meshing implementation.
+The worker-local room graph cache is keyed by exact XY bounds and stratum. Grouping samples by the
+legacy cache regions reduced a profiled LOD3 tile from 1,084 graph builds to 229. At coarser sizes,
+plane-first traversal still thrashes the same XY windows across a tall Z run, so tiles wider than 16
+chunks keep each XY column's Z samples contiguous. The profiled 64-tile LOD6 case fell from 36,480
+graph builds (13.189 s in graph construction) to 12,160 (5.210 s), with the same density hash.
+
+Cache capacity is also scale-dependent: a 32-chunk LOD5 tile benefits from two cached stratum
+windows (26,926 to 9,700 builds across 64 tiles), while a 64-chunk LOD6 tile regresses with two
+slots (12,160 to 23,104), so the current generator selects two only for 32-chunk tile contexts and
+one otherwise. Replacing the room-graph `std::function` recursion then reduced profiled LOD5
+RoomGraphBuild from 4.531 to 4.417 s and LOD6 from 5.210 to 5.038 s across the same 64 tiles,
+without changing graph build counts or density hashes. These timings are profiling evidence, not a
+cross-host benchmark. The shared API
+should own the traversal and cache policy, with parity tests covering both traversal paths while
+writing the same original density-grid indices.
+
+The compatibility-map and allocator change did not move the original LOD0 run; those changes remove
+known linear-lookup and heap-growth gaps, but stage profiling put the cost in density evaluation and
+room-graph cache churn. The hot room-graph walks also used an owning `std::function` compatibility
+alias for recursion; replacing them with generic self-recursive lambdas kept the same depth-first
+order and removed repeated type-erasure/allocation overhead. Keep that code in the core rather than
+carrying the UE-style callback alias into it. Viewer lighting and its palette remain host renderer
+choices and stay outside the shared generator.

 ## Compatibility boundary

-The library should expose standard C++ value types and explicit immutable world inputs. Unreal
-adapters convert `UObject` settings/assets into those inputs and convert mesh results back to
-Unreal types. The standalone loader maps its JSON snapshot into the same inputs. No Unreal headers,
-reflection macros, `UObject`, or host allocator should cross the library boundary.
+The core should expose standard C++ value types, explicit immutable resolved inputs, and owned mesh
+results. No Unreal headers, reflection macros, `UObject`, Unreal containers, renderer types, or host
+allocator should cross the boundary. The plugin adapter converts settings and loaded assets into
+the resolved input snapshot, then converts the returned vertices, indices, and normals into Unreal
+mesh types. The standalone JSON loader maps its exported snapshot into the same input snapshot.
+
+Do not move scheduling into the library. Each host should pass an immutable generation context to a
+worker and receive an owned result; this lets the plugin retain its task admission and game-thread
+rules while the viewer retains its worker pool and bounded D3D11 upload pump.

 ## Plugin integration

-Build the core as one C++20 static library and link it from both the standalone viewer and the
-VoxelForge Unreal module. The plugin keeps ownership of asset loading, worker admission,
-cancellation, collision, game-thread mesh application, and per-frame upload budgets; workers call
-the shared density and meshing API with immutable contexts. The standalone keeps its D3D11 renderer
-and JSON loader and calls the same API.
+Add a `VoxelForgeCore` C++20 target and make both builds compile or link that target from the same
+source files. For Unreal, first prototype the Unreal Build Tool integration: either a small core
+module that compiles the library sources or a platform-built static library with explicit build
+rules. Prefer source integration if packaging a binary per Unreal platform would add maintenance.
+The plugin then replaces its copied evaluator and mesher calls with the core API behind a thin
+adapter. Preserve plugin-owned UObject resolution, task cancellation, collision, game-thread mesh
+application, and upload budgets. The standalone keeps its JSON loader and D3D11 renderer.

 ## Preventing drift before extraction

-Until the move lands, keep the standalone copies synchronized from the plugin source with a small
-manifest of shared files and a checked sync script. The script should compare hashes and refuse an
-ambiguous two-sided edit. Run the exported 3,436-sample bitwise parity corpus after every density
-change, then remove the duplicate source only when both hosts build against the shared library.
+Until the common target exists, keep a manifest of mirrored files and a sync/check script that names
+the canonical source for each file, compares hashes, and refuses when both copies have changed.
+Avoid hand-editing only one copy without recording that the edit is pending synchronization. After
+each density or meshing change, run the 3,436-sample bitwise corpus and the selected-tile density
+hash check. Keep `/fp:precise`; do not accept evaluator changes that alter density bits as a speed
+shortcut. A standalone-only optimization must be marked as pending synchronization until it is
+ported or explicitly rejected for the canonical core. Delete the duplicate standalone source only
+after both host builds use the common target and the parity corpus passes through each adapter.
@@ -1,6 +1,7 @@
 #include "OwnerWorldLoader.h"
 #include "StreamingWorld.h"
 #include "StandaloneResearchTunnels.h"
+#include "VoxelDensityProfile.h"

 #ifdef TEXT
 #undef TEXT
@@ -18,6 +19,8 @@
 #include <windowsx.h>

 #include <bit>
+#include <cstdio>
+#include <cstdlib>
 #include <filesystem>
 #include <fstream>
 #include <iomanip>
@@ -105,9 +108,16 @@ namespace
    };
    struct FCrossingWork
    {
+        uint64 SubmittedJobs=0;
+        uint64 PendingJobs=0;
+        uint64 FinishedJobs=0;
+        uint64 CancelledQueuedJobs=0;
+        uint64 StaleResults=0;
        uint64 Tiles=0;
        uint64 Samples=0;
        double WorkerSeconds=0.0;
+        std::chrono::steady_clock::time_point StartTime{};
+        bool bReported=false;
    };

    FStandaloneBatchKey BatchForTile(const FVoxelTileKey& Tile)
@@ -133,6 +143,9 @@ namespace
        std::unique_ptr<FStandaloneOwnerWorld> World;
        std::unique_ptr<FStandaloneStreamingGenerator> Streamer;
        std::string DataPath;
+        bool bHasCameraOverride=false;
+        V3 CameraPositionOverride{};
+        float CameraYawOverride=0.0f,CameraPitchOverride=0.0f;
        std::unordered_map<FVoxelTileKey,FResidentTileMesh,FStandaloneTileKeyHash> ResidentTiles;
        std::unordered_map<FStandaloneBatchKey,std::vector<FVoxelTileKey>,FStandaloneBatchKeyHash> BatchMembers;
        std::unordered_map<FStandaloneBatchKey,FStreamingGpuBatch,FStandaloneBatchKeyHash> StreamingBatches;
@@ -143,6 +156,7 @@ namespace
        uint64 GeneratedByLevel[MaxLevels+1]{};
        uint64 SamplesByLevel[MaxLevels+1]{};
        double WorkerSecondsByLevel[MaxLevels+1]{};
+        uint64 ProfileCyclesByLevel[MaxLevels+1][VoxelDensityProfile::BucketCount]{};
        std::vector<double> TileSecondsByLevel[MaxLevels+1];
        uint64 CrossingIndex=0;
        bool bGenerationRunning=false;
@@ -154,11 +168,23 @@ namespace
        bool bAutoCaptureWritten=false;
        bool bFullDetailReported=false;
        bool bInitialClipmapReported=false;
+        bool bProfileAttribution=false;
+        bool bCompareEvaluators=false;
+        bool bParityOnly=false;
+        bool bDensityHash=false;
+        bool bTestOneChunkCrossing=false;
+        bool bCrossingMoveStarted=false;
+        bool bCrossingMoveFinished=false;
+        bool bCrossingClosePosted=false;
+        std::chrono::steady_clock::time_point CrossingMoveStopTime{};
+        bool bCaptureAfterClipmap=false;
+        bool bQuitAfterClipmap=false;
        std::string CapturePath;
        std::chrono::steady_clock::time_point GenerationStart{};
        bool bShowLevel[MaxLevels+1]={true,true,true,true,true,true,true,true,true};
        bool Keys[256]{};
        int32 Seed=0;
+        int32 WorkerCount=6;
        int32 Width=1280,Height=800;
        float Yaw=0.0f,Pitch=0.0f,MoveSpeed=8.0f;
        V3 Position{};
@@ -190,7 +216,9 @@ namespace
            StreamingBatches.clear();BatchMembers.clear();ResidentTiles.clear();
            DirtyBatchSet.clear();DesiredTiles.clear();CrossingWork.clear();
            for(int32 I=0;I<=MaxLevels;++I)
-            {DesiredByLevel[I]=0;GeneratedByLevel[I]=0;SamplesByLevel[I]=0;WorkerSecondsByLevel[I]=0.0;TileSecondsByLevel[I].clear();}
+            {DesiredByLevel[I]=0;GeneratedByLevel[I]=0;SamplesByLevel[I]=0;WorkerSecondsByLevel[I]=0.0;
+             std::fill(std::begin(ProfileCyclesByLevel[I]),std::end(ProfileCyclesByLevel[I]),0);
+             TileSecondsByLevel[I].clear();}
            TotalVertices=0;TotalIndices=0;
        }

@@ -335,9 +363,12 @@ O VS(I i){O o;o.p=mul(float4(i.p,1),VP);o.world=i.p;o.n=normalize(i.n);return o;
 float4 PS(O i):SV_TARGET{
 float3 N=normalize(i.n);float3 V=normalize(Camera-i.world);float3 L=normalize(-LightDir);
 float Diffuse=saturate(dot(N,L));float Head=saturate(dot(N,V));
- float Dist=length(Camera-i.world);float Atten=1.0-saturate(Dist/95.0);Head=Head*Head*Atten;
- float3 Rock=lerp(float3(0.28,0.34,0.38),float3(0.43,0.47,0.48),saturate(N.z*0.5+0.5));
- float3 Lit=Rock*(Ambient+0.88*Diffuse)+float3(0.62,0.76,0.9)*Head*1.25;
+ float Dist=length(Camera-i.world);float Atten=1.0-saturate(Dist/95.0);
+ float3 Rock=lerp(float3(0.30,0.22,0.16),float3(0.48,0.43,0.34),saturate(N.z*0.5+0.5));
+// Keep the camera fill broad and low-amplitude.  The old squared, high-intensity
+// lobe made the 25 cm density-lattice normal variation read as horizontal bands.
+Head=Head*Atten;
+float3 Lit=Rock*(Ambient+0.88*Diffuse)+float3(0.50,0.69,0.82)*Head*0.75;
 float Fog=saturate((Dist-FogStart)/(FogEnd-FogStart));
 Lit=lerp(Lit,float3(0.10,0.14,0.19),Fog);return float4(Lit,1);
 })";
@@ -361,9 +392,9 @@ float4 PS(O i):SV_TARGET{
            return Resize(Width,Height);
        }

