wip(step3): 12.5 cm voxels, fall-through fixed, brick evaluator in game, organic tunnels -- speed gate FAILS

Stopped by the owner's decision (in-game lit captures were eating the round); he judges the look in
game instead.
1. Fall-through with the research field: fixed. Deep drop, gate on: 2 holds (0.50 s), min foot
   clearance +43 cm; gate off: penetrates (-278 cm). Stress fixture offsets converted to metres.
2. VOXEL_SIZE 25 -> 12.5 cm. Physical sizes kept in metres (player, steps, gate, insets, research
   field sizes, dig radii, test thresholds); class defaults that count chunks/voxels doubled
   (ClipRadius 6, RenderDistanceChunks 1536, ViewDistance*, StrateViewMargin 6,
   PawnCollisionRadius 2, SkirtCells 4, EdgeSealThickness 128, decoration spacings, density volume
   res 256). Asset values override these defaults: the owner must double his own.
   Ring scale, six workers, idle, step-2f field before the look changes: 64 m cube (519^3) 4.55 s;
   32 m cube (259^3) 0.95 s.
3. In game: tiles go through the brick evaluator (proved by a log marker); profile switch rebuilds
   the loaded world (~5,800 tiles); full-detail fill 12.4 s after spawn in the visible-window test.
4. Look: round winding tunnels, per-column ellipsoid/lattice caches. All step-2f geometry gates pass
   at 12.5 cm (0 crumbs, 1 solid component per profile, solid share 69-87%). 259^3 ring, six workers:
   1.08-2.02 s (worst HyperComplex) -- the 600 ms gate FAILS. Lit captures failed (hidden/off-screen
   window, then blank frames); not pursued.

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

CODEMAP.md

@@ -28,7 +28,7 @@ A **density-field voxel terrain** plugin for Unreal Engine, built around undergr
 ### Core terminology
 | Term | Meaning |
 |------|---------|
-| Chunk | 32×32×32 voxels (`CHUNK_SIZE`). Voxel = 25 cm (`VOXEL_SIZE`). |
+| Chunk | 32×32×32 voxels (`CHUNK_SIZE`). Voxel = 12.5 cm (`VOXEL_SIZE`), so a chunk is 4 m. |
 | Density | Scalar field. `< 0` solid, `> 0` air, `0` = surface (`IsoLevel`). |
 | Strate | Vertical layer of the world, stacked downward. Has its own generator + params. |
 | Generator type | Archetype per strate: `TunnelNetwork`, `FlatPlain`, `CrystalChamber`, `Maze`, `SurfaceWorld`, `VerticalShafts`, `FloatingIslands`, `Underwater`. See §8. |
@@ -76,17 +76,45 @@ what generator type, and is there a passage/elevator SDF near here?"
 
 Paths relative to `Source/VoxelForge/`. `Public/` = headers, `Private/` = impl.
 
-### 3.0 Human-scale generation contract (2026-09-06)
-`VoxelStrateTypes.h` is now authored against the confirmed player capsule, not against naked voxel
-counts: one voxel is **0.25 m**, the capsule is **1.36 voxels / 0.34 m radius** and **7.04 voxels /
-1.76 m tall**, and the default strate envelope is **8 chunks / 256 voxels / 64 m**. Core defaults are
-therefore: room diameters **8-20 m**, origin hub **24 m**, room cells **32 m**, round tunnel bores
-**3-4 m**, maze cells **16 m** with a **6.25 m nominal bore** (**3.125 m** radius; **3.0625 m**
-worst-case usable floor after roughness), slab open span **22.4 m**, shaft diameters
-**4-7 m**, island diameters **12-24 m**, and ordinary inter-strate bores **3-4 m**. These are
-different derivations, not a global multiplier: passage widths follow the 2.5-4 m walking/fight
-target, room/cell spacing preserves a rock web, and the 64 m vertical envelope reserves space for
-tall rooms, surface sky, and vertical traversal.
+### 3.0 Human-scale generation contract (2026-10-01)
+`VOXEL_SIZE` is **12.5 cm**; 32³ voxels make each chunk **4 m**. The gameplay capsule remains
+**0.34 m radius / 1.76 m tall** (**2.72 / 14.08 voxels**). The default strate envelope is
+**16 chunks / 512 voxels / 64 m**; serialized owner strate assets that still say 8 chunks must be
+changed to 16. Research-field dimensions are authored in metres and converted at runtime. The live
+strate generator continues to read voxel-valued parameters from its assets, so those authored
+dimensions become half as large in metres unless their asset values are doubled.
