VoxelForge/Standalone/Research/WorldAssemblerM3/Round10/REPORT.md

World assembler S6 / M3 Round 10 — in-game cold-view time

Result: the 251 ms staircase was trace sampling cadence, not a scheduler tick. On the staged 2,578-key SurfaceWorld actor, the default legacy path reached full residency in 17.412 s. Gated coarse-first admission reached it in 11.704 s while the walkable ring time stayed at 0.627 s versus 0.632 s. A 12-mesh / 14 ms startup apply budget did not improve the legacy timeline. The remaining floor is on worker-side work supply/admission, not the game-thread apply budget; the trace does not isolate why available worker slots stay idle. The 1.57 s aim remains unmet. This is null-RHI residency evidence, not rendered enclosure proof.

1. Cadence and comparison with S5

AVoxelWorld::UpdateChunksAroundPosition emits streaming_sample only when at least 0.25 s has elapsed since its previous sample. S5’s near-251 ms per-LOD staircase was quantized by that observer; it was not a one-LOD-per-tick scheduler throttle. S6 records direct tile-result/apply timestamps and per-frame ProcessPendingChunks samples, so the completion table below does not depend on that sampler.

S5 reported 3.352–6.367 s for its per-LOD residency samples. The Round 9 report did not retain the exact BuildHost command line, executable hash, or raw direct-tile trace for that run. S6 uses the same staged DA_Settings SurfaceWorld plan, 25 cm Core scale, 2,578 desired keys, 12-task host setup, and legacy-order controls, but its direct full-set result is 17.412 s. The 250 ms sampling interval cannot explain this larger difference. Treat S5’s 3.35–6.37 s values as a different/unidentified measurement route, not a matched baseline; S6 preserves the direct command lines and hashes below. The Standalone 3.58–3.64 s S5 measurement was CPU-only over 2,553 keys, so it is also not a matched BuildHost apply/collision measurement.

2. Matched BuildHost actor runs

Fresh sequential processes used the staged owner plan, CoreVoxelSizeCm=25, 2,578 desired tiles, -corelimit=12 (10 active task slots in the actor trace), null RHI, a manually ticked UWorld with a 16 ms step, and a +Y 600 cm/s move beginning at 30 s. There was no overlapping Unreal process or build. Full-set time is the actor’s initial desired-set milestone. “LOD0 full” means all 512 visual LOD0 keys, separate from the 9/9 collision walkable gate.

Configuration	Full set ready	Walkable ring 9/9	Full 512-key LOD0	Count-cap frames / backlog frames	Visible meshes/frame p95/max	Apply CPU p95/max	UWorld::Tick spawn p95/max	UWorld::Tick walk p95/max
Legacy order, authored 4 / 2 ms budget	17.412 s	0.632 s	2.252 s	78 / 22	4 / 4	0.77 / 1.62 ms	1.406 / 5.059 ms	1.771 / 29.247 ms
Legacy order, startup boost on (12 / 14 ms)	17.346 s	0.611 s	2.190 s	0 / 0	4 / 6	0.70 / 1.49 ms	1.301 / 5.429 ms	1.798 / 28.489 ms
Coarse-first after ring gate, authored 4 / 2 ms	11.704 s	0.627 s	10.201 s	118 / 48	4 / 4	0.82 / 1.44 ms	1.396 / 5.027 ms	1.911 / 28.276 ms

Coarse-first shortened the full desired-set tail by 5.708 s (32.8%) without a walkable-gate regression. It did not make all near detail ready sooner: full LOD0 moved from 2.252 s to 10.201 s. The option therefore stays opt-in pending a rendered view/enclosure check. The apply boost removed cap frames but changed full-set time by only 66 ms (0.4%), so it also stays opt-in. Normal play continues to use the authored 4 visible-mesh / 2 ms limit. These are UWorld::Tick CPU samples in a manual null-RHI harness, not full rendered frame times. Walk p95 stayed below 2 ms, but the observed ~28–29 ms maxima remain and were not isolated to a single subsystem. Per-frame apply maxima were 1.62 ms legacy and 1.44 ms gated coarse-first.

Per-LOD cold set

Each pair is legacy → gated coarse-first, from direct tile completion events up to the initial desired-set milestone. Queue values are cumulative tile-seconds, not elapsed wall seconds. The initial set contains 2,578 tile results; empty results are included in the result count and apply time.