-        void PrintParity()
+        size_t PrintParity()
        {
-            if(bParityChecked||!World)return;
+            if(bParityChecked||!World)return 0;
            bParityChecked=true;size_t MismatchCount=0;
            std::map<std::string,size_t> Kinds;
            for(size_t I=0;I<World->DensitySamples.size();++I)
@@ -384,6 +415,7 @@ float4 PS(O i):SV_TARGET{
            for(const auto& [Kind,Count]:Kinds)std::printf("  %-24s %zu\n",Kind.c_str(),Count);
            std::fflush(stdout);
            std::fflush(stdout);
+            return MismatchCount;
        }

        bool ConfigureWorld(std::string& Error)
@@ -428,12 +460,22 @@ float4 PS(O i):SV_TARGET{
            BuildTargetChunk=StandaloneWorldMesher::CameraChunk(CameraVoxels);
            GenerationStart=std::chrono::steady_clock::now();
            bFullDetailReported=false;bInitialClipmapReported=false;bAutoCaptureWritten=false;
+            if(bProfileAttribution)
+            {
+                VoxelDensityProfile::SetMode(VoxelDensityProfile::EMode::Attribution,128);
+                VoxelDensityProfile::CalibrateTimer();
+                VoxelDensityProfile::Reset();
+                std::printf("Density stage profiling enabled (Attribution, sample interval 128).\n");
+            }
            const FStandaloneDesiredTileSet Desired=StandaloneWorldMesher::BuildDesiredSet(*World,CameraVoxels);
            DesiredTiles=Desired.Tiles;
            for(int32 I=0;I<=MaxLevels;++I)DesiredByLevel[I]=0;
            for(const FVoxelTileKey& Tile:DesiredTiles)++DesiredByLevel[Tile.Level];
-            Streamer=std::make_unique<FStandaloneStreamingGenerator>(*World,6);
+            Streamer=std::make_unique<FStandaloneStreamingGenerator>(*World,WorkerCount);
            const FStandaloneStreamingUpdate Update=Streamer->Start(CameraVoxels,{},CrossingIndex);
+            FCrossingWork& InitialWork=CrossingWork[CrossingIndex];
+            InitialWork.SubmittedJobs=Update.NewJobs;InitialWork.PendingJobs=Update.NewJobs;
+            InitialWork.StartTime=GenerationStart;
            bGenerationRunning=true;bRebuildPending=false;
            std::printf("Streaming seed %d%s: %d desired tiles, %d full-detail tiles, %d workers; upload budget %d batches / %.1f ms per frame.\n",
                Seed,bRoundTunnels?" with research round/winding tunnels":"",Update.DesiredTiles,
@@ -495,6 +537,15 @@ float4 PS(O i):SV_TARGET{
            for(const auto& Pair:ResidentTiles)ResidentKeys.push_back(Pair.first);
            ++CrossingIndex;
            const FStandaloneStreamingUpdate Update=Streamer->UpdateDesired(CameraVoxels,ResidentKeys,CrossingIndex);
+            for(const auto& Pair:Update.CancelledByCrossing)
+            {
+                FCrossingWork& Work=CrossingWork[Pair.first];
+                Work.PendingJobs=Work.PendingJobs>=Pair.second?Work.PendingJobs-Pair.second:0;
+                Work.CancelledQueuedJobs+=Pair.second;
+            }
+            FCrossingWork& Work=CrossingWork[CrossingIndex];
+            Work.SubmittedJobs=Update.NewJobs;Work.PendingJobs=Update.NewJobs;
+            Work.StartTime=std::chrono::steady_clock::now();
            BuildTargetChunk=Update.CenterChunk;bGenerationRunning=true;
            std::printf("[ChunkCrossing] id=%llu center=(%d,%d,%d) delta=(%d,%d,%d) desired=%d retained=%d new_jobs=%d dropped_resident=%zu cancelled_queued=%d full_detail=%d\n",
                static_cast<unsigned long long>(CrossingIndex),Update.CenterChunk.X,Update.CenterChunk.Y,Update.CenterChunk.Z,
@@ -502,6 +553,7 @@ float4 PS(O i):SV_TARGET{
                Update.DesiredTiles,Update.RetainedResidentTiles,Update.NewJobs,Leavers.size(),
                Update.CancelledQueuedJobs,Update.FullResTiles);
            std::fflush(stdout);
+            ReportCrossingWork();
        }

        bool RebuildBatch(const FStandaloneBatchKey& Key)
@@ -621,10 +673,44 @@ float4 PS(O i):SV_TARGET{
                std::printf("[TileTiming] LOD=%d generated=%llu target=%d samples=%.2fM worker_sum=%.3fs avg=%.3fms p50=%.3fms p95=%.3fms\n",
                    L,static_cast<unsigned long long>(GeneratedByLevel[L]),DesiredByLevel[L],
                    static_cast<double>(SamplesByLevel[L])/1.0e6,WorkerSecondsByLevel[L],Average*1000.0,P50*1000.0,P95*1000.0);
+                if(bProfileAttribution)
+                {
+                    std::printf("[TileProfile] LOD=%d",L);
+                    for(size_t I=0;I<VoxelDensityProfile::BucketCount;++I)
+                    {
+                        const double Ms=static_cast<double>(ProfileCyclesByLevel[L][I])
+                            *FPlatformTime::GetSecondsPerCycle64()*1000.0;
+                        if(Ms>0.001)std::printf(" %s=%.3fms",
+                            VoxelDensityProfile::BucketName(static_cast<VoxelDensityProfile::EBucket>(I)),Ms);
+                    }
+                    std::printf("\n");
+                }
                Times.clear();
            }
        }

+        void ReportCrossingWork()
+        {
+            for(auto& Pair:CrossingWork)
+            {
+                FCrossingWork& Work=Pair.second;
+                if(Work.bReported||Work.PendingJobs!=0)continue;
+                Work.bReported=true;
+                const double AverageMs=Work.FinishedJobs>0
+                    ?Work.WorkerSeconds*1000.0/static_cast<double>(Work.FinishedJobs):0.0;
+                const double WallSeconds=Work.StartTime.time_since_epoch().count()!=0
+                    ?std::chrono::duration<double>(std::chrono::steady_clock::now()-Work.StartTime).count():0.0;
+                std::printf("[ChunkCrossingWork] id=%llu submitted=%llu finished=%llu generated=%llu cancelled_queued=%llu stale=%llu samples=%.2fM worker_sum=%.3fs wall=%.3fs avg=%.3fms\n",
+                    static_cast<unsigned long long>(Pair.first),
+                    static_cast<unsigned long long>(Work.SubmittedJobs),
+                    static_cast<unsigned long long>(Work.FinishedJobs),
+                    static_cast<unsigned long long>(Work.Tiles),
+                    static_cast<unsigned long long>(Work.CancelledQueuedJobs),
+                    static_cast<unsigned long long>(Work.StaleResults),
+                    static_cast<double>(Work.Samples)/1.0e6,Work.WorkerSeconds,WallSeconds,AverageMs);
+            }
+        }
+
        void PumpGeneration()
        {
            if(bRebuildPending){StartGeneration();return;}
@@ -632,11 +718,16 @@ float4 PS(O i):SV_TARGET{
            std::vector<FStandaloneCpuTile> Ready=Streamer->TakeReady();
            for(FStandaloneCpuTile& Tile:Ready)
            {
-                if(!Streamer->IsDesired(Tile.Key)||ResidentTiles.contains(Tile.Key))continue;
                const int32 Level=Tile.Key.Level;
                FCrossingWork& Crossing=CrossingWork[Tile.RequestedCrossing];
-                ++Crossing.Tiles;Crossing.Samples+=Tile.DensitySamples;Crossing.WorkerSeconds+=Tile.WorkerSeconds;
+                if(Crossing.PendingJobs>0)--Crossing.PendingJobs;
+                ++Crossing.FinishedJobs;Crossing.Samples+=Tile.DensitySamples;Crossing.WorkerSeconds+=Tile.WorkerSeconds;
+                if(!Streamer->IsDesired(Tile.Key)||ResidentTiles.contains(Tile.Key))
+                {++Crossing.StaleResults;continue;}
+                ++Crossing.Tiles;
                ++GeneratedByLevel[Level];SamplesByLevel[Level]+=Tile.DensitySamples;WorkerSecondsByLevel[Level]+=Tile.WorkerSeconds;
+                if(bProfileAttribution)for(size_t I=0;I<Tile.ProfileCycles.size();++I)
+                    ProfileCyclesByLevel[Level][I]+=Tile.ProfileCycles[I];
                TileSecondsByLevel[Level].push_back(Tile.WorkerSeconds);
                FResidentTileMesh Resident;Resident.Vertices=std::move(Tile.Vertices);Resident.Indices=std::move(Tile.Indices);
                Resident.Samples=Tile.DensitySamples;Resident.WorkerSeconds=Tile.WorkerSeconds;
@@ -644,7 +735,20 @@ float4 PS(O i):SV_TARGET{
                const FStandaloneBatchKey Batch=BatchForTile(Tile.Key);
                BatchMembers[Batch].push_back(Tile.Key);MarkBatchDirty(Batch);
            }
+            ReportCrossingWork();
            UploadDirtyBatches();
+            if(bTestOneChunkCrossing&&bCrossingMoveFinished&&!bCrossingClosePosted
+                &&DirtyBatchSet.empty())
+            {
+                const auto Crossing=CrossingWork.find(1);
+                if(Crossing!=CrossingWork.end()&&Crossing->second.bReported)
+                {
+                    bCrossingClosePosted=true;
+                    std::printf("[CrossingTestComplete] crossing work and its bounded uploads drained; closing this test viewer.\n");
+                    std::fflush(stdout);
+                    PostMessageA(Window,WM_CLOSE,0,0);
+                }
+            }
            if(!bFullDetailReported&&IsFullDetailUploaded())
            {
                bFullDetailReported=true;
@@ -661,13 +765,15 @@ float4 PS(O i):SV_TARGET{
                    Elapsed,ResidentTiles.size(),static_cast<unsigned long long>(TotalVertices),
                    static_cast<unsigned long long>(TotalIndices),StreamingBatches.size());
                ReportTileTimings();std::fflush(stdout);
+                if(bQuitAfterClipmap&&CapturePath.empty())PostMessageA(Window,WM_CLOSE,0,0);
            }
            bGenerationRunning=ResidentTiles.size()!=DesiredTiles.size()||!DirtyBatchSet.empty();
        }

