wip(assembler): round S5 -- streaming gates + trace; coarse-first order kept opt-in (no win)

Walkable/enclosed/refined gates defined; JSONL trace joins request, worker start,
build/apply and discards. Coarse-first + dynamic near reserve behind
voxel.StreamingCoarseFirst / voxel.StreamingDynamicNearReserve (default 0 =
S4 order): matched runs showed no walkability or full-view win and slower
LOD0/1 refinement. Instruments settled: 3.1 s standalone reproduced (16.7 s was
another generation route); black early captures were clear-only frames.
Hashes equal S4; Core-off, M2, SurfaceWorld/FloatingIslands BuildHost pass.

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

CODEMAP.md

@@ -328,6 +328,9 @@ the four MPSC result queues `CriticalProcessQueue` / `EditedProcessQueue` / `Nea
 `ProcessQueue`, and `PendingTiles` (TSet) (VoxelWorld.h:908-912).
 **Async state:** `bShuttingDown`, `ActiveTaskCount` (atomics), `GenerationEpoch`, and the
 game-thread `PendingTileCancellation` map whose per-request atomic flags are read-only on workers.
+S5 assigns request and worker-start ordinals to each tile, preserving epoch, resident-at-request,
+priority, queue-wait and build timing in the opt-in startup trace. All four result queues remain MPSC;
+EndPlay still waits for workers and drains every queue.
 