LOD	Last tile apply (s)	Worker task (s)	Worker queue (tile-s)	Result queue (tile-s)	GT apply (ms) / results
0	2.252 → 10.201	3.39 → 3.39	0.028 → 0.033	5.94 → 5.87	27.9 → 35.0 / 512
1	4.785 → 11.765	3.21 → 3.05	0.007 → 0.006	4.43 → 4.79	17.0 → 20.4 / 448
2	6.609 → 8.567	0.70 → 0.71	0.012 → 0.006	4.77 → 4.90	10.9 → 13.0 / 320
3	8.162 → 7.460	1.81 → 1.78	0.005 → 0.005	2.86 → 3.07	10.5 → 11.8 / 272
4	10.317 → 6.530	6.02 → 6.08	0.004 → 0.006	2.16 → 2.62	26.5 → 24.9 / 272
5	11.494 → 5.182	1.42 → 1.42	0.003 → 0.003	2.20 → 2.40	18.4 → 15.9 / 208
6	12.483 → 4.438	1.85 → 1.90	0.002 → 0.002	1.58 → 1.65	15.2 → 12.8 / 144
7	13.198 → 3.852	1.74 → 1.75	0.001 → 0.001	0.88 → 1.02	16.7 → 13.5 / 96
8	17.476 → 3.432	17.50 → 19.29	0.005 → 0.465	2.59 → 2.62	95.0 → 87.0 / 306

Per-frame visible mesh counts and the largest per-LOD apply sample were:

LOD	Visible meshes/frame p95/max, legacy → coarse	Max LOD apply/frame, legacy → coarse
0	2 / 4 → 3 / 4	0.762 → 0.934 ms
1	1 / 4 → 1 / 4	0.630 → 0.809 ms
2	0 / 4 → 1 / 4	0.533 → 0.639 ms
3	0 / 4 → 1 / 4	0.515 → 0.570 ms
4	2 / 4 → 3 / 4	0.605 → 1.354 ms
5	1 / 3 → 2 / 4	1.272 → 0.742 ms
6	0 / 4 → 1 / 4	0.687 → 0.708 ms
7	0 / 4 → 0 / 4	0.974 → 0.732 ms
8	2 / 4 → 3 / 4	1.625 → 1.439 ms

The 10-task-slot slack estimate was 132.1 slot-seconds on legacy order and 73.3 on coarse-first, computed as ten slots × (first request to full-set ready) minus measured worker-task seconds. The shared pool averaged 3.27 / 4.21 running workers on frames with estimated pending results; one legacy backlog frame had no running worker, versus none in the coarse-first run. Idle worker time and apply-cap frames are shared-pool/world properties, so a per-LOD split would double-count them. Legacy used 37.64 worker-seconds and coarse-first 39.36; pre-start worker-queue wait was only 0.067 / 0.527 tile-seconds. Result-queue wait summed to 27.41 / 28.94 tile-seconds. Game-thread tile-apply time was about 0.24 / 0.23 seconds for the cold set. The boost run had no apply-cap or result-backlog frames and still took 17.346 s. Thus the apply cap is not the end-to-end floor. The remaining floor is worker-side work supply/admission; the trace does not isolate why available worker slots stay idle or identify a safe single worker-side change for this round.

Collision is currently required only on the LOD0 path. The trace captured 60 body-update events in each cold run; 60 of 125 collision-desired entries were ready by the full-set milestone, while non-required/empty entries were classified separately. RealtimeMesh submit-to-body-update latency (not isolated Chaos CPU cooking) was legacy p50/p95/max 17.02 / 17.89 / 18.04 ms, boost 16.80 / 17.51 / 17.70 ms, and coarse-first 17.08 / 18.01 / 18.30 ms. The nine required ring tiles were ready at the gate times above. LOD1–8 had no collision-ready events in this trace.

3. Shared change and standalone crossing instrument

voxel.StreamingCoarseFirst=1 now starts far-level admissions only after all nine required walkable-ring collision gates are ready. voxel.StreamingDynamicNearReserve=1 keeps the authored near reserve while moving or before that gate; after a stationary gate it lets coarse work use the reserved slots. Both are explicit, default-off controls. voxel.StreamingStartupApplyBoost=1 allows 12 visible-mesh applies / 14 ms during the initial desired-set and walkable-ring phase, with a 30 s ceiling; zero keeps the authored budget. It is not preemptive within one apply, so it remains opt-in. This behavior is in shared AVoxelWorld::UpdateChunksAroundPosition / ProcessPendingChunks and does not change density, tile output, or collision policy.

The standalone 60 m/s flight instrument now counts forward +Y chunk crossings using the actual center delta, including multi-chunk jumps. The valid viewer route completed 250/250 center samples, with 226 center updates and max forward delta +2. It reports crossing_deltas=forward-+Y, resident 2,553/2,553 and no missing tiles. Latency still fails: nearest-ready p50/p95/max was 9.087 / 20.594 / 128.785 ms and maximum crossing wall time 42.931 s. The earlier --stream-benchmark and --stream-pipeline invocations emitted no Flight markers; only the viewer route is used for this result.