        void CaptureRenderedFrame()
        {
-            if(CapturePath.empty()||bAutoCaptureWritten||!IsFullDetailUploaded())return;
+            const bool bReadyToCapture=bCaptureAfterClipmap?bInitialClipmapReported:IsFullDetailUploaded();
+            if(CapturePath.empty()||bAutoCaptureWritten||!bReadyToCapture)return;
            bAutoCaptureWritten=SaveBackBufferPng(CapturePath);
            if(bAutoCaptureWritten)
            {
@@ -698,6 +804,21 @@ float4 PS(O i):SV_TARGET{
        {
            PumpGeneration();
            const auto Now=std::chrono::steady_clock::now();
+            if(bTestOneChunkCrossing&&bFullDetailReported&&!bCrossingMoveStarted&&!bCrossingMoveFinished)
+            {
+                bCrossingMoveStarted=true;
+                CrossingMoveStopTime=Now+std::chrono::milliseconds(350);
+                Keys[VK_SHIFT]=true;Keys['W']=true;
+                std::printf("[CrossingTestMove] moving forward for 350 ms after full detail.\n");
+                std::fflush(stdout);
+            }
+            if(bCrossingMoveStarted&&Now>=CrossingMoveStopTime)
+            {
+                Keys[VK_SHIFT]=false;Keys['W']=false;
+                bCrossingMoveStarted=false;bCrossingMoveFinished=true;
+                std::printf("[CrossingTestMove] movement stopped; waiting for crossing work and uploads.\n");
+                std::fflush(stdout);
+            }
            const float Dt=std::clamp(std::chrono::duration<float>(Now-PreviousFrame).count(),0.0f,0.1f);
            PreviousFrame=Now;FrameMs=Dt*1000.0;
            const V3 F=Forward(),R=Normalize(Cross(F,{0,0,1})),U=Normalize(Cross(R,F));
@@ -767,23 +888,241 @@ float4 PS(O i):SV_TARGET{
        const std::filesystem::path Exe(Module);
        return Exe.parent_path().parent_path()/L"data"/L"owner_world.json";
    }
+
+    int RunTileBenchmark(FStandaloneOwnerWorld& World,V3 CameraPosition,int32 Level,
+                         int32 Count,bool bProfile,bool bCompareEvaluators,bool bDensityHash)
+    {
+        const FVector CameraVoxels(CameraPosition.X*4.0,CameraPosition.Y*4.0,CameraPosition.Z*4.0);
+        const FStandaloneDesiredTileSet Desired=StandaloneWorldMesher::BuildDesiredSet(World,CameraVoxels);
+        std::vector<FVoxelTileKey> Tiles;Tiles.reserve(static_cast<size_t>(std::max(Count,1)));
+        for(const FVoxelTileKey& Tile:Desired.Tiles)
+            if(Tile.Level==Level&&static_cast<int32>(Tiles.size())<Count)Tiles.push_back(Tile);
+        if(Tiles.empty())
+        {std::fprintf(stderr,"No LOD %d tiles are in the current desired set.\n",Level);return 4;}
+
+        VoxelGenLOD::InitializeConsoleSwitches();
+        UVoxelMarchingCubesMesher::InitializeConsoleSwitches();
+        int32 TileCacheWindow=1,TileCacheWindowMinLOD=3,RoomGraphCacheSlots=1;
+        FParse::Value(FCommandLine::Get(),"voxel.TileCacheWindow=",TileCacheWindow);
+        FParse::Value(FCommandLine::Get(),"voxel.TileCacheWindowMinLOD=",TileCacheWindowMinLOD);
+        FParse::Value(FCommandLine::Get(),"voxel.RoomGraphCacheSlots=",RoomGraphCacheSlots);
+        std::printf("[BenchmarkConfig] tile_cache_window=%d fused_min_lod=%d room_graph_cache_slots=%d command='%s'\n",
+            TileCacheWindow,TileCacheWindowMinLOD,RoomGraphCacheSlots,FCommandLine::Get());
+        VoxelDensityProfile::SetMode(bProfile?VoxelDensityProfile::EMode::Attribution
+            :VoxelDensityProfile::EMode::Disabled,128);
+        if(bProfile)VoxelDensityProfile::CalibrateTimer();
+        VoxelDensityProfile::Reset();
+        const VoxelDensityProfile::FSnapshot Before=VoxelDensityProfile::SnapshotCurrentThread();
+
+        UVoxelMarchingCubesMesher Mesher;
+        Mesher.SetGenerator(World.Generator.get());Mesher.bGenerateSkirts=World.bGenerateSkirts;
+        Mesher.SkirtCells=static_cast<float>(World.SkirtCells);Mesher.LODOctaveDrop=World.LODOctaveDrop;
+        const int32 BandMin=World.StrateContentCutMinLevel<9&&Desired.bHasStrate
+            ?Desired.StrateBottomChunkZ*CHUNK_SIZE:MIN_int32;
+        const int32 BandMax=World.StrateContentCutMinLevel<9&&Desired.bHasStrate
+            ?(Desired.StrateTopChunkZ+1)*CHUNK_SIZE:MAX_int32;
+        if(bCompareEvaluators)
+        {
+            const bool bOriginalFused=VoxelGenLOD::IsUseFusedEvaluatorEnabled();
+            uint64 Mismatches=0,ComparedSamples=0,MeshCountMismatches=0;
+            float MaxAbsoluteDifference=0.0f;
+            size_t FirstMismatch=SIZE_MAX;
+            FVoxelTileKey FirstMismatchTile{};
+            VoxelGenLOD::SetUseFusedEvaluatorEnabled(true);
+            for(const FVoxelTileKey& Tile:Tiles)
+            {
+                const int32 Cells=Level<World.FullResClipLevels?CHUNK_SIZE:FMath::Max(1,World.CoarseTileCells);
+                const int32 Step=FMath::Max(1,Tile.ExtentVoxels()/Cells);
+                std::vector<float> FusedDensities,OpStackDensities;
+                int64 FusedSampleCalls=0,OpStackSampleCalls=0;
+                FVoxelMeshData FusedMesh=Mesher.GenerateMesh(Tile.OriginVoxels(),Step,Cells,nullptr,
+                    BandMin,BandMax,&FusedSampleCalls,nullptr,nullptr,&FusedDensities);
+                VoxelGenLOD::SetUseFusedEvaluatorEnabled(false);
+                FVoxelMeshData OpStackMesh=Mesher.GenerateMesh(Tile.OriginVoxels(),Step,Cells,nullptr,
+                    BandMin,BandMax,&OpStackSampleCalls,nullptr,nullptr,&OpStackDensities);
+                VoxelGenLOD::SetUseFusedEvaluatorEnabled(true);
+                if(FusedMesh.Vertices.Num()!=OpStackMesh.Vertices.Num()
+                    || FusedMesh.Triangles.Num()!=OpStackMesh.Triangles.Num())++MeshCountMismatches;
+                const size_t SharedSamples=std::min(FusedDensities.size(),OpStackDensities.size());
+                ComparedSamples+=SharedSamples;
+                for(size_t I=0;I<SharedSamples;++I)
+                {
+                    if(std::bit_cast<uint32>(FusedDensities[I])!=std::bit_cast<uint32>(OpStackDensities[I]))
+                    {
+                        if(FirstMismatch==SIZE_MAX){FirstMismatch=I;FirstMismatchTile=Tile;}
+                        ++Mismatches;
+                        MaxAbsoluteDifference=std::max(MaxAbsoluteDifference,
+                            std::abs(FusedDensities[I]-OpStackDensities[I]));
+                    }
+                }
+                Mismatches+=static_cast<uint64>(std::max(FusedDensities.size(),OpStackDensities.size())-SharedSamples);
+            }
+            std::printf("[EvaluatorCompare] LOD=%d tiles=%zu compared_samples=%llu density_mismatches=%llu first_tile=(%d,%d,%d) first_index=%llu max_abs=%.9g mesh_count_mismatch_tiles=%llu original_fused=%d\n",
+                Level,Tiles.size(),static_cast<unsigned long long>(ComparedSamples),
+                static_cast<unsigned long long>(Mismatches),
+                FirstMismatchTile.Coord.X,FirstMismatchTile.Coord.Y,FirstMismatchTile.Coord.Z,
+                static_cast<unsigned long long>(FirstMismatch),MaxAbsoluteDifference,
+                static_cast<unsigned long long>(MeshCountMismatches),bOriginalFused?1:0);
+            VoxelGenLOD::SetUseFusedEvaluatorEnabled(bOriginalFused);
+            std::fflush(stdout);
+            return Mismatches==0?0:5;
+        }
+        std::vector<double> TileSeconds;TileSeconds.reserve(Tiles.size());
+        uint64 Samples=0,Vertices=0,Indices=0;
+        uint64 DensityHash=14695981039346656037ull;
+        const auto TotalStart=std::chrono::steady_clock::now();
+        for(const FVoxelTileKey& Tile:Tiles)
+        {
+            const auto TileStart=std::chrono::steady_clock::now();
+            const int32 Cells=Level<World.FullResClipLevels?CHUNK_SIZE:FMath::Max(1,World.CoarseTileCells);
+            const int32 Step=FMath::Max(1,Tile.ExtentVoxels()/Cells);
+            int64 TileSamples=0;
+            std::vector<float> DebugDensitySamples;
+            FVoxelMeshData Mesh=Mesher.GenerateMesh(Tile.OriginVoxels(),Step,Cells,nullptr,
+                BandMin,BandMax,&TileSamples,nullptr,nullptr,
+                bDensityHash?&DebugDensitySamples:nullptr);
+            Samples+=static_cast<uint64>(FMath::Max<int64>(0,TileSamples));
+            Vertices+=static_cast<uint64>(Mesh.Vertices.Num());
+            Indices+=static_cast<uint64>(Mesh.Triangles.Num());
+            if(bDensityHash)
+            {
+                for(const float Density:DebugDensitySamples)
+                {
+                    DensityHash^=std::bit_cast<uint32>(Density);
+                    DensityHash*=1099511628211ull;
+                }
+            }
+            TileSeconds.push_back(std::chrono::duration<double>(
+                std::chrono::steady_clock::now()-TileStart).count());
+        }
+        const double TotalSeconds=std::chrono::duration<double>(
+            std::chrono::steady_clock::now()-TotalStart).count();
+        const VoxelDensityProfile::FSnapshot After=VoxelDensityProfile::SnapshotCurrentThread();
+        std::sort(TileSeconds.begin(),TileSeconds.end());
+        const double Average=TotalSeconds/static_cast<double>(Tiles.size());
+        std::printf("[TileBenchmark] LOD=%d generated=%zu samples=%.2fM worker_sum=%.3fs avg=%.3fms p50=%.3fms p95=%.3fms vertices=%llu indices=%llu profiling=%s\n",
+            Level,Tiles.size(),static_cast<double>(Samples)/1.0e6,TotalSeconds,Average*1000.0,
+            TileSeconds[(TileSeconds.size()-1)*50/100]*1000.0,
+            TileSeconds[(TileSeconds.size()-1)*95/100]*1000.0,
+            static_cast<unsigned long long>(Vertices),static_cast<unsigned long long>(Indices),
+            bProfile?"Attribution":"off");
+        if(bDensityHash)
+        {
+            std::printf("[TileDensityHash] LOD=%d tiles=%zu hash=0x%016llx\n",Level,Tiles.size(),
+                static_cast<unsigned long long>(DensityHash));
+        }
+        if(bProfile)
+        {
+            std::printf("[TileProfile] LOD=%d",Level);
+            for(int32 I=0;I<VoxelDensityProfile::BucketCount;++I)
+            {
+                const uint64 CycleDelta=After.Cycles[I]>=Before.Cycles[I]
+                    ?After.Cycles[I]-Before.Cycles[I]:0;
+                const double Ms=static_cast<double>(CycleDelta)
+                    *FPlatformTime::GetSecondsPerCycle64()*1000.0;
+                if(Ms>0.001)std::printf(" %s=%.3fms",VoxelDensityProfile::BucketName(
+                    static_cast<VoxelDensityProfile::EBucket>(I)),Ms);
+            }
+            std::printf("\n");
+            const auto CounterDelta=[&](VoxelDensityProfile::ECounter Counter)
+            {
+                const size_t I=static_cast<size_t>(Counter);
+                return After.Counters[I]>=Before.Counters[I]
+                    ?After.Counters[I]-Before.Counters[I]:0;
+            };
+            std::printf("[TileCounters] sdf_builds=%llu graph_tile_builds=%llu graph_core_builds=%llu op_block_builds=%llu\n",
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfCacheBuild)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::RoomGraphBuildGeneratorTile)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::RoomGraphBuildGeneratorTunnelCore)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::OpBlockBuilds)));
+            std::printf("[SdfRebuildCauses] strate=%llu seed=%llu fingerprint=%llu layout=%llu manager=%llu terrain_ops=%llu fused_policy=%llu tile_policy=%llu tile_identity=%llu bounds_x=%llu bounds_y=%llu tile_window_accepted=%llu tile_window_rejected=%llu\n",
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildStrate)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildSeed)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildFingerprint)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildLayout)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildManagerLifetime)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildTerrainOps)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildFusedPolicy)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildTilePolicy)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildTileIdentity)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildBoundsX)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfRebuildBoundsY)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfTileWindowAccepted)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::SdfTileWindowRejected)));
+            std::printf("[TilePolicyKeyChanges] origin=%llu step=%llu cells=%llu fingerprint=%llu layout=%llu manager=%llu terrain_ops=%llu fused_path=%llu context=%llu\n",
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyOrigin)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyStep)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyCells)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyFingerprint)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyLayout)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyManager)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyTerrainOps)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyFusedPath)),
+                static_cast<unsigned long long>(CounterDelta(VoxelDensityProfile::ECounter::TilePolicyKeyContext)));
+        }
+        std::fflush(stdout);
+        return 0;
+    }
 }