 **⭐ 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` /
@@ -359,7 +362,7 @@ casse cette arithmetique *structurellement* — `BeginPlay` loggue une **Error**
 | `AdvanceHeadlessDigMotionTest` / `RecordHeadlessDigMotionFrame` | 3060 / 3217 | Explicit `-voxel.TestDigMotion=1` repro: optional near digs, then repeats underfoot digs and samples `Generator->GetDensityAt` immediately after each edit, on its first subsequent simulated frame, and on every following observation frame. `-voxel.TestDigMotionFieldOverride=1` is a measurement-only prototype that clears `CurrentFloor` and switches CMC to falling when the sampled post-edit density is air; it is inactive by default. The test counts false support, false air and downward velocity beyond one gravity step. `-voxel.TestDigMotionUnderfootDigCount` is clamped to 1–100; `-voxel.TestDigMotionUnderfootIntervalSeconds=0` performs per-frame rapid edits. The accompanying gap audit is scoped to the dig session. |
 | `MaybeFinishHeadlessStreamingTest` | — | The existing `voxel.TestExitSeconds` diagnostic emits `VoxelForgeStreamingExitDiag` at most every 30 elapsed seconds, including threshold and early-return test flags; no output when the timer is unconfigured. |
 | `UpdateVisualStalenessAudit` | 5034 | Test-only camera-frustum coverage audit: tracks up to 24 real pre-skirt isosurface points per visible component, then counts frames where no visible tile component covers a previously visible, still-desired point (excluding edited bounds). `voxel.TestVisualGapInjectEarlyHide=1` hides a valid component for three frames and must be caught. `voxel.TestGapAuditLegacyHandoff=1` restores old handoff culling for A/B baselines. |
-| `ProcessPendingChunks` | 7435 | Drains critical, edited, near, then ordinary result queues under the 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`. Collision profiling timestamps the result's arrival here to separate worker completion from game-thread apply wait. Step 1's `voxel.Step1PrequeueVisibleLOD0` holds completed results until the requested non-empty L0 count is ready, then releases them through this same drain and emits the `VFStep1UploadCorpusReleased` CSV event for UE's delayed CSV capture. |
+| `ProcessPendingChunks` | 7435 | Drains `CriticalProcessQueue`, `EditedProcessQueue`, `NearProcessQueue`, then `ProcessQueue` under the per-frame budget; reads the per-request cancellation token before removing its map entry, traces stale/obsolete/aborted discards with epoch and request ordinal, then applies each live result via `ApplyTileResult`. Collision profiling timestamps result arrival separately from game-thread apply wait. Step 1's `voxel.Step1PrequeueVisibleLOD0` still releases through this drain and emits `VFStep1UploadCorpusReleased`. |
 | `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 raw stack sampler, off by default.** A sampler thread captures registered generation workers' stacks (`CaptureThreadStackBackTrace`) every N µs and tags each with the tile's LOD. It writes bounded raw PCs to CSV plus a symbolization-free summary under `Saved/`; symbol resolution is deferred until after process exit so the measured process never initializes/calls DbgHelp for this diagnostic. Lifecycle teardown is serialized and registration TLS is session-tagged. 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]`. |
 | 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). |
@@ -367,7 +370,7 @@ casse cette arithmetique *structurellement* — `BeginPlay` loggue une **Error**
 | `RecordTileHash` / `WriteTileHashDump` | — | Headless `-voxel.TileHashDump=<absolute path>` parity instrument. Records the exact `FVoxelMeshData` used by the streaming result, including empty/classified tiles, then writes sorted absolute tile keys, LOD step/cells, content-band limits, settings digest context, and the shared `VoxelGeometryHash` at `EndPlay`. |
 | `SyncRemeshTile` | — | **INSTANT DIG**: level-0 same-frame re-mesh on the game thread. Uses `PreEditDensityCache` when the tile is resident in the bounded local cache, replays the current DiffLayer over that exact procedural lattice, then meshes/applies inline so a carve is visible THIS frame. A cache miss uses normal generation; it captures the unedited lattice if the tile needs meshing, while uniform-tile proofs remain enabled. |
 | `PreEditDensityCache` / `VoxelForgePreEditDensityCache::BuildEditedGrid` | — | Per-world cache of the pawn's current L0 tile: exact 35³ pre-edit samples, max **171,500 bytes**. It is pruned when the pawn crosses chunk boundaries and cleared on `RegenerateAllChunks`. `BuildEditedGrid` snapshots touched diff chunks once and reproduces `GetDensityAt`'s final `pre-edit - EvaluateMods` operation before the per-call shared-grid mesher path. |
-| `UpdateChunksAroundPosition` | 8603 | 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 (`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. |
+| `UpdateChunksAroundPosition` | 8603 | Shared Core/legacy scheduler. Builds the desired set and delta-culls leavers; admits the occupied/support tile plus collision-required LOD0 tiles in the pawn's 3×3 XY ring as `CollisionCritical`. S5 coarse-first admission is opt-in (`voxel.StreamingCoarseFirst=1`): it submits remaining coarse levels from max LOD down through `max(2, FullResClipLevels)`, then LOD0/1 refinement; ties favor travel direction, then distance. Default `0` preserves S4's LOD0-to-LOD8 order. `voxel.StreamingDynamicNearReserve=1` opts into holding the authored near-worker reserve while moving or until the stationary walkable ring is ready; default `0`. The clipmap stays non-overlapping and existing load-before-unload / `TransitionHold` reconciliation remains authoritative; S5 adds no overlapping placeholder layer. Standalone defaults to legacy; use `--coarse-first-stream-order` to opt in. §8.10 + §8.26. |