4. Hashes, regression checks and async audit

Standalone W1/W8/W16 fresh processes produced identical full-world and all nine per-LOD hashes; all match S5. Full world identity: density c10697a8f48380c7, mesh 430d2be2bdd8b988, payload b7b33e14092225c9, topology c5c46d66606a819e.

• Core-off legacy at 12.5 cm: PASS, VOXELFORGE_CORE_OFF_LEGACY_PASS.
• M2 V1 and V2: PASS, five fixtures plus same-LOD border for each.
• SurfaceWorld and FloatingIslands BuildHost at 25 cm: PASS. FloatingIslands W1/W8/W16 matched Standalone; the final Standalone W1/W8/W16 matrix also matched across all nine LODs and S5.
• S4 cap census/agreement and ray coverage: PASS; 551 requested/emitted cap bands, 1,584/1,584 coverage rays, 128-point cap agreement max 0.0504 cm against 2 cm tolerance, negative control detected.
• S3 ground and S4 LOD7/8 ground agreement: PASS; max ground delta 0.0427 cm, cross-LOD 0.0 cm, and the negative control was detected.
• Water remains fixed for the season; player carve/fill stays unrestricted.

Diff audit: GenerationEpoch is carried and stale epochs are rejected; the four result queues remain EQueueMode::Mpsc; EndPlay sets shutdown, waits for active work and drains the queues; workers check bShuttingDown; ModsLock behavior is unchanged; all UObject/RealtimeMesh component creation and mutation stays on the game thread. No density-sign or determinism path changed. The added C++ comments are French and English. No generated header, Binaries/, or Intermediate/ source edits.

5. Exact commands, binaries and run ledger

BuildHost actor runs used E:\Program Files\Epic Games\UE_5.7\Engine\Binaries\Win64\UnrealEditor-Cmd.exe with -nullrhi -unattended -corelimit=12 -run=VoxelForgeExplore -seed=0 -archetype=TunnelNetwork -coreworldtest=1 -corevoxelsizecm=25 -settingsref=/Game/VoxelForge/DA_Settings.DA_Settings -coreworldstreamtrace=1 -voxel.TestMove=1 -voxel.TestMoveSpeedCmPerSecond=600 -voxel.TestMoveStartSeconds=30 -voxel.TestMoveDirection=0,1,0 -voxel.StartupTraceUntilEndPlay=1 -voxel.CollisionGate=0. The exact expanded LogInit: Command Line: including project, DDC, trace and output paths is preserved verbatim in each legacy run, boost run, and gated coarse run. Their CVar overrides were respectively:

Run	StreamingCoarseFirst	StreamingDynamicNearReserve	StreamingStartupApplyBoost
Legacy negative control	0	0	0
Apply-budget A/B	0	0	1
Gated coarse-first	1	1	0

Standalone 60 m/s exact command (also in commands.txt):

Standalone\bin\VoxelForgeCoreStandalone.exe --world-assembler-prototype --force-generate --workers=12 --clip-radius=3 --max-clip-level=8 --camera=0,-35,-487.94,0,-22 --test-flight-crossings=250 --flight-speed-mps=60 --process-priority=Normal --process-affinity=4095 --frame-limit=60

Executable / module	SHA-256
BuildHost UnrealEditor-Cmd.exe	ECCE9996191F7F7E69B3D104EECDD0E678AE69176FDCCD79828C3E75A0654D6D
BuildHost UnrealEditor-VoxelForge.dll	1C083446ED360FBD798EAE4C4D3CA9E5994E906770B5017A51418BFEF46C30FA
BuildHost UnrealEditor-VoxelForgeCore.dll	54259552340FE81EA05640C622E9E5E5C0AC68C95F87DC718326D25EF133B2AC
Standalone VoxelForgeCoreStandalone.exe	1F445760AFDE264E2F28F8F2FB4CB68DBD0C8988AACD5A493B065E74C1827529

BuildHost UBT passed twice at 12 actions / Normal priority. & .\Standalone\Build.ps1 -Target All -CompileWorkers 12 -BuildPriority Normal passed. Existing vswhere/C4838 warnings did not fail the Standalone build. The S6 hash and 60 m/s logs are in Saved/BuildHost/S6/standalone.

Failed/retried attempts, including setup failures:

• An earlier full actor attempt reached its 600 s timeout in the Standalone-reference preamble; it left a partial trace. A first trace-only command omitted its required startup-trace output path and exited before the actor run; the helper now makes StreamingTraceOnly imply trace output.
• Two fresh-shell memory preflights lacked the local Add-Type helper and stopped before launching Unreal. The corrected GlobalMemoryStatusEx preflight ran before later launches.
• Early coarse-first attempts did not parse the requested CVar in headless BeginPlay; trace values showed legacy LOD