 int main(int Argc,char** Argv)
 {
    FApp App;App.DataPath=DefaultDataPath().string();
+    std::string JoinedCommandLine;
+    for(int32 I=1;I<Argc;++I)
+    {if(!JoinedCommandLine.empty())JoinedCommandLine.push_back(' ');JoinedCommandLine+=Argv[I];}
+    FCommandLine::Set(JoinedCommandLine);
+    int32 BenchmarkLevel=-1,BenchmarkCount=64;
    for(int32 I=1;I<Argc;++I)
    {
        const std::string_view Arg=Argv[I];
        if(Arg.starts_with("--capture="))App.CapturePath=std::string(Arg.substr(10));
+        else if(Arg=="--profile-attribution")App.bProfileAttribution=true;
+        else if(Arg=="--compare-evaluators")App.bCompareEvaluators=true;
+        else if(Arg=="--parity-only")App.bParityOnly=true;
+        else if(Arg=="--density-hash")App.bDensityHash=true;
+        else if(Arg=="--round-tunnels")App.bRoundTunnels=true;
+        else if(Arg=="--test-one-chunk-crossing")App.bTestOneChunkCrossing=true;
+        else if(Arg=="--capture-after-clipmap")App.bCaptureAfterClipmap=true;
+        else if(Arg=="--quit-after-clipmap")App.bQuitAfterClipmap=true;
+        else if(Arg.starts_with("--workers="))
+        {
+            const int32 Parsed=std::atoi(std::string(Arg.substr(10)).c_str());
+            App.WorkerCount=std::clamp(Parsed,1,32);
+        }
+        else if(Arg.starts_with("--benchmark-level="))
+            BenchmarkLevel=std::atoi(std::string(Arg.substr(18)).c_str());
+        else if(Arg.starts_with("--benchmark-count="))
+            BenchmarkCount=std::max(1,std::atoi(std::string(Arg.substr(18)).c_str()));
+        else if(Arg.starts_with("--camera="))
+        {
+            const std::string Values(Arg.substr(9));
+            float X=0,Y=0,Z=0,YawDegrees=0,PitchDegrees=0;
+            if(std::sscanf(Values.c_str(),"%f,%f,%f,%f,%f",&X,&Y,&Z,&YawDegrees,&PitchDegrees)==5)
+            {
+                App.bHasCameraOverride=true;App.CameraPositionOverride={X,Y,Z};
+                App.CameraYawOverride=YawDegrees*0.017453292519943295f;
+                App.CameraPitchOverride=PitchDegrees*0.017453292519943295f;
+            }
+            else {std::fprintf(stderr,"Invalid --camera value: %s\n",Values.c_str());return 2;}
+        }
        else if(!Arg.empty()&&Arg.front()!='-')App.DataPath=std::string(Arg);
    }
    std::string Error;
    App.World=std::make_unique<FStandaloneOwnerWorld>();
    if(!StandaloneOwnerLoader::Load(App.DataPath,*App.World,Error)){std::fprintf(stderr,"Owner data load failed: %s\n",Error.c_str());return 2;}
    App.Seed=App.World->Seed;
+    if(App.bParityOnly)
+    {
+        VoxelGenLOD::InitializeConsoleSwitches();
+        return App.PrintParity()==0?0:1;
+    }
    App.Position={static_cast<float>(App.World->ViewerStartVoxels.X*0.25),
        static_cast<float>(App.World->ViewerStartVoxels.Y*0.25),static_cast<float>(App.World->ViewerStartVoxels.Z*0.25)};
+    if(App.bHasCameraOverride){App.Position=App.CameraPositionOverride;App.Yaw=App.CameraYawOverride;App.Pitch=App.CameraPitchOverride;}
+    if(BenchmarkLevel>=0)
+        return RunTileBenchmark(*App.World,App.Position,BenchmarkLevel,BenchmarkCount,
+            App.bProfileAttribution,App.bCompareEvaluators,App.bDensityHash);
    App.World.reset();

    WNDCLASSA Class{};Class.lpfnWndProc=WindowProc;Class.hInstance=GetModuleHandleA(nullptr);Class.lpszClassName="VoxelForgeStandaloneWorld";
@@ -3,6 +3,7 @@
 #include "WorldMesher.h"

 #include <condition_variable>
+#include <array>
 #include <unordered_set>

 struct FStandaloneTileKeyHash
@@ -30,6 +31,7 @@ struct FStandaloneCpuTile
    std::vector<uint32> Indices;
    uint64 DensitySamples=0;
    double WorkerSeconds=0.0;
+    std::array<uint64,VoxelDensityProfile::BucketCount> ProfileCycles{};
 };

 struct FStandaloneStreamingUpdate
@@ -40,6 +42,7 @@ struct FStandaloneStreamingUpdate
    int32 NewJobs=0;
    int32 CancelledQueuedJobs=0;
    int32 FullResTiles=0;
+    std::vector<std::pair<uint64,uint64>> CancelledByCrossing;
 };

 struct FStandaloneTileJob
@@ -48,6 +51,20 @@ struct FStandaloneTileJob
    int32 BandZMin=MIN_int32;
    int32 BandZMax=MAX_int32;
    uint64 RequestedCrossing=0;
+    double PriorityDistanceSquared=0.0;
+};
+
+struct FStandaloneTileJobHeapOrder
+{
+    bool operator()(const FStandaloneTileJob& A,const FStandaloneTileJob& B)const
+    {
+        if(A.PriorityDistanceSquared!=B.PriorityDistanceSquared)
+            return A.PriorityDistanceSquared>B.PriorityDistanceSquared;
+        if(A.Key.Level!=B.Key.Level)return A.Key.Level>B.Key.Level;
+        if(A.Key.Coord.Z!=B.Key.Coord.Z)return A.Key.Coord.Z>B.Key.Coord.Z;
+        if(A.Key.Coord.Y!=B.Key.Coord.Y)return A.Key.Coord.Y>B.Key.Coord.Y;
+        return A.Key.Coord.X>B.Key.Coord.X;
+    }
 };

 namespace StandaloneWorldMesher
@@ -99,6 +116,7 @@ public:
    explicit FStandaloneStreamingGenerator(FStandaloneOwnerWorld& InWorld,int32 InWorkerCount=6)
        :World(&InWorld),WorkerCount(FMath::Clamp(InWorkerCount,1,32))
    {
+        VoxelGenLOD::InitializeConsoleSwitches();
        UVoxelMarchingCubesMesher::InitializeConsoleSwitches();
    }
    FStandaloneStreamingGenerator(const FStandaloneStreamingGenerator&)=delete;
@@ -153,12 +171,20 @@ public:
        Desired.clear();Desired.reserve(NewDesired.Tiles.size()*2+1);
        for(const FVoxelTileKey& Key:NewDesired.Tiles)Desired.insert(Key);
        FullResTileCount=Stats.FullResTiles;
+        std::unordered_map<uint64,uint64> CancelledByCrossing;
        auto QueueEnd=std::remove_if(Queue.begin(),Queue.end(),[&](const FStandaloneTileJob& Job)
        {
            if(Desired.contains(Job.Key))return false;
-            Outstanding.erase(Job.Key);++Stats.CancelledQueuedJobs;return true;
+            Outstanding.erase(Job.Key);++Stats.CancelledQueuedJobs;
+            ++CancelledByCrossing[Job.RequestedCrossing];return true;
        });
        Queue.erase(QueueEnd,Queue.end());
+        Stats.CancelledByCrossing.reserve(CancelledByCrossing.size());
+        for(const auto& Pair:CancelledByCrossing)Stats.CancelledByCrossing.push_back(Pair);
+
+        for(FStandaloneTileJob& Job:Queue)
+            Job.PriorityDistanceSquared=StandaloneWorldMesher::TileDistanceSquaredChunks(
+                Job.Key,NewDesired.CenterChunk);