 | `AdvanceHeadlessCollisionGateStressTest` / `ObserveHeadlessCollisionGateStressTest` / `MaybeFinishHeadlessCollisionGateStressTest` / `MeasureHeadlessCollisionGateStressTerrain` | — | Explicit `-voxel.TestCollisionGateStress=1` harness only. The normal spawn-height cvar offsets from the measured density surface; `voxel.TestCollisionGateStressDeepDrop=1` carves a process-local box shaft from above the top to just above the floor tile, extending along the post-landing route so the pawn does not run into the shaft sidewall, temporarily raises only the transient test DiffLayer brush-radius allowance for that fixture, removes only the local pawn's redundant streaming anchor for this isolated repro, proves every sample in two complete outside-ring level-0 tiles is air before seeding their known-empty status, prewarms the known-empty in-ring column, and asserts that the support walk reaches the in-layout floor outside the aligned level-0 render ring before the pawn drops. The log reports whether the fixed gate retains that otherwise out-of-ring target. Other streaming anchors remain active. The fixture rejects a target or floor outside the layout, a non-air empty-tile proof, or a floor inside the aligned level-0 ring. The `GetDensityAt` column at capsule feet defines the local air/solid surface; the deep-drop column crossing is cached after its first exact scan. Each frame records actor position, velocity, direct foot tile and gate support readiness, gate state, signed feet clearance, and a ≤0.5-voxel sweep between positions. PASS requires no measured clearance below −5 cm; deep-drop gate-on also requires horizontal travel during a hold. Result logs include the terminal density column, first crossing below spawn height, hold count/time, and gated horizontal progress. |
 | `BuildDesiredTiles` | 6482 | Builds `DesiredSorted`+`DesiredStamped` (player clipmap shells), then adds streaming anchors and the pending player's collision-only support column before the leaver sweep. It orders from the true local pawn position and builds `CriticalDesiredTiles` for the unresolved support target, occupied/support and heading tiles. Also rebuilds `DesiredTransitionDescendants` (ancestor → desired descendants, including face/edge-touching replacements) for the reconcile pass. |
 | `CollisionGateDesiredTiles` / `RequestCollisionGateSupportTile` | 5062 | Retains an unresolved support target and up to three level-0 tiles below it across desired-set rebuilds. The additional keys are hidden collision-only tiles, queued ahead of ordinary streaming, and pruned once resolved or outside the active column. |
@@ -380,7 +383,8 @@ casse cette arithmetique *structurellement* — `BeginPlay` loggue une **Error**
 | `AddAnchorDesiredTiles` | — | **§9.3 multi-anchor:** folds each `FVoxelStreamingAnchor`'s thin level-0 box (`XYRadiusChunks`/`ZBelowChunks`/`ZAboveChunks`) into the SAME desired set and the independent collision set, 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. |
-| `LoadTile` | 9033 | 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`). `voxel.TestFullResClipLevels` is a measurement-only runtime override; default -1 uses the authored setting. 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). §8.10. |
+| `LoadTile` / `LoadCoreTile` | 9033 / 11633 | Budget check → background generation; RAII task guard. Both paths preserve `Epoch`, desired/modification epochs, cancellation state and MPSC result routing. Requests receive a monotonic request ordinal; workers assign a generation-start ordinal. `VoxelStartupTrace` joins those with priority, resident-at-request, queue wait, worker build, result wait, apply time and discard reason. Legacy `LoadTile` retains its classify/mesh and strate-band behavior; Core delegates to the resolved immutable snapshot. `EndPlay` still drains workers/results. §8.10. |
+| `VoxelStartupTrace::RecordRequest` / `RecordTile` | — | Opt-in JSONL via `-voxel.StartupTraceFile=...`; `-voxel.StartupTraceUntilEndPlay=1` keeps tracing through movement, otherwise it stops at initial settle. Each request is keyed by epoch and ordinal; applied tile records add worker-start order, priority, queue/build/apply timings. Discard events identify stale epoch, obsolete, no-longer-desired and aborted results. |
 | `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` | 9838 | 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. Ordinary same-tile re-cooks retain the installed body while it is valid. A carve intersecting the pawn's current support tile and capsule bounds suppresses that old body until the replacement cook is installed. `voxel.ProfileCollisionCookStages` reports generation/apply waits, combined RMC time, completion hops, triangle/skirt counts and collision settings. With `voxel.Step1PrequeueVisibleLOD0>0`, logs apply frame and waits for `GetLastRenderTime` to advance for first-visible frame; `voxel.Step1UploadDisableCollision=1` isolates render upload. Routes each terrain base through shared MIDs; disabled volume MIDs bind neutral `VolP*` and null `VolTex*`. **§9.4:** `SetVisibility(false)` when the tile is in `CollisionOnlyTiles` (anchor-only, hidden — collision still cooks). §8.10 + ARCHITECTURE SurfaceWorld row. |
 | `HandleRealtimeMeshCollisionBodyUpdated` / `HandleTileCollisionCookComplete` / `PollStep1SupportSweepProbes` | 6964 / 7192 / — | Correlate RMC body-install event and collision-future completion. `voxel.ProfileCollisionCookStages=1` logs worker queue/gen, result wait, apply/upload/config submission, RMC submit→body update, body event→future, future→game-thread hop, submit frames, pending VoxelForge cooks, priority, collision flags, and triangle/skirt totals; it also reports mesh-generation-end→ready. `voxel.Step1CollisionProbe=1` sweeps the pawn capsule against the exact component pointer, logs both the first hit and whether it is the target body, and checks every body in the captured support cohort. RMC's end-frame dequeue, async worker start, geometry extraction and core cook remain inside the external plugin and are not individually timestampable from this module. |