+
+The intended metre-scale families remain: room diameters **8-20 m**, origin hub **24 m**, room cells
+**32 m**, round tunnel bores **3-4 m**, maze cells **16 m** with a **6.25 m nominal bore** (**3.125 m**
+radius; **3.0625 m** worst-case usable floor after roughness), slab open span **22.4 m**, shaft
+diameters **4-7 m**, island diameters **12-24 m**, and ordinary inter-strate bores **3-4 m**. Source
+new fixed physical dimensions are authored in metres and converted to voxel counts at the boundary.
+The owner's voxel-valued live strate assets are not migrated by C++: unchanged saved values therefore
+produce half-size rooms, tunnels, features and noise periods in metres. See the step-3 migration list
+in the handoff report for the exact `DA_Settings` fields and values.
+
+**Owner asset migration (no packages were edited):** preserve `DA_Settings` distances by changing
+`ViewDistanceXY 32→64`, `ViewDistanceUp/Down 10→20`, `PawnCollisionRadiusChunks 2→4`,
+`StrateViewMarginChunks 6→12`, `ClipRadius 6→12`, `RenderDistanceChunks 1536→3072`,
+`EdgeSealThickness 128→256 voxels`, decoration radius/near radius `12/6→24/12 chunks`, region
+size `8→16 cells`, spacing/far spacing `8→16 voxels`, march step `4→8 voxels`, column depth
+`320→640 voxels`, `OriginSpineRadius 28→56 voxels`, `MaxBrushRadius 30→60 voxels`, and
+`SkirtCells 1→2`. If the density volume is enabled, set resolution `256→512` and debug radius
+`12→24 cells`; it is currently disabled. Keep `MaxClipLevel=6` and zero-valued disabled/unbounded
+limits unchanged; double finite world radii and gaps, and multiply finite voxel-volume caps by 8.
+The loaded `Crystal Cave` slot in `DA_Strate3` is 64 chunks tall; use 128 chunks to preserve its
+former 512 m span. Its voxel-authored live-generator lengths remain half-size until that asset is
+re-authored; the editor-only research field is the step-3 target.
+
+Step-3 six-worker scale benchmark (idle machine, `Saved/Research/step3/scale/step3_scale_cacheperf_v1`):
+the 519³ field took **7.133834–7.825292 s** across three runs (median **7.347496 s**); the 259³
+field took **1.098521–1.194640 s** (median **1.169567 s**). The inclusive sample lattices span
+64.75 m and 32.25 m between end samples respectively (the requested approximately 64 m / 32 m
+fields). Both runs were deterministic across three trials. This records evaluator time only, not game
+streaming. Six-worker 4³-brick proof, sound-bound audit, seam injection, and 1-vs-6-worker hashes
+pass. Reusing exact detail-lattice hash corners across adjacent mixed bricks reduced the earlier
+step-3 259³ median from 1.502908 s to 1.169567 s; the separate step-2f 600 ms worst-profile timing
+bar still does not pass at the finer voxel size.
 
 The reachable roughness envelope is `roughness * VOXEL_NOISE_SCALE * 1.5`, or `roughness * 1.875`
 voxels. The new defaults keep it below the relevant feature radius: **0.625** of the minimum
@@ -117,11 +145,12 @@ remains `0`, and lateral regions remain gated off.
 |------|------|
 | `../../VoxelForge.uplugin` | Plugin manifest. Runtime module `VoxelForge` plus the Editor-only `VoxelForgeEditor` commandlet module. Beta. |
 | `VoxelForge.Build.cs` | Deps: Core, CoreUObject, Engine, **GameplayTags**, **RealtimeMeshComponent**. |
-| `Tools/VoxelForgeTest.ps1` | The single guarded test entry point. `-Scenario canonical|owner|probe|perf|tests|surface-fall|crossing|gate-stress|visual-edit|visual-lod|visual-gap-walk|visual-gap-inject|dig-motion|parity|horizon|horizon-render|determinism|region-eval|step2d|step2e|step2f|step2f-capture`, optional `-Build`, `-Assets owner|default`, `-Cvars`, absolute `-Out`, `-Label`, `-HorizonLevel`, `-HorizonRenderDistanceChunks`, `-Step2fProfile=0..7`, `-Step2fCaptureView=inside|wall-closeup`, `-ParityExportSize=32|64|128`, `-ParityExportTile=x,y,z`, `-SharedDDC`, `-SampleDensityStacks`, `-StackSampleIntervalUs`, and `-RegionHashProbe -RegionWorkers N -RegionReferenceHash HASH`; stages the host/assets, isolates user/DDC/config state, serializes one Unreal launch at a time, enforces the 8 GiB/30-minute guards, waits for 6 GiB free RAM before launches, and emits compact `result.json`/`summary.txt` plus