        const int32 BandMin=World->StrateContentCutMinLevel<9&&NewDesired.bHasStrate
            ? NewDesired.StrateBottomChunkZ*CHUNK_SIZE:MIN_int32;
@@ -167,16 +193,12 @@ public:
        for(const FVoxelTileKey& Key:NewDesired.Tiles)
        {
            if(Resident.contains(Key)||Outstanding.contains(Key))continue;
-            Queue.push_back({Key,BandMin,BandMax,CrossingIndex});
+            Queue.push_back({Key,BandMin,BandMax,CrossingIndex,
+                StandaloneWorldMesher::TileDistanceSquaredChunks(Key,NewDesired.CenterChunk)});
            Outstanding.insert(Key);++Stats.NewJobs;
        }
        CurrentDesired=std::move(NewDesired);
-        std::sort(Queue.begin(),Queue.end(),[&](const FStandaloneTileJob& A,const FStandaloneTileJob& B)
-        {
-            const double DA=StandaloneWorldMesher::TileDistanceSquaredChunks(A.Key,CurrentDesired.CenterChunk);
-            const double DB=StandaloneWorldMesher::TileDistanceSquaredChunks(B.Key,CurrentDesired.CenterChunk);
-            return DA==DB?A.Key.Level<B.Key.Level:DA<DB;
-        });
+        std::make_heap(Queue.begin(),Queue.end(),FStandaloneTileJobHeapOrder{});
        Wake.notify_all();
        return Stats;
    }
@@ -215,8 +237,21 @@ private:
        const int32 Cells=Level<World->FullResClipLevels?CHUNK_SIZE:FMath::Max(1,World->CoarseTileCells);
        const int32 Step=FMath::Max(1,Extent/Cells);
        int64 SampleCount=0;
+        const bool bProfile=VoxelDensityProfile::GetMode()!=VoxelDensityProfile::EMode::Disabled;
+        const VoxelDensityProfile::FSnapshot Before=bProfile
+            ? VoxelDensityProfile::SnapshotCurrentThread():VoxelDensityProfile::FSnapshot{};
        FVoxelMeshData Mesh=Mesher.GenerateMesh(Job.Key.OriginVoxels(),Step,Cells,nullptr,
            Job.BandZMin,Job.BandZMax,&SampleCount);
+        if(bProfile)
+        {
+            const VoxelDensityProfile::FSnapshot After=VoxelDensityProfile::SnapshotCurrentThread();
+            for(size_t I=0;I<Out.ProfileCycles.size();++I)
+            {
+                const size_t Bucket=I;
+                Out.ProfileCycles[I]=After.Cycles[Bucket]>=Before.Cycles[Bucket]
+                    ?After.Cycles[Bucket]-Before.Cycles[Bucket]:0;
+            }
+        }
        Out.DensitySamples=static_cast<uint64>(FMath::Max<int64>(0,SampleCount));
        Out.Vertices.reserve(static_cast<size_t>(Mesh.Vertices.Num()));
        for(int32 I=0;I<Mesh.Vertices.Num();++I)
@@ -241,7 +276,8 @@ private:
                std::unique_lock<std::mutex> Lock(Mutex);
                Wake.wait(Lock,[&](){return bStop||!Queue.empty();});
                if(bStop)return;
-                Job=Queue.front();Queue.erase(Queue.begin());
+                std::pop_heap(Queue.begin(),Queue.end(),FStandaloneTileJobHeapOrder{});
+                Job=Queue.back();Queue.pop_back();
            }
            FStandaloneCpuTile Tile=Generate(Job);
            {
@@ -7750,7 +7750,7 @@ void VoxelCaveMorphology::BuildChunkCache(
    GraphLowLink.Init(INDEX_NONE, NumRooms);
    GraphBridges.Init(0u, GraphEdges.Num());
    int32 GraphDiscoveryTime = 0;
-    TFunction<void(int32, int32)> VisitGraph = [&](int32 Node, int32 ParentEdge)
+    const auto VisitGraph = [&](auto&& VisitGraphSelf, int32 Node, int32 ParentEdge) -> void
    {
        GraphDiscovery[Node] = GraphDiscoveryTime;
        GraphLowLink[Node] = GraphDiscoveryTime;
@@ -7765,7 +7765,7 @@ void VoxelCaveMorphology::BuildChunkCache(
            const int32 Other = Edge.RoomA == Node ? Edge.RoomB : Edge.RoomA;
            if (GraphDiscovery[Other] == INDEX_NONE)
            {
-                VisitGraph(Other, EdgeIndex);
+                VisitGraphSelf(VisitGraphSelf, Other, EdgeIndex);
                GraphLowLink[Node] = FMath::Min(
                    GraphLowLink[Node], GraphLowLink[Other]);
                if (GraphLowLink[Other] > GraphDiscovery[Node])
@@ -7784,7 +7784,7 @@ void VoxelCaveMorphology::BuildChunkCache(
    {
        if (GraphDiscovery[Node] == INDEX_NONE)
        {
-            VisitGraph(Node, INDEX_NONE);
+            VisitGraph(VisitGraph, Node, INDEX_NONE);
        }
    }
    for (int32 EdgeIndex = 0; EdgeIndex < GraphEdges.Num(); ++EdgeIndex)
@@ -7803,7 +7803,7 @@ void VoxelCaveMorphology::BuildChunkCache(
    TArray<uint8, TInlineAllocator<128>> GraphForestEdges;
    GraphForestVisited.Init(0u, NumRooms);
    GraphForestEdges.Init(0u, GraphEdges.Num());
-    TFunction<void(int32)> VisitWalkForest = [&](int32 Node)
+    const auto VisitWalkForest = [&](auto&& VisitWalkForestSelf, int32 Node) -> void
    {
        GraphForestVisited[Node] = 1u;
        for (const int32 EdgeIndex : GraphAdjacency[Node])
@@ -7815,14 +7815,14 @@ void VoxelCaveMorphology::BuildChunkCache(
                continue;
            }
            GraphForestEdges[EdgeIndex] = 1u;
-            VisitWalkForest(Other);
+            VisitWalkForestSelf(VisitWalkForestSelf, Other);
        }
    };
    for (int32 Node = 0; Node < NumRooms; ++Node)
    {
        if (GraphForestVisited[Node] == 0u)
        {
-            VisitWalkForest(Node);
+            VisitWalkForest(VisitWalkForest, Node);
        }
    }
    for (int32 EdgeIndex = 0; EdgeIndex < GraphEdges.Num(); ++EdgeIndex)
@@ -704,6 +704,28 @@ namespace VoxelDensityProfile
        case ECounter::MesherOutputAllocatedBytes: return TEXT("MesherOutputAllocatedBytes");
        case ECounter::OpBlockScratchBytes: return TEXT("OpBlockScratchBytes");
        case ECounter::OpBlockCopyBytes: return TEXT("OpBlockCopyBytes");
+        case ECounter::SdfRebuildStrate: return TEXT("SdfRebuildStrate");
+        case ECounter::SdfRebuildSeed: return TEXT("SdfRebuildSeed");
+        case ECounter::SdfRebuildFingerprint: return TEXT("SdfRebuildFingerprint");
+        case ECounter::SdfRebuildLayout: return TEXT("SdfRebuildLayout");
+        case ECounter::SdfRebuildManagerLifetime: return TEXT("SdfRebuildManagerLifetime");
+        case ECounter::SdfRebuildTerrainOps: return TEXT("SdfRebuildTerrainOps");
+        case ECounter::SdfRebuildFusedPolicy: return TEXT("SdfRebuildFusedPolicy");
+        case ECounter::SdfRebuildTilePolicy: return TEXT("SdfRebuildTilePolicy");
+        case ECounter::SdfRebuildTileIdentity: return TEXT("SdfRebuildTileIdentity");
+        case ECounter::SdfRebuildBoundsX: return TEXT("SdfRebuildBoundsX");
+        case ECounter::SdfRebuildBoundsY: return TEXT("SdfRebuildBoundsY");
+        case ECounter::SdfTileWindowAccepted: return TEXT("SdfTileWindowAccepted");
+        case ECounter::SdfTileWindowRejected: return TEXT("SdfTileWindowRejected");
+        case ECounter::TilePolicyKeyOrigin: return TEXT("TilePolicyKeyOrigin");
+        case ECounter::TilePolicyKeyStep: return TEXT("TilePolicyKeyStep");
+        case ECounter::TilePolicyKeyCells: return TEXT("TilePolicyKeyCells");
+        case ECounter::TilePolicyKeyFingerprint: return TEXT("TilePolicyKeyFingerprint");
+        case ECounter::TilePolicyKeyLayout: return TEXT("TilePolicyKeyLayout");
+        case ECounter::TilePolicyKeyManager: return TEXT("TilePolicyKeyManager");
+        case ECounter::TilePolicyKeyTerrainOps: return TEXT("TilePolicyKeyTerrainOps");
+        case ECounter::TilePolicyKeyFusedPath: return TEXT("TilePolicyKeyFusedPath");
+        case ECounter::TilePolicyKeyContext: return TEXT("TilePolicyKeyContext");
        case ECounter::Count:              break;
        }
        return TEXT("Unknown");
@@ -32,6 +32,7 @@
 #endif

 #include <atomic>
+#include <array>

 namespace VoxelMath
 {
@@ -139,6 +140,24 @@ namespace
        TEXT("voxel.TileCacheWindowOpStackMinLOD"),
        GVoxelForgeTileCacheWindowOpStackMinLOD,
        TEXT("Minimum SampleStep LOD for the op-stack tile-sized room cache window."));
+    int32 GVoxelForgeRoomGraphCacheSlots = 1;
+    FAutoConsoleVariableRef CVarVoxelForgeRoomGraphCacheSlots(
+        TEXT("voxel.RoomGraphCacheSlots"),
+        GVoxelForgeRoomGraphCacheSlots,
+        TEXT("Per-worker exact room-graph windows to retain during density sampling (1-4)."));
+    bool GVoxelForgeRoomGraphCacheSlotsParsed = false;
+
+    void VF_ParseRoomGraphCacheSlots()
+    {
+        if (GVoxelForgeRoomGraphCacheSlotsParsed) return;
+        GVoxelForgeRoomGraphCacheSlotsParsed = true;
+        int32 CommandLineValue = GVoxelForgeRoomGraphCacheSlots;
+        if (FParse::Value(FCommandLine::Get(), TEXT("voxel.RoomGraphCacheSlots="), CommandLineValue))
+        {
+            GVoxelForgeRoomGraphCacheSlots = CommandLineValue;
+        }
+        GVoxelForgeRoomGraphCacheSlots = FMath::Clamp(GVoxelForgeRoomGraphCacheSlots, 1, 4);
+    }

    int32 GVoxelForgeSpatialIndex = 1;
    FAutoConsoleVariableRef CVarVoxelForgeSpatialIndex(
@@ -879,10 +898,23 @@ void VoxelGenLOD::InitializeConsoleSwitches()
    VF_ParseFusedEvaluatorSwitch();
    VF_ParseWormBlockSkipSwitch();
    VF_ParseTileCacheWindowSwitch();
+    VF_ParseRoomGraphCacheSlots();
    VF_ParseSpatialIndexSwitch();
    VF_ParseChunkRoutingCacheSwitch();
 }

