perf(streaming): item 3 bought nothing, item 5 removed dead weight -- both worth knowing

Two WORK-NEXT items, one null result and one confirmed removal. Item 4 untouched.

ITEM 3 -- CANCEL OBSOLETE WORK AND PRIORITISE THE FLOOR: NULL RESULT
Clean -nullrhi, diagnostics off, 800 cm/s, 200 m of travel:

                          before            after
  request->ready p95      0.369117 s        0.367612 s
  queue wait p95          0.066949 s        0.064483 s
  generation p95          0.287926 s        0.283642 s
  p95 travel distance     2.952938 m        2.940894 m
  obsolete aborts                 0                12

Flat. Aborting obsolete tiles, sorting from the true pawn position and prioritising the
occupied/next tile changed request-to-ready by 0.4%.

⭐ AND THE NULL RESULT IS ITSELF THE FINDING. Look at the split: queue wait is 65 ms of a
368 ms p95; generation is 284 ms. **The queue was never the problem, so reordering it
cannot help.** Only 12 tiles in a 200 m run were obsolete enough to abort. The floor is
late because generation is slow, not because work is badly ordered -- item 3 establishes
that by exclusion, which is worth more than the 0.4%.

The machinery is kept: it is correct, it costs nothing, and it will matter once
generation is fast enough for ordering to be the binding constraint.

ITEM 5 -- VERDICT CACHE REMOVED. It was pure overhead.
Clean moving-session audit:

  outer  calls/probes/hits    2,109 / 2,109 / 0
  nested calls/probes/hits   19,904 / 19,904 / 0
  enclosing hits                              0
  stores 15,877, evictions 15,365

⇒ It stored 15,877 proofs, evicted 15,365 of them, and hit ZERO times. A moving player
visits unique regions; the cache was filling and thrashing a 512-entry table under a
global lock for no benefit. Same-binary A/B: request p95 0.367645 vs 0.365978,
generation p95 0.285559 vs 0.285992 -- neutral within noise, so its cost was the lock and
the linear scans rather than time we can see.

Removed: the global cache, its lock, the entry table, the invalidation path, the
cache-only profiler fields and the synthetic probe. The proof classifier stays.

⚠️ I asked for this cache two days ago on the strength of a commandlet measurement that
re-visited the same region repeatedly. That access pattern is nothing like a moving
player, and I should have said so when specifying it.

VERIFIED BY ME: build 0 errors; export 0.563 / 0.565 / 0.566 s (35de59b was 0.545-0.563,
flat as expected); geometry SHA-256 b3e5f4c3..9b377 UNCHANGED; capability 20,830 /
10,909, connected.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012aheo2crrZmHuYtPpdY2Cb

CODEMAP.md

@@ -47,7 +47,7 @@ A **density-field voxel terrain** plugin for Unreal Engine, built around undergr
                          Tick → UpdateChunksAroundPosition
                                    │  (load/unload around player, by distance + LOD)
                                    ▼
-                         LoadChunk → UE::Tasks::Launch  ──────────► background thread
+                         LoadTile → UE::Tasks::Launch  ──────────► background thread
                                                                          │
    UVoxelMarchingCubesMesher::GenerateMesh(chunk, step) ◄────────────────┘
         │ samples density per cell corner
@@ -265,12 +265,14 @@ redesign; tile identity lives in `FVoxelTileKey` (VoxelWorld.h).
 
 ### 3.5 World orchestrator — `Public/VoxelWorld.h` + `Private/VoxelWorld.cpp`
 `AVoxelWorld : AActor` — owns everything, drives streaming. Also `FChunkResult` struct
-(VoxelWorld.h:35) = async task payload (coord, chunk, meshdata, LOD, **Epoch**).
+(VoxelWorld.h:35) = async task payload (coord, chunk, meshdata, LOD, **Epoch**, request timing,
+desired-set epoch and obsolete/cancellation state).
 
 **Owned objects (UPROPERTY):** `Settings`, `Generator`, `Mesher`, `StrateManager`,
 `DiffLayer` (VoxelWorld.h:55-78). **Storage:** `Chunks` map, `ChunkMeshes` map,
 `ChunkLODs`, `ProcessQueue` (TQueue), `PendingChunkCoord` (TSet) (VoxelWorld.h:85-312).
-**Async state:** `bShuttingDown`, `ActiveTaskCount` (atomics), `GenerationEpoch`.
+**Async state:** `bShuttingDown`, `ActiveTaskCount` (atomics), `GenerationEpoch`, and the
+game-thread `PendingTileCancellation` map whose per-request atomic flags are read-only on workers.
 