full logs. Parity can render its game leg with `-EnableRender` and compares the selected export region against the exact streamed tile hashes. `visual-gap-walk` records camera-frustum surface-point coverage at 6000 cm/s; `visual-gap-inject` proves an early component hide is detected; `dig-motion` logs per-frame vertical velocity, gate, support readiness, and movement mode. `-Cvars @{ 'voxel.TestGapAuditLegacyHandoff'=1 }` replays the old visibility handoff for A/B audit baselines. Horizon traces report desired/loaded tiles, resident stream geometry, worker seconds, steady wall time, and peak private memory; `horizon-render` captures the owner view; `step2f-capture` selects one of eight parameter sets, finds a positive-density camera point and aims at a sampled rock wall for an inside or wall-closeup view, then saves it under `Saved/Research/step2f/captures` (pass the persistent render DDC with `-SharedDDC`); `determinism` compares per-tile hashes across the default CRT path and `-voxel.CrtFma3=0`; `region-eval` compares the live path with the default-off editor chunk program on the 259³ lattice and five-seed probe set; `step2d` retains the earlier independent field; `step2e` retains its earlier benchmark; `step2f` compares frozen step 2e and the coherent field for all eight profiles on the same 259³ ring, applies geometry gates, and checks sound bounds, seams, and 1-vs-6-worker hashes. `-RegionHashProbe -RegionWorkers N -RegionReferenceHash HASH` checks a full-ring hash at another partition count. `-SampleDensityStacks` captures profiles, and `-StackSampleIntervalUs` sets the sampling interval.
+| `Tools/VoxelForgeTest.ps1` | The single guarded test entry point. `-Scenario canonical|owner|probe|perf|tests|surface-fall|crossing|gate-stress|visual-edit|visual-lod|visual-gap-walk|visual-gap-inject|dig-motion|parity|horizon|horizon-render|determinism|region-eval|step2d|step2e|step2f|step2f-capture|step3-scale|step3-game`, optional `-Build`, `-Assets owner|default`, `-Cvars`, absolute `-Out`, `-Label`, `-HorizonLevel`, `-HorizonRenderDistanceChunks`, `-Step2fProfile=0..7`, `-Step2fCaptureView=inside|wall-closeup`, `-ParityExportSize=32|64|128`, `-ParityExportTile=x,y,z`, `-SharedDDC`, `-SampleDensityStacks`, `-StackSampleIntervalUs`, and `-RegionHashProbe -RegionWorkers N -RegionReferenceHash HASH`; stages the host/assets, isolates user/DDC/config state, serializes one Unreal launch at a time, enforces the 8 GiB/30-minute guards, waits for 6 GiB free RAM before launches, and emits compact `result.json`/`summary.txt` plus full logs. Parity can render its game leg with `-EnableRender` and compares the selected export region against the exact streamed tile hashes. `visual-gap-walk` records camera-frustum surface-point coverage at 6000 cm/s; `visual-gap-inject` proves an early component hide is detected; `dig-motion` logs per-frame vertical velocity, gate, support readiness, and movement mode. `-Cvars @{ 'voxel.TestGapAuditLegacyHandoff'=1 }` replays the old visibility handoff for A/B audit baselines. Horizon traces report desired/loaded tiles, resident stream geometry, worker seconds, steady wall time, and peak private memory; `horizon-render` captures the owner view; `step2f-capture` selects one of eight parameter sets, finds a positive-density camera point and aims at a sampled rock wall for an inside or wall-closeup view, then saves it under `Saved/Research/step2f/captures` (pass the persistent render DDC with `-SharedDDC`); `determinism` compares per-tile hashes across the default CRT path and `-voxel.CrtFma3=0`; `region-eval` compares the live path with the default-off editor chunk program on the 259³ lattice and five-seed probe set; `step2d` retains the earlier independent field; `step2e` retains its earlier benchmark; `step2f` compares frozen step 2e and the coherent field for all eight profiles on the same 259³ ring, applies geometry gates, and checks sound bounds, seams, and 1-vs-6-worker hashes. `-RegionHashProbe -RegionWorkers N -RegionReferenceHash HASH` checks a full-ring hash at another partition count. `-SampleDensityStacks` captures profiles, and `-StackSampleIntervalUs` sets the sampling interval.