+void VoxelGenLOD::SetUseFusedEvaluatorEnabled(bool bEnabled)
+{
+    VF_ParseFusedEvaluatorSwitch();
+    GVoxelForgeUseFusedEvaluator = bEnabled ? 1 : 0;
+}
+
+bool VoxelGenLOD::IsUseFusedEvaluatorEnabled()
+{
+    VF_ParseFusedEvaluatorSwitch();
+    return GVoxelForgeUseFusedEvaluator != 0;
+}
+
 bool VoxelGenLOD::IsTileCacheWindowEnabled(bool bFusedPath)
 {
    if (GVoxelForgeTileCacheWindow == 0)
@@ -4171,38 +4203,31 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo
    //
    // Declared before the RoomDensity if-block so SDFCache and NearestRoomIdx
    // are also visible to the terrain ops block further below.
-    thread_local FChunkSDFCache SDFCache;
-    // Cache validity is tracked by the SEARCH BOX the cache was built for, NOT by chunk
-    // equality. Gradient-normal sampling queries at WorldX±1 (and warp displacement) can
-    // step a voxel outside the chunk; with chunk-equality keying that flipped the key and
-    // rebuilt the (now expensive) cache every boundary cell. Since the stored rooms cover
-    // the search box + MaxInfluence, any query INSIDE the box is correct — so we only
-    // rebuild when the query actually leaves the box. Result: one build per chunk, no thrash.
-    thread_local float CachedSMinX = 1.0f, CachedSMaxX = -1.0f;  // start invalid (min > max)
-    thread_local float CachedSMinY = 0.0f, CachedSMaxY = 0.0f;
-    thread_local int32 CachedStrate = INT32_MIN;
-    thread_local uint32 CachedSeed = 0;
-    // ⚠️ AUDIT §C2 — `CachedFingerprint` et `CachedLayout` ci-dessous sont OBLIGATOIRES dans cette clé.
-    // La clé ci-dessus décrit la GÉOMÉTRIE de la fenêtre (boîte, strate, seed) et rien de ce qui
-    // détermine les PARAMS avec lesquels les salles ont été cuites. Comme `GetGenerationParams`
-    // blende à l'intérieur d'une strate (`Alpha` = f(chunk Z), et f(chunk XY) aussi en
-    // `Interleaved`), deux chunks voisins produisent la MÊME clé avec des params DIFFÉRENTS, et le
-    // deuxième se sert des salles du premier. Non déterministe entre pairs, parce que l'ordre des
-    // workers décide lequel est « le premier » — exactement ce que §2.6.1 interdit.
-    //
-    // Pourquoi ça ne casse PAS l'invariant de perf de §8.10 : la clé reste une BOÎTE, donc les
-    // sondes de gradient à `WorldX ± 1` ne font toujours pas tourner le cache. Ce qui le fait
-    // tourner en plus, c'est un changement RÉEL de params — une fois par chunk dans une bande de
-    // transition, ce qui est exactement le nombre de reconstructions nécessaire.
-    thread_local uint32 CachedFingerprint = 0xFFFFFFFFu;
-    thread_local uint32 CachedLayout      = 0xFFFFFFFFu;
-    thread_local bool CachedUsesFusedCacheWindow = false;
-    thread_local bool CachedUsesTileCacheWindow = false;
-    thread_local FIntVector CachedTileOrigin = FIntVector::ZeroValue;
-    thread_local int32 CachedTileStep = 0;
-    thread_local int32 CachedTileCells = 0;
-    thread_local uint64 CachedManagerLifetimeId = 0;
-    thread_local const TArray<FStrateTerrainOpEntry>* CachedTerrainOps = nullptr;
+    // Keep a small bounded set of exact room-graph windows. Tall 32-chunk tiles need two slots
+    // for their two strata; other tile extents start with one. Every slot uses the full cache
+    // identity/bounds.
+    struct FRoomGraphCacheSlot
+    {
+        FChunkSDFCache Cache;
+        float SMinX = 1.0f, SMaxX = -1.0f;
+        float SMinY = 0.0f, SMaxY = 0.0f;
+        int32 Strate = INT32_MIN;
+        uint32 Seed = 0;
+        uint32 Fingerprint = 0xFFFFFFFFu;
+        uint32 Layout = 0xFFFFFFFFu;
+        bool bUsesFusedCacheWindow = false;
+        bool bUsesTileCacheWindow = false;
+        FIntVector TileOrigin = FIntVector::ZeroValue;
+        int32 TileStep = 0;
+        int32 TileCells = 0;
+        uint64 ManagerLifetimeId = 0;
+        const TArray<FStrateTerrainOpEntry>* TerrainOps = nullptr;
+        uint64 LastUse = 0;
+        bool bValid = false;
+    };
+    static thread_local std::array<FRoomGraphCacheSlot, 4> RoomGraphCacheSlots{};
+    static thread_local uint64 RoomGraphCacheUseSerial = 0;
+    VF_ParseRoomGraphCacheSlots();
    // The envelope proof is constant for a cache-policy key.  Do not run the complete room/edge
    // reach calculation for every density sample: that would turn a build-time guard into a
    // per-sample cost on the LOD0 floor.
@@ -4267,6 +4292,32 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo
        || CachedTilePolicyTerrainOps != TerrainOps;
    if (bTilePolicyKeyChanged)
    {
+        if (VoxelDensityProfile::AreCountersEnabled())
+        {
+            const auto CountPolicyKeyChange = [](bool bCondition, VoxelDensityProfile::ECounter Counter)
+            {
+                if (bCondition) VoxelDensityProfile::AddCounter(Counter);
+            };
+            CountPolicyKeyChange(CachedTilePolicyOrigin != RequestedTileOrigin,
+                VoxelDensityProfile::ECounter::TilePolicyKeyOrigin);
+            CountPolicyKeyChange(CachedTilePolicySampleStep != RequestedTileStep
+                    || CachedTilePolicyStep != RequestedTileStep,
+                VoxelDensityProfile::ECounter::TilePolicyKeyStep);
+            CountPolicyKeyChange(CachedTilePolicyCells != RequestedTileCells,
+                VoxelDensityProfile::ECounter::TilePolicyKeyCells);
+            CountPolicyKeyChange(CachedTilePolicyFingerprint != ParamsFingerprint,
+                VoxelDensityProfile::ECounter::TilePolicyKeyFingerprint);
+            CountPolicyKeyChange(CachedTilePolicyLayout != LayoutVersion,
+                VoxelDensityProfile::ECounter::TilePolicyKeyLayout);
+            CountPolicyKeyChange(CachedTilePolicyManagerLifetimeId != CurrentManagerLifetimeId,
+                VoxelDensityProfile::ECounter::TilePolicyKeyManager);
+            CountPolicyKeyChange(CachedTilePolicyTerrainOps != TerrainOps,
+                VoxelDensityProfile::ECounter::TilePolicyKeyTerrainOps);
+            CountPolicyKeyChange(CachedTilePolicyFusedPath != bUseFusedCacheWindow,
+                VoxelDensityProfile::ECounter::TilePolicyKeyFusedPath);
+            CountPolicyKeyChange(CachedTilePolicyHasContext != (bTileWindowEligible && bHasTileWindowContext),
+                VoxelDensityProfile::ECounter::TilePolicyKeyContext);
+        }
        CachedTilePolicyValid = true;
        CachedTilePolicyHasContext = bTileWindowEligible && bHasTileWindowContext;
        CachedTilePolicyFusedPath = bUseFusedCacheWindow;
@@ -4280,6 +4331,12 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo
        CachedTilePolicyTerrainOps = TerrainOps;
        CachedTilePolicyResult = CachedTilePolicyHasContext
            && VoxelCaveMorphology::IsRoomGraphWindowInvariant(Params, TerrainOps);
+        if (VoxelDensityProfile::AreCountersEnabled() && CachedTilePolicyHasContext)
+        {
+            VoxelDensityProfile::AddCounter(CachedTilePolicyResult
+                ? VoxelDensityProfile::ECounter::SdfTileWindowAccepted
+                : VoxelDensityProfile::ECounter::SdfTileWindowRejected);
+        }
    }
    const bool bUseTileCacheWindow = CachedTilePolicyResult;

@@ -4290,36 +4347,137 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo

    float CaveSDF = FLT_MAX;

+    // Get strate index for unique caves per strate. The index changes only at chunk-Z
+    // boundaries and is also part of the worker-local graph-cache identity.
+    int32 StrateIdx = 0;
+    if (StrateManager)
+    {
+        thread_local int32  SI_ChunkZ  = INT32_MAX;
+        thread_local uint32 SI_Version = 0xFFFFFFFFu;
+        thread_local int32  SI_Index   = 0;
+        const int32 QZ = FMath::FloorToInt(WorldZ / (float)CHUNK_SIZE);
+        const uint32 LV = StrateManager->GetLayoutVersion();
+        if (QZ != SI_ChunkZ || LV != SI_Version)
+        {
+            SI_ChunkZ  = QZ;
+            SI_Version = LV;
+            SI_Index = StrateManager->GetStrateIndex(
+                ((float)QZ + 0.5f) * CHUNK_SIZE * VOXEL_SIZE);
+        }
+        StrateIdx = SI_Index;
+    }
+
+        const float ChunkMinX = CacheRegionX * static_cast<float>(CacheRegionSize);
+        const float ChunkMinY = CacheRegionY * static_cast<float>(CacheRegionSize);
+        const float ChunkMaxX = ChunkMinX + static_cast<float>(CacheRegionSize);
+        const float ChunkMaxY = ChunkMinY + static_cast<float>(CacheRegionSize);
+        const float WarpMargin = FMath::Abs(Params.CaveWarpStrength)
+            * VOXEL_NOISE_SCALE * 1.5f;
+        float DesiredSMinX = ChunkMinX;
+        float DesiredSMinY = ChunkMinY;
+        float DesiredSMaxX = ChunkMaxX;
+        float DesiredSMaxY = ChunkMaxY;
+        if (bUseTileCacheWindow)
+        {
+            const float TileMinX = static_cast<float>(RequestedTileOrigin.X);
+            const float TileMinY = static_cast<float>(RequestedTileOrigin.Y);
+            const float TileExtent = static_cast<float>(
+                static_cast<int64>(RequestedTileStep)
+                * static_cast<int64>(RequestedTileCells));
+            const float TileMaxX = TileMinX + TileExtent;
+            const float TileMaxY = TileMinY + TileExtent;
+            const float TileHalo = static_cast<float>(RequestedTileStep) + 2.0f;
+            const float Expansion = WarpMargin + TileHalo;
+            DesiredSMinX = TileMinX - Expansion;
+            DesiredSMinY = TileMinY - Expansion;
+            DesiredSMaxX = TileMaxX + Expansion;
+            DesiredSMaxY = TileMaxY + Expansion;
+        }
+        else
+        {
+            const float Expansion = WarpMargin + 2.0f;
+            DesiredSMinX = ChunkMinX - Expansion;
+            DesiredSMinY = ChunkMinY - Expansion;
+            DesiredSMaxX = ChunkMaxX + Expansion;
+            DesiredSMaxY = ChunkMaxY + Expansion;
+        }
+
+        const int64 RequestedTileExtent = static_cast<int64>(RequestedTileStep)
+            * static_cast<int64>(RequestedTileCells);
+        const bool bNeedsTwoStrataWindows = bHasTileWindowContext
+            && RequestedTileExtent == static_cast<int64>(CHUNK_SIZE) * 32;
+        const int32 ActiveCacheSlots = FMath::Max(
+            FMath::Clamp(GVoxelForgeRoomGraphCacheSlots, 1, 4),
+            bNeedsTwoStrataWindows ? 2 : 1);
+        FRoomGraphCacheSlot* CacheSlot = nullptr;
+        FRoomGraphCacheSlot* EvictionSlot = &RoomGraphCacheSlots[0];
+        for (int32 SlotIndex = 0; SlotIndex < ActiveCacheSlots; ++SlotIndex)
+        {
+            FRoomGraphCacheSlot& Candidate = RoomGraphCacheSlots[SlotIndex];
+            if (!Candidate.bValid)
+            {
+                EvictionSlot = &Candidate;
+                break;
+            }
+            if (Candidate.LastUse < EvictionSlot->LastUse)
+            {
+                EvictionSlot = &Candidate;
+            }
+            const bool bSameKey = Candidate.Strate == StrateIdx
+                && Candidate.Seed == static_cast<uint32>(Seed)
+                && Candidate.Fingerprint == ParamsFingerprint
+                && Candidate.Layout == LayoutVersion
+                && Candidate.ManagerLifetimeId == CurrentManagerLifetimeId
+                && Candidate.TerrainOps == TerrainOps
+                && Candidate.bUsesFusedCacheWindow == bUseFusedCacheWindow
+                && Candidate.bUsesTileCacheWindow == bUseTileCacheWindow
+                && (!bUseTileCacheWindow
+                    || (Candidate.TileOrigin == RequestedTileOrigin
+                        && Candidate.TileStep == RequestedTileStep
+                        && Candidate.TileCells == RequestedTileCells));
+            const bool bSameBounds = Candidate.SMinX == DesiredSMinX
+                && Candidate.SMaxX == DesiredSMaxX
+                && Candidate.SMinY == DesiredSMinY
+                && Candidate.SMaxY == DesiredSMaxY;
+            const bool bContainsQuery = WarpedX >= Candidate.SMinX
+                && WarpedX <= Candidate.SMaxX
+                && WarpedY >= Candidate.SMinY
+                && WarpedY <= Candidate.SMaxY;
+            if (bSameKey && bSameBounds && bContainsQuery)
+            {
+                CacheSlot = &Candidate;
+                break;
+            }
+        }
+        if (CacheSlot == nullptr)
+        {
+            CacheSlot = EvictionSlot;
+        }
+        CacheSlot->LastUse = ++RoomGraphCacheUseSerial;
+        FChunkSDFCache& SDFCache = CacheSlot->Cache;
+        float& CachedSMinX = CacheSlot->SMinX;
+        float& CachedSMaxX = CacheSlot->SMaxX;
+        float& CachedSMinY = CacheSlot->SMinY;
+        float& CachedSMaxY = CacheSlot->SMaxY;
+        int32& CachedStrate = CacheSlot->Strate;
+        uint32& CachedSeed = CacheSlot->Seed;
+        uint32& CachedFingerprint = CacheSlot->Fingerprint;
+        uint32& CachedLayout = CacheSlot->Layout;
+        bool& CachedUsesFusedCacheWindow = CacheSlot->bUsesFusedCacheWindow;
+        bool& CachedUsesTileCacheWindow = CacheSlot->bUsesTileCacheWindow;
+        FIntVector& CachedTileOrigin = CacheSlot->TileOrigin;
+        int32& CachedTileStep = CacheSlot->TileStep;
+        int32& CachedTileCells = CacheSlot->TileCells;
+        uint64& CachedManagerLifetimeId = CacheSlot->ManagerLifetimeId;
+        const TArray<FStrateTerrainOpEntry>*& CachedTerrainOps = CacheSlot->TerrainOps;
+
    if (Params.RoomDensity > 0.0f && Params.RoomSpacing > 0.0f)
    {
-        // Get strate index for unique caves per strate.
-        // MEMOISED: the index only changes at chunk-Z boundaries, yet this ran the strate layout's
-        // linear scan PER VOXEL. Key = (chunk-Z, layout version) — the version guards against editor
-        // rebuilds (RebuildStrates / live edit) serving a stale index. The lookup queries the BAND
-        // CENTRE so the result is a pure function of the key (order-independent, window-invariant).
-        int32 StrateIdx = 0;
-        if (StrateManager)
-        {
-            thread_local int32  SI_ChunkZ  = INT32_MAX;
-            thread_local uint32 SI_Version = 0xFFFFFFFFu;
-            thread_local int32  SI_Index   = 0;
-            const int32 QZ = FMath::FloorToInt(WorldZ / (float)CHUNK_SIZE);
-            const uint32 LV = StrateManager->GetLayoutVersion();
-            if (QZ != SI_ChunkZ || LV != SI_Version)
-            {
-                SI_ChunkZ  = QZ;
-                SI_Version = LV;
-                // GetStrateIndex expects Unreal world units (divides by VOXEL_SIZE internally);
-                // WorldZ here is in voxels, so multiply by VOXEL_SIZE.
-                SI_Index = StrateManager->GetStrateIndex(((float)QZ + 0.5f) * CHUNK_SIZE * VOXEL_SIZE);
-            }
-            StrateIdx = SI_Index;
-        }