 **⭐ Espace ACTEUR / ACTOR SPACE (2026-08-17).** Le champ de densite est defini en espace **acteur** :
 `(0,0)` est l origine de l acteur `AVoxelWorld`, pas celle du monde Unreal. `WorldToLocalCm` /
@@ -295,19 +297,19 @@ casse cette arithmetique *structurellement* — `BeginPlay` loggue une **Error**
 | `OnObjectModifiedInEditor` | 58 | Regenerates when a strate asset is edited (if `bLiveEditStrates`). |
 | `EndPlay` | 140 | Sets `bShuttingDown`, **waits for `ActiveTaskCount`→0**, unbinds delegate. |
 | `BeginPlay` | 177 | Constructs Generator/Mesher/StrateManager/DiffLayer, wires services, seeds. |
-| `Tick` | 220 | `UpdateChunksAroundPosition(player)` + `ProcessPendingChunks()`. |
+| `Tick` | 220 | `UpdateChunksAroundPosition(player, heading)` + `ProcessPendingChunks()`; the clean LOD0 latency summary is flushed at `EndPlay`. |
 | `GetPlayerPosition` | 231 | Pawn position or zero. (`GetLODForChunk`/`LODToStep`/`IsChunkInRange` removed — dead since the clipmap.) |
-| `ProcessPendingChunks` | 301 | Drains ProcessQueue under per-frame budget; applies each via `ApplyTileResult`. |
-| `ApplyTileResult` | — | **Shared game-thread apply** for one `FChunkResult` (async drain + sync carve): discards stale epochs, 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). |
+| `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). |
 | `GenerateTileResult` | — | **Shared worker-side gen** for one tile (async `LoadTile` task + sync `SyncRemeshTile`): ClassifyTile (T1.d) → `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, sorts by distance, loads/unloads, 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. |
-| `BuildDesiredTiles` | ~722 | Builds `DesiredSorted`+`DesiredStamped` (player clipmap shells + F18 sheet ring) then `AddAnchorDesiredTiles()` before the leaver sweep. |
+| `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. |
+| `BuildDesiredTiles` | ~722 | Builds `DesiredSorted`+`DesiredStamped` (player clipmap shells + F18 sheet ring) then `AddAnchorDesiredTiles()` before the leaver sweep; it orders from the true local pawn position and builds `CriticalDesiredTiles` for the occupied/support and heading tiles. |
 | `AddAnchorDesiredTiles` | — | **§9.3 multi-anchor:** folds each `FVoxelStreamingAnchor`'s thin level-0 box (`XYRadiusChunks`/`ZBelowChunks`/`ZAboveChunks`) into the SAME desired set (deduped by stamp) so AI/remote players keep collision loaded around them; delta cull releases them on move/unregister. **§9.4:** a tile only a CollisionOnly anchor wants (clipmap didn't stamp it) → `CollisionOnlyTiles` → hidden at apply. Zero cost when no anchors. |
 | `ReconcileAnchorTileVisibility` | — | **§9.4:** after each rebuild, toggle `SetVisibility` on ALREADY-LOADED tiles that flipped render↔collision-only (diff `CollisionOnlyTiles` vs prev — bounded, no O(loaded) scan; unhide only if still desired). No-op without CollisionOnly anchors. |
 | `RegisterStreamingAnchor` / `UnregisterStreamingAnchor` | — | **BlueprintCallable §9.3:** add/remove an actor as a streaming anchor (`EVoxelAnchorPolicy` CollisionOnly/FullVisual + XY/ZBelow/ZAbove box). Idempotent; forces a rebuild next Tick. §9.4: CollisionOnly tiles cook collision but are hidden (no draw/VSM) unless the player clipmap wants them too. |
-| `LoadChunk` | 445 | Budget check → `UE::Tasks::Launch` background gen+mesh; RAII task guard. Worker runs `Generator->ClassifyTile` first (T1.d): AllSolid/AllAir ⇒ skip `GenerateMesh`, tile stays empty (capture tiles always generate). STRATE CONTENT CUT: tiles ≥ `StrateContentCutMinLevel` pass the player-strate band (`MeshBandChunkLo/Hi` → voxels) to `GenerateMesh` + stamp it on `FChunkResult::BandChunkLo/Hi`; band change re-queues via `BandRemeshQueue` (see `UpdateChunksAroundPosition`). TOO-COARSE SKIP: if one cell is taller than the band (`Step > band height` — level ≥7 territory) the tile is enqueued EMPTY without launching a task (cell-granular cut could only render garbage). F18: `Tile.Level > MaxClipLevel` = SHEET tile → routed to `GenerateSheetMesh` (band mid-chunk = strate ref; band unarmed ⇒ empty; MC-ring sampling density, cells capped 128/axis; carries the XY hole `SheetHole*Vox` — hole moves ⇒ overlapping sheets re-queue via `BandRemeshQueue`, see `UpdateChunksAroundPosition`). §8.10. |