+| Step-3 scale and game validation | `step3-scale` measures the same 519³ and 259³ six-worker lattices at 12.5 cm. The cached medians are 7.347496 s and 1.169567 s respectively; determinism and brick sound-proof audits pass. `step3-game` routes in-game generation through the brick evaluator and records tile timings, spawn fill, moving frames, and profile-switch rebuilds. The visible D3D12 run measured tile generation p50/p95/max 3.326/31.045/169.867 ms, full-detail L0 fill 12.369 s for 2,765 tiles, and moving-frame p50/p95/p99/max 4.856/8.922/13.090/80.575 ms. Switching profiles rebuilt 5,815 loaded tiles. Both high-resolution screenshot requests were logged, but the harness found only one fresh PNG and it is a blank pale viewport; no lit comparison capture is accepted. Raw output: `Saved/TestHarness/step3_game_visible_final`; scale output: `Saved/Research/step3/scale/step3_scale_cacheperf_v1`. |
 | `Public/VoxelForgeModule.h` / `Private/VoxelForgeModule.cpp` | `FVoxelForgeModule` boilerplate (Startup/Shutdown just log). |
-| `Source/VoxelForgeEditor/VoxelForgeEditor.Build.cs` / `Private/VoxelForgeExploreCommandlet.*` | Editor/commandlet-only `-run=VoxelForgeExplore`; owns bounded render/`VF_` walk/OBJ export orchestration and is not linked by Game/Shipping. Optional `-exporttilex/y/z` sets an exact L0 tile origin for controlled export/game parity; omitting all three preserves the existing center-derived origin. `-regioneval=1` retains the exact-field step-2c research harness. `-step2d` and `-step2e` retain their earlier benchmarks. `-step2f` compares frozen step 2e and the coherent field for all eight profiles on the same 259³ ring, counts sub-1 m³ connected components, measures mesh dihedrals, detail spectrum, LOD surface distance (p95 threshold 0.5 lattice samples), chamber-pitch autocorrelation, roof/floor coverage, and six-worker timing. It also runs the sound/undersized-bound audit, seam injection, and 1-vs-6-worker hash checks and writes profile slices and JSON under `Saved/Research/step2f`. |
-| `Source/VoxelForgeEditor/Private/VoxelForgeRegionEvalExperiment.*` | Six-worker lattice drivers for step 2c through step 2f. The original path still compares live `GetDensityAt` with the editor chunk program. Research modes assign contiguous batches dynamically so surface-heavy work balances across workers; steps 2e and 2f use 4³ proof bricks and stage each evaluated brick before copying contiguous X rows, while step 2d retains 8³. Each skip requires a macro Lipschitz bound plus the complete final-detail supremum; uncertain bricks retain all 25 cm detail samples. Timings split classification, point evaluation, and uniform fill; step 2f compares all profiles and hashes workers 1 and 6. |
-| `Public/VoxelRedefinedDensityField.h` / `Private/VoxelRedefinedDensityField.cpp` | Two default-off `WITH_EDITOR` research fields, selected by `voxel.ResearchRedefinedDensityField` (step 2d) and `voxel.ResearchParameterizedCave` (steps 2e/2f). Step 2f retains frozen 2e parameters for its rejection audit, while the new profiles use strongly jittered, individually rotated elliptical rooms capped against nearest neighbors, leaving a solid web joined by narrow tree/loop tunnels and capped shafts. Room heights and feature rolls vary by cell; ledges, occasional pillars, stalactites/stalagmites, basin floors, and bounded hanging masses remain features. Point and brick detail sampling use the same independently rotated 48/24/12-sample bands and smooth ridge response; the twelve-sample shortest wavelength and 3.822-unit final-field detail bound are preserved. The thread-local cell cache and per-column distance reuse remain. In an editor game with this field enabled, sealed-solid, out-of-layout, and outer-classifier verdicts are declined without changing their console switches. No reflected class layout or serialized assets change. |
+| `Source/VoxelForgeEditor/VoxelForgeEditor.Build.cs` / `Private/VoxelForgeExploreCommandlet.*` | Editor/commandlet-only `-run=VoxelForgeExplore`; owns bounded render/`VF_` walk/OBJ export orchestration and is not linked by Game/Shipping. Optional `-exporttilex/y/z` sets an exact L0 tile origin for controlled export/game parity; omitting all three preserves the existing center-derived origin. `-regioneval=1` retains the exact-field step-2c research harness. `-step2d` and `-step2e` retain their earlier benchmarks. `-step2f` compares the metre-scaled step 2e baseline and the coherent field for all eight profiles on the same 259³ ring (32.25 m at this lattice), counts sub-1 m³ connected components in a 128-sample padded volume, measures mesh dihedrals, detail spectrum, LOD surface distance (p95 threshold 1 sample = 12.5 cm), chamber-pitch autocorrelation, roof/floor coverage, and six-worker timing. It also runs the sound/undersized-bound audit, seam injection, and 1-vs-6-worker hash checks and writes profile slices and JSON under `Saved/Research/step2f`. |