        // Rebuild only when the WARPED query (what EvaluateSDFCached uses) leaves the
        // cached search box, or the strate/seed changed.
        const bool bNeedRebuild =
-            StrateIdx != CachedStrate || (uint32)Seed != CachedSeed ||
+            !CacheSlot->bValid || StrateIdx != CachedStrate || (uint32)Seed != CachedSeed ||
            ParamsFingerprint != CachedFingerprint || LayoutVersion != CachedLayout ||
            CurrentManagerLifetimeId != CachedManagerLifetimeId ||
            TerrainOps != CachedTerrainOps ||
@@ -4334,58 +4492,44 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo

        if (bNeedRebuild)
        {
-            // The lowered evaluator uses the same window-invariance contract as the prepared
-            // room-graph source.  With the switch enabled, the search box is the exact requesting
-            // tile footprint plus the one-sample normal halo and deterministic warp margin.  A
-            // tile worker can visit neighbouring chunk keys in an arbitrary order without
-            // rebuilding the graph at every coarse sample.  Point queries without a tile context
-            // retain the historical fused four-chunk / ordinary one-chunk fallback.
-            const float ChunkMinX = CacheRegionX * (float)CacheRegionSize;
-            const float ChunkMinY = CacheRegionY * (float)CacheRegionSize;
-            const float ChunkMaxX = ChunkMinX + (float)CacheRegionSize;
-            const float ChunkMaxY = ChunkMinY + (float)CacheRegionSize;
-            const float WarpMargin = FMath::Abs(Params.CaveWarpStrength)
-                * VOXEL_NOISE_SCALE * 1.5f;
-            float SMinX = ChunkMinX;
-            float SMinY = ChunkMinY;
-            float SMaxX = ChunkMaxX;
-            float SMaxY = ChunkMaxY;
-            if (bUseTileCacheWindow)
+            if (VoxelDensityProfile::AreCountersEnabled())
            {
-                const float TileMinX = static_cast<float>(RequestedTileOrigin.X);
-                const float TileMinY = static_cast<float>(RequestedTileOrigin.Y);
-                const float TileExtent = static_cast<float>(
-                    static_cast<int64>(RequestedTileStep)
-                    * static_cast<int64>(RequestedTileCells));
-                const float TileMaxX = TileMinX + TileExtent;
-                const float TileMaxY = TileMinY + TileExtent;
-                const float TileHalo = static_cast<float>(RequestedTileStep) + 2.0f;
-                const float Expansion = WarpMargin + TileHalo;
-                SMinX = TileMinX - Expansion;
-                SMinY = TileMinY - Expansion;
-                SMaxX = TileMaxX + Expansion;
-                SMaxY = TileMaxY + Expansion;
+                const auto CountIf = [](bool bCondition, VoxelDensityProfile::ECounter Counter)
+                {
+                    if (bCondition) VoxelDensityProfile::AddCounter(Counter);
+                };
+                CountIf(StrateIdx != CachedStrate, VoxelDensityProfile::ECounter::SdfRebuildStrate);
+                CountIf(static_cast<uint32>(Seed) != CachedSeed, VoxelDensityProfile::ECounter::SdfRebuildSeed);
+                CountIf(ParamsFingerprint != CachedFingerprint, VoxelDensityProfile::ECounter::SdfRebuildFingerprint);
+                CountIf(LayoutVersion != CachedLayout, VoxelDensityProfile::ECounter::SdfRebuildLayout);
+                CountIf(CurrentManagerLifetimeId != CachedManagerLifetimeId,
+                    VoxelDensityProfile::ECounter::SdfRebuildManagerLifetime);
+                CountIf(TerrainOps != CachedTerrainOps, VoxelDensityProfile::ECounter::SdfRebuildTerrainOps);
+                CountIf(bUseFusedCacheWindow != CachedUsesFusedCacheWindow,
+                    VoxelDensityProfile::ECounter::SdfRebuildFusedPolicy);
+                CountIf(bUseTileCacheWindow != CachedUsesTileCacheWindow,
+                    VoxelDensityProfile::ECounter::SdfRebuildTilePolicy);
+                CountIf(bUseTileCacheWindow
+                        && (RequestedTileOrigin != CachedTileOrigin
+                            || RequestedTileStep != CachedTileStep
+                            || RequestedTileCells != CachedTileCells),
+                    VoxelDensityProfile::ECounter::SdfRebuildTileIdentity);
+                CountIf(WarpedX < CachedSMinX || WarpedX > CachedSMaxX,
+                    VoxelDensityProfile::ECounter::SdfRebuildBoundsX);
+                CountIf(WarpedY < CachedSMinY || WarpedY > CachedSMaxY,
+                    VoxelDensityProfile::ECounter::SdfRebuildBoundsY);
            }
-            else
-            {
-                const float Expansion = WarpMargin + 2.0f;
-                SMinX = ChunkMinX - Expansion;
-                SMinY = ChunkMinY - Expansion;
-                SMaxX = ChunkMaxX + Expansion;
-                SMaxY = ChunkMaxY + Expansion;
-            }
-
            VoxelCaveMorphology::BuildChunkCache(
                SDFCache,
-                SMinX, SMinY, SMaxX, SMaxY,
+                DesiredSMinX, DesiredSMinY, DesiredSMaxX, DesiredSMaxY,
                Params, (uint32)Seed, StrateIdx,
                TerrainOps,
                ERoomGraphBuildSite::GeneratorTile
            );
            UpdateTunnelCoreTileReach(SDFCache, bUseTileCacheWindow);

-            CachedSMinX = SMinX; CachedSMaxX = SMaxX;
-            CachedSMinY = SMinY; CachedSMaxY = SMaxY;
+            CachedSMinX = DesiredSMinX; CachedSMaxX = DesiredSMaxX;
+            CachedSMinY = DesiredSMinY; CachedSMaxY = DesiredSMaxY;
            CachedStrate = StrateIdx;
            CachedSeed = (uint32)Seed;
            CachedFingerprint = ParamsFingerprint;
@@ -4398,6 +4542,7 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo
            CachedTileCells = bUseTileCacheWindow ? RequestedTileCells : 0;
            CachedManagerLifetimeId = CurrentManagerLifetimeId;
            CachedTerrainOps = TerrainOps;
+            CacheSlot->bValid = true;
        }

