perf(generation): gate the tile cache window by LOD; LOD0 back to switch-off speed

2de603c's tile-sized room cache cut worker CPU, but LOD0 generation was ~9% slower. The per-tile
split shows the loss was evaluation (more candidates per sample in a bigger window), not cache
building. Policy, measured per path and LOD:
- fused path: tile window at LOD >= voxel.TileCacheWindowMinLOD (default 3)
- op-stack path: tile window at LOD >= voxel.TileCacheWindowOpStackMinLOD (default 4)
- lower LODs keep the 4-chunk worker window
- voxel.SpatialIndex: -1 auto (off at fused LOD0, on elsewhere), 0 or 1

Proof caveat closed: VF_ComputeTerrainOpEnvelope adds authored terrain-op extents (pits, chimneys,
detail overrides, ...) to the invariance envelope. Ops beyond the proven 256-voxel extent, or
malformed/unloaded ones, force the legacy window, so window invariance holds for every asset.

Sampler fix: the leaf symbol kept pointers into a TMap that FindOrAdd can rehash while later
frames resolve; the leaf is now re-found by program counter after insertion.

Game path, two runs each, sampler off (Codex; checked against its report):
                       LOD0 request p50/p95    generation p50/p95    worker s
  static  optimized    106.3 / 175.9 ms        76.6 / 124.3 ms       39.2
  static  switch-off   106.3 / 173.0           75.9 / 124.1          58.5   (-33%)
  moving  optimized     79.4 / 158.5           72.4 / 123.7         113.7
  moving  switch-off    82.8 / 160.4           73.2 / 125.4         141.9   (-20%)
LOD0 is flat within noise static and slightly better moving; worker CPU keeps most of the gain.
Field exact: 841/841 per-tile triangle identity, export OBJ SHA-256 unchanged (b3e5f4c3..9b377),
capability 20,830 / 10,909.

Caution: compare these numbers only within this round. The machine is in daytime use, so
absolute figures drift between rounds.

Current LOD0 map (sampler, moving, inclusive): tunnel SDF 12.5% + tunnel core 12.3% (the same
tunnel in two coordinate spaces); passage carving 11.1%, structural posts 11.4%, native floor
4.1%, with floor projection at 8.9% inside them; cave warp ~8% exclusive. Development-build-only
overheads: TArray RangeCheck ~6.5%, per-sample FScopedTimer ~2.6%.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y