-| `UnloadTile` | — | Clears tile state; the component is PARKED in the pool (T2.c), not destroyed. |
+| `LoadTile` | 445 | Budget check → `UE::Tasks::Launch` background gen+mesh; RAII task guard. Worker runs `Generator->ClassifyTile` first (T1.d): AllSolid/AllAir ⇒ skip `GenerateMesh`, tile stays empty (capture tiles always generate). It carries `GenerationEpoch`/desired epoch and checks `bShuttingDown`, generation pause, and its read-only cancellation token between expensive stages; obsolete work returns through `ProcessQueue`. STRATE CONTENT CUT: tiles ≥ `StrateContentCutMinLevel` pass the player-strate band (`MeshBandChunkLo/Hi` → voxels) to `GenerateMesh` + stamp it on `FChunkResult::BandChunkLo/Hi`; band change re-queues via `BandRemeshQueue` (see `UpdateChunksAroundPosition`). TOO-COARSE SKIP: if one cell is taller than the band (`Step > band height` — level ≥7 territory) the tile is enqueued EMPTY without launching a task (cell-granular cut could only render garbage). F18: `Tile.Level > MaxClipLevel` = SHEET tile → routed to `GenerateSheetMesh` (band mid-chunk = strate ref; band unarmed ⇒ empty; MC-ring sampling density, cells capped 128/axis; carries the XY hole `SheetHole*Vox` — hole moves ⇒ overlapping sheets re-queue via `BandRemeshQueue`, see `UpdateChunksAroundPosition`). §8.10. |
+| `UnloadTile` | — | Clears tile state; marks any matching generation request obsolete without removing its pending slot, so a late result cannot clear a replacement request; the component is PARKED in the pool (T2.c), not destroyed. |
 | `ApplyMeshToTile` | — | Upload geometry. One component per tile (clipmap keeps count low; supersedes the old region batching); worker-built streams (T1.f) → `CreateSectionGroup(MoveTemp)`. Reuses the component's existing `URealtimeMesh` (no per-apply mesh alloc). **F17: two polygroups** (0 ground / 1 sky-cap, per-triangle class from the mesher) → RMC auto-section per non-empty group; slot 0 = override/default material, slot 1 = `CeilingMaterial` (fallback ground); config gated by `FChunkResult::bHasGroundTris/bHasCeilingTris`. Takes `FChunkResult&`; strate lookups clamp Z into `Result.BandChunkLo/Hi` (strate content cut) — ground at clamped bottom chunk, cap at clamped top (mid = gap fallback). Collision level-0 only (T1.c) both groups; shadow per SECTION: ground casts at level≤1, cap never. **§9.4:** `SetVisibility(false)` when the tile is in `CollisionOnlyTiles` (anchor-only, hidden — collision still cooks). §8.10 + ARCHITECTURE SurfaceWorld row. |
 | `AcquireTileComponent` / `ReleaseTileComponent` | — | **T2.c component pool** (`TileComponentPool`, bounded): park on unload (geometry+collision stripped, hidden, stays registered), pop on apply — no `NewObject`/`RegisterComponent`/GC churn during travel & regen. §8.10. |
 | `GetStrateAtPosition` | 965 | Gameplay query → strate index. |
@@ -768,7 +770,7 @@ Stage order (negative=solid throughout). Each stage's anchor:
 | View distance / task budget / LOD distances | `VoxelSettings.h` (no recompile of logic — data asset). |
 | LOD step mapping | `AVoxelWorld::LODToStep` VoxelWorld.cpp:268; `GetLODForChunk` :242. |
 | How chunks stream in/out | `UpdateChunksAroundPosition` VoxelWorld.cpp:362. |
-| Async threading / stale-result handling | `LoadChunk` :445, `ProcessPendingChunks` :301, Epoch logic. |
+| Async threading / stale-result handling | `LoadTile` :445, `ProcessPendingChunks` :301, generation Epoch + per-request cancellation token. |
 | Add a new cave feature / terrain op | Add enum in `VoxelTerrainOpDefinition.h:36`, params there, `ApplyTo` (.cpp:6), transport fields in `FStrateGenerationParams`, consume it in a new Step inside `GetDensityWithParams`. |
 | Tweak room/tunnel shapes | `VoxelCaveMorphology.cpp` `BuildChunkCache` :47 / `EvaluateSDFCached` :757. |
 | Worm tunnel behavior | `GetDensityWithParams` Step 5, VoxelGenerator.cpp:1241. |