        // Evaluate SDF using cached rooms and tunnels (WARPED coordinates).
@@ -103,13 +103,15 @@ FVoxelMeshData UVoxelMarchingCubesMesher::GenerateMesh(FIntVector OriginVoxels,
                                                       int32 BandZMinVox, int32 BandZMaxVox,
                                                       int64* OutDensitySampleCount,
                                                       FVoxelSharedDensityGrid* OutPreEditDensityGrid,
-                                                       const FVoxelSharedDensityGrid* InSharedDensityGrid)
+                                                       const FVoxelSharedDensityGrid* InSharedDensityGrid,
+                                                       std::vector<float>* OutDebugDensitySamples)
 {
    TRACE_CPUPROFILER_EVENT_SCOPE(VoxelForge_MesherGenerateMesh);
    FVoxelMeshData MeshData;
    if (OutCaptureGrid) { OutCaptureGrid->Reset(); }
    if (OutPreEditDensityGrid) { OutPreEditDensityGrid->Reset(); }
    if (OutDensitySampleCount) { *OutDensitySampleCount = 0; }
+    if (OutDebugDensitySamples) { OutDebugDensitySamples->clear(); }
    if (!Generator) return MeshData;
    VoxelDensityProfile::FScopedTimer MesherProfileTimer(
        VoxelDensityProfile::EBucket::MesherGenerateMesh);
@@ -284,7 +286,7 @@ FVoxelMeshData UVoxelMarchingCubesMesher::GenerateMesh(FIntVector OriginVoxels,
    //
    // CellsPerAxis = nombre de cellules par axe ; GridDim = points de grille (+1).
    // Le point de grille (gx,gy,gz) correspond au voxel monde (gx,gy,gz)*Step.
-    // Ordre de remplissage z→y→x : garde le cache SDF (search-box) bien chaud.
+    // The grid fill order is selected below by tile extent to preserve room-cache locality.
    // On échantillonne avec un anneau de marge de 1 point de chaque côté (indices -1..GridDim)
    // pour pouvoir calculer les normales par GRADIENT DE GRILLE (T1.b) — différences centrales
    // sur la grille au lieu de 6 appels densité frais par vertex. La marge utilise les mêmes
@@ -869,56 +871,114 @@ FVoxelMeshData UVoxelMarchingCubesMesher::GenerateMesh(FIntVector OriginVoxels,
        }
 #endif
    }
-    else for (int32 gz = GzLo; gz <= GzHi; gz++)
+    else
    {
-        if (ShouldAbortWork())
+        auto BuildAxisGroups = [&](int32 AxisOrigin, int32 RegionSize)
        {
-            if (bUseOperatorBlock) { Generator->EndDensityBlock(); }
-            return FVoxelMeshData();
-        }
-        for (int32 gy = -1; gy <= GridDim; gy++)
-        {
-            if (ShouldAbortWork())
+            std::vector<std::vector<int32>> Groups;
+            int32 PreviousRegion = INT32_MIN;
+            for (int32 Index = -1; Index <= GridDim; ++Index)
            {
-                if (bUseOperatorBlock) { Generator->EndDensityBlock(); }
-                return FVoxelMeshData();
-            }
-            for (int32 gx = -1; gx <= GridDim; gx++)
-            {
-                const int32 LocalIndex = ((gz + 1) * MDim + (gy + 1)) * MDim + (gx + 1);
-                if (bUseSharedDensityGrid)
+                const int64 WorldCoordinate = static_cast<int64>(AxisOrigin)
+                    + static_cast<int64>(Index) * Step;
+                const int32 Region = FMath::FloorToInt(
+                    static_cast<float>(WorldCoordinate) / static_cast<float>(RegionSize));
+                if (Groups.empty() || Region != PreviousRegion)
                {
-                    const int32 SharedX = SharedDeltaX + gx;
-                    const int32 SharedY = SharedDeltaY + gy;
-                    const int32 SharedZ = SharedDeltaZ + gz;
-                    DensityGrid[LocalIndex] = ReuseGrid->Samples[
-                        (SharedZ * ReuseGrid->Dim + SharedY) * ReuseGrid->Dim + SharedX];
+                    Groups.emplace_back();
+                    PreviousRegion = Region;
                }
-                else
+                Groups.back().push_back(Index);
+            }
+            return Groups;
+        };
+        const int32 RoomCacheRegionSize = CHUNK_SIZE
+            * (VoxelGenLOD::IsUseFusedEvaluatorEnabled() ? 4 : 1);
+        const auto XGroups = BuildAxisGroups(OriginVoxels.X,RoomCacheRegionSize);
+        const auto YGroups = BuildAxisGroups(OriginVoxels.Y,RoomCacheRegionSize);
+        auto SampleGridPoint = [&](int32 gx, int32 gy, int32 gz)
+        {
+            const int32 LocalIndex = ((gz + 1) * MDim + (gy + 1)) * MDim + (gx + 1);
+            if (bUseSharedDensityGrid)
+            {
+                const int32 SharedX = SharedDeltaX + gx;
+                const int32 SharedY = SharedDeltaY + gy;
+                const int32 SharedZ = SharedDeltaZ + gz;
+                DensityGrid[LocalIndex] = ReuseGrid->Samples[
+                    (SharedZ * ReuseGrid->Dim + SharedY) * ReuseGrid->Dim + SharedX];
+            }
+            else
+            {
+                const int32 ReachBlockX = FMath::Clamp(gx, 0, CellsPerAxis - 1)
+                    / VoxelGenLOD::TileReachBlockCells;
+                const int32 ReachBlockY = FMath::Clamp(gy, 0, CellsPerAxis - 1)
+                    / VoxelGenLOD::TileReachBlockCells;
+                const int32 ReachBlockZ = FMath::Clamp(gz, 0, CellsPerAxis - 1)
+                    / VoxelGenLOD::TileReachBlockCells;
+                VoxelGenLOD::TileReachBlockIndex =
+                    ((ReachBlockZ * ReachBlocksPerAxis) + ReachBlockY)
+                        * ReachBlocksPerAxis + ReachBlockX;
+                float PreEditDensity = 0.0f;
+                DensityGrid[LocalIndex] = Generator->GetDensityAt(
+                    OriginVoxels.X + gx * Step,
+                    OriginVoxels.Y + gy * Step,
+                    OriginVoxels.Z + gz * Step,
+                    bCapturePreEditGrid ? &PreEditDensity : nullptr);
+                if (bCapturePreEditGrid)
                {
-                    // World voxel = tile origin + grid offset scaled by the cell size (Step).
-                    const int32 ReachBlockX = FMath::Clamp(gx, 0, CellsPerAxis - 1)
-                        / VoxelGenLOD::TileReachBlockCells;
-                    const int32 ReachBlockY = FMath::Clamp(gy, 0, CellsPerAxis - 1)
-                        / VoxelGenLOD::TileReachBlockCells;
-                    const int32 ReachBlockZ = FMath::Clamp(gz, 0, CellsPerAxis - 1)
-                        / VoxelGenLOD::TileReachBlockCells;
-                    VoxelGenLOD::TileReachBlockIndex =
-                        ((ReachBlockZ * ReachBlocksPerAxis) + ReachBlockY)
-                            * ReachBlocksPerAxis + ReachBlockX;
-                    float PreEditDensity = 0.0f;
-                    DensityGrid[LocalIndex] = Generator->GetDensityAt(
-                        OriginVoxels.X + gx * Step,
-                        OriginVoxels.Y + gy * Step,
-                        OriginVoxels.Z + gz * Step,
-                        bCapturePreEditGrid ? &PreEditDensity : nullptr);
-                    if (bCapturePreEditGrid)
+                    OutPreEditDensityGrid->Samples[LocalIndex] = PreEditDensity;
+                }
+                if (OutDensitySampleCount)
+                {
+                    ++(*OutDensitySampleCount);
+                }
+            }
+        };
+        // Very coarse tiles revisit the same XY cache window across a tall Z run.  Keep that
+        // vertical run contiguous; SampleGridPoint writes by lattice index, so output bits do
+        // not depend on this traversal.  Finer tiles retain plane-first traversal, which has
+        // lower overhead when a tile spans only one room-graph region.
+        const int64 TileExtentVoxels = static_cast<int64>(Step) * CellsPerAxis;
+        const bool bUseVerticalCacheTraversal = TileExtentVoxels > CHUNK_SIZE * 16;
+        if (bUseVerticalCacheTraversal)
+        {
+            for (const std::vector<int32>& YGroup : YGroups)
+            for (const std::vector<int32>& XGroup : XGroups)
+            for (const int32 gy : YGroup)
+            for (const int32 gx : XGroup)
+            {
+                for (int32 gz = GzLo; gz <= GzHi; ++gz)
+                {
+                    if (ShouldAbortWork())
                    {
-                        OutPreEditDensityGrid->Samples[LocalIndex] = PreEditDensity;
+                        if (bUseOperatorBlock) { Generator->EndDensityBlock(); }
+                        return FVoxelMeshData();
                    }
-                    if (OutDensitySampleCount)
+                    SampleGridPoint(gx,gy,gz);
+                }
+            }
+        }
+        else
+        {
+            for (int32 gz = GzLo; gz <= GzHi; ++gz)
+            {
+                if (ShouldAbortWork())
+                {
+                    if (bUseOperatorBlock) { Generator->EndDensityBlock(); }
+                    return FVoxelMeshData();
+                }
+                for (const std::vector<int32>& YGroup : YGroups)
+                for (const std::vector<int32>& XGroup : XGroups)
+                for (const int32 gy : YGroup)
+                {
+                    if (ShouldAbortWork())
                    {
-                        ++(*OutDensitySampleCount);
+                        if (bUseOperatorBlock) { Generator->EndDensityBlock(); }
+                        return FVoxelMeshData();
+                    }
+                    for (const int32 gx : XGroup)
+                    {
+                        SampleGridPoint(gx,gy,gz);
                    }
                }
            }
@@ -943,6 +1003,19 @@ FVoxelMeshData UVoxelMarchingCubesMesher::GenerateMesh(FIntVector OriginVoxels,
    {
        Generator->EndDensityBlock();
    }
+    if (OutDebugDensitySamples)
+    {
+        const size_t SampleCount = static_cast<size_t>(GzHi - GzLo + 1)
+            * static_cast<size_t>(MDim) * static_cast<size_t>(MDim);
+        OutDebugDensitySamples->reserve(SampleCount);
+        for (int32 gz = GzLo; gz <= GzHi; ++gz)
+        for (int32 gy = -1; gy <= GridDim; ++gy)
+        for (int32 gx = -1; gx <= GridDim; ++gx)
+        {
+            const int32 LocalIndex = ((gz + 1) * MDim + (gy + 1)) * MDim + (gx + 1);
+            OutDebugDensitySamples->push_back(DensityGrid[LocalIndex]);
+        }
+    }
    MesherDensityGridTimer.End();

    if (bUseProofBlocks)
@@ -172,6 +172,28 @@ namespace VoxelDensityProfile
        MesherOutputAllocatedBytes,
        OpBlockScratchBytes,
        OpBlockCopyBytes,
+        SdfRebuildStrate,
+        SdfRebuildSeed,
+        SdfRebuildFingerprint,
+        SdfRebuildLayout,
+        SdfRebuildManagerLifetime,
+        SdfRebuildTerrainOps,
+        SdfRebuildFusedPolicy,
+        SdfRebuildTilePolicy,
+        SdfRebuildTileIdentity,
+        SdfRebuildBoundsX,
+        SdfRebuildBoundsY,
+        SdfTileWindowAccepted,
+        SdfTileWindowRejected,
+        TilePolicyKeyOrigin,
+        TilePolicyKeyStep,
+        TilePolicyKeyCells,
+        TilePolicyKeyFingerprint,
+        TilePolicyKeyLayout,
+        TilePolicyKeyManager,
+        TilePolicyKeyTerrainOps,
+        TilePolicyKeyFusedPath,
+        TilePolicyKeyContext,
        Count
    };

@@ -433,6 +433,8 @@ private:

 namespace VoxelGenLOD
 {
+    VOXELFORGE_API void SetUseFusedEvaluatorEnabled(bool bEnabled);
+    VOXELFORGE_API bool IsUseFusedEvaluatorEnabled();
    // -1 uses the measured path/LOD policy; 0/1 force the spatial broad phase off/on for an A/B
    // run. The policy affects candidate enumeration only, never cache contents or ordering.
    VOXELFORGE_API bool ShouldUseSpatialIndex(bool bFusedPath);
@@ -13,6 +13,7 @@

 #include "CoreMinimal.h"
 #include <atomic>
+#include <vector>
 #include "VoxelTypes.h"   // Pour FVoxelMeshData, CHUNK_SIZE, VOXEL_SIZE, etc.
 #include "VoxelGenerator.h"
 #include "VoxelMarchingCubesMesher.generated.h"
@@ -78,7 +79,8 @@ public:
                                int32 BandZMinVox = INT32_MIN, int32 BandZMaxVox = INT32_MAX,
                                int64* OutDensitySampleCount = nullptr,
                                FVoxelSharedDensityGrid* OutPreEditDensityGrid = nullptr,
-                                const FVoxelSharedDensityGrid* InSharedDensityGrid = nullptr);
+                                const FVoxelSharedDensityGrid* InSharedDensityGrid = nullptr,
+                                std::vector<float>* OutDebugDensitySamples = nullptr);

    //=========================================================================
    // SERVICES (injectés par AVoxelWorld)