CODEMAP.md

@@ -302,7 +302,7 @@ casse cette arithmetique *structurellement* — `BeginPlay` loggue une **Error**
 | `ProcessPendingChunks` | 301 | Drains ProcessQueue under per-frame budget; reads the per-request cancellation token before removing its map entry, drops obsolete/aborted results, then applies each live result via `ApplyTileResult`. |
 | `ApplyTileResult` | — | **Shared game-thread apply** for one `FChunkResult` (async drain + sync carve): discards stale generation epochs and keys no longer desired, marks loaded, ingests capture, EMPTY releases the tile's existing component (a re-gen can flip content→empty on band change — old geometry must not linger), else `ApplyMeshToTile`. Returns true iff a visible mesh uploaded (counts the budget). Doesn't touch `PendingTiles` (caller's). |
 | `FVoxelStackSampler` / `FScopedVoxelStackRegistration` (`VoxelStackSampler.h/.cpp`) | — | **Diagnostic stack-sampling profiler (2026-09-13), off by default.** A sampler thread captures registered generation workers' stacks (`CaptureThreadStackBackTrace`) every N µs and tags each with the tile's LOD. It symbolises after the run (inline frames too) into CSV plus a summary under `Saved/`. Registration is per tile TASK in `LoadTile` (and the export's tasks), never per sample; disabled costs one relaxed atomic load. Game `-voxel.SampleStacks=<us>`, commandlet `-samplestacks[=us]`. ~4-5% overhead at 1 ms, so trust large shares, not single-digit deltas. Report: `PERF-SAMPLED.md`. |
-| Room-cache performance (2026-09-13): `VF_FindPlayerFitPointForRoomMemoized` / `GPlayerFitMemo` (`VoxelCaveMorphology.cpp`), `voxel.TileCacheWindow` + `VoxelGenLOD::TileOriginVoxels/TileCellsPerAxis` (`VoxelGenerator.h`), `FChunkSDFSpatialIndex` | — | **Player-fit memo:** exact 97-word key, process-wide, `voxel.PlayerFitMemo`. **Tile cache window:** the room cache covers the requesting tile's footprint plus a fixed collect margin. It is set via `TGuardValue` in `GenerateMesh`/classify, and is window-invariant by construction (the margin does not grow with the tile). **Spatial index:** ordered XY buckets over cache primitives, exact rejection. The `static_assert` in `VoxelStrateTypes.h` ties `VF_STRATE_PARAM_FIELDS` to `FStrateGenerationParams` (83 fields + 8 B padding). |
+| Room-cache performance (2026-09-13): `VF_FindPlayerFitPointForRoomMemoized` / `GPlayerFitMemo` (`VoxelCaveMorphology.cpp`), `voxel.TileCacheWindow` + `VoxelGenLOD::TileOriginVoxels/TileCellsPerAxis` (`VoxelGenerator.h`), `FChunkSDFSpatialIndex` | — | **Player-fit memo:** exact 97-word key, process-wide, `voxel.PlayerFitMemo`. **Tile cache window:** the room cache covers the requesting tile's footprint plus a fixed collect margin. It is set via `TGuardValue` in `GenerateMesh`/classify, and is window-invariant by construction (the margin does not grow with the tile). **LOD-gated:** fused path at LOD ≥ `voxel.TileCacheWindowMinLOD` (3), op-stack at LOD ≥ `voxel.TileCacheWindowOpStackMinLOD` (4); lower LODs keep the 4-chunk worker window, because the tile window slowed LOD0 evaluation ~6%. Strates whose authored terrain ops exceed the proven 256-voxel extent (or are malformed/unloaded) fall back to the legacy window (`VF_ComputeTerrainOpEnvelope`). `voxel.SpatialIndex` is -1 auto (off at fused LOD0, on elsewhere), 0 or 1. **Spatial index:** ordered XY buckets over cache primitives, exact rejection. The `static_assert` in `VoxelStrateTypes.h` ties `VF_STRATE_PARAM_FIELDS` to `FStrateGenerationParams` (83 fields + 8 B padding). |
 | `GenerateTileResult` | — | **Shared worker-side gen** for one tile (async `LoadTile` task + sync `SyncRemeshTile`): [ClassifyTile (T1.d) + exact validation, only when `voxel.OuterClassifierMode=1`; **default 0 since 2026-09-13**] → `GenerateMesh`/`GenerateSheetMesh` → `BuildTileStreamSet`. Reads Generator/Mesher only → safe on a worker or the game thread; fills `FChunkResult`, no enqueue. |
 | `SyncRemeshTile` | — | **INSTANT DIG**: level-0 same-frame re-mesh on the game thread (`GenerateTileResult` + `ApplyTileResult` inline, Cells=CHUNK_SIZE/Step=1 + strate band, no capture). Used for the tile under the brush centre so a carve is visible THIS frame; one full-res gen on the game thread. |
 | `UpdateChunksAroundPosition` | 362 | Builds desired set, cancels pending keys that left it, sorts from the **true pawn position**, submits the occupied/support LOD0 tile and the next tile along the pawn heading as an absolute `BackgroundHigh` prefix, then loads/unloads and handles LOD changes. **Delta cull**: `BuildDesiredTiles` returns the LEAVERS (stamped `DesiredStamped` map, one sweep) — only those + `TransitionHold` are considered per crossing, not every loaded tile. `BuildDesiredTiles` also applies `RenderDistanceChunks`: the outermost shell widens to cover the distance — as level-MaxClipLevel MC tiles, or (F18 `bFarSheetRing`) as a SHEET ring at level MaxClipLevel+span (covered-check vs the MaxClipLevel box; `VF_OuterShell` shared with `IsTileInClipRange` so the cull sees the same horizon), dz pre-clamped to the vertical band. **§9.3 anchors:** also prunes dead `StreamingAnchors` + detects their chunk crossings → rebuilds the desired set when an anchor moves/(un)registers (`bAnchorsMoved`/`bForceDesiredRebuild`), same cadence as player movement. §8.10. |