+| `Source/VoxelForgeEditor/Private/VoxelForgeRegionEvalExperiment.*` | Six-worker lattice drivers for step 2c through step 2f. The original path still compares live `GetDensityAt` with the editor chunk program. Research modes assign contiguous batches dynamically so surface-heavy work balances across workers; steps 2e and 2f use 4³ proof bricks and stage each evaluated brick before copying contiguous X rows, while step 2d retains 8³. Each skip requires a macro Lipschitz bound plus the complete final-detail supremum; uncertain bricks retain every 12.5 cm detail sample. Timings split classification, point evaluation, and uniform fill; step 2f compares all profiles and hashes workers 1 and 6. |
+| `Public/VoxelRedefinedDensityField.h` / `Private/VoxelRedefinedDensityField.cpp` | Two default-off `WITH_EDITOR` research fields, selected by `voxel.ResearchRedefinedDensityField` (step 2d) and `voxel.ResearchParameterizedCave` (steps 2e/2f). Step 2f retains a metre-scaled 2e baseline for its rejection audit. New profiles use strongly jittered, individually rotated elliptical rooms capped against nearest neighbors and curved cubic-Bezier passages with deterministic centerline and width variation along each edge. Room heights and feature rolls vary by cell; ledges, occasional pillars, stalactites/stalagmites, basin floors, and bounded hanging masses remain features. Point and brick detail sampling share independently rotated 96/48/24-sample bands and a smooth ridge response, preserving 12/6/3 m wavelengths at 12.5 cm per sample. The parameterized mesher evaluates the exact field through prepared 4³ bricks and the same sound skip proof in commandlet and editor-game meshing. In-editor and in-game profile changes invalidate and rebuild all loaded tiles. In an editor game with this field enabled, sealed-solid, out-of-layout, and outer-classifier verdicts are declined without changing their console switches. |
 | `Private/Tests/VoxelForgeParameterizedGenerationTest.cpp` | Separate three-seed functional sweep for the step 2f parameterized field; passage landing air/floor brackets and arrival-to-exit walking routes are checked without changing the live-world tests. |
 | `Private/Tests/VoxelForgeRedefinedDensityFieldTest.cpp` | Three-seed functional sweep for the research field: authored passage endpoints must bracket open air and solid floor, and each tested strate must have a walkable arrival-to-exit route. |
 | `Public/VoxelGeometryHash.h` | Header-only geometry digest shared by commandlet export and runtime tile-hash dump; hashes mesh counts and the vertex/normal/UV/color/triangle arrays with the historical CRC order. |
@@ -131,7 +160,7 @@ remains `0`, and lateral regions remain gated off.
 | Symbol | Line | Notes |
 |--------|------|-------|
 | `CHUNK_SIZE` (32) | 211 | Chunk dimensions. 64³ would cut draw calls but makes streaming too bursty; fps is fixed render-side instead. |
-| `VOXEL_SIZE` (25.0f cm) | 213 | World scale. |
+| `VOXEL_SIZE` (12.5f cm) | 213 | World scale; one 32³ chunk is 4 m. |
 | `EVoxelFace` enum | 266 | 6 cube faces. |
 | `WorldToChunkCoord` | 287 | World cm → chunk coord (handles negatives via floor). |
 | `VoxelMath::IsFiniteFast` / `IsFinite` | 24 / 45 | Exact IEEE bit test (`voxel.FastIsFinite` A/B switch). |
@@ -494,8 +523,8 @@ context. It never touches a generator, operator stack, manager, cache, or mutabl
 `FVoxelPassageLanding`, `VF_BuildPassageLanding`, `VF_EvaluatePassageLandingSDF`, and
 `VF_IsPassageLandingFloor` (h:57/184-199, implemented near `.cpp:6468`) turn each inter-strate mouth
 into a deterministic rounded room plus a hard support slab. The room is sized from the player
-capsule (minimum 12 voxels / 3 m flat floor, 12 voxels / 3 m clear height, and at least 14 voxels /
-3.5 m authored width); the `(0,0)` origin landing is enlarged for the reserved future shaft, while
+capsule (minimum 24 voxels / 3 m flat floor and clear height); the `(0,0)` origin landing is enlarged
+for the reserved future shaft, while
 inter-strate mouths remain local rooms and deliberately have no landing-to-root connector road.
 The manager's shared thread-local passage shortlist serves tube, room, and floor evaluation; no
 landing search occurs per voxel. `GeneratePassages` stores the full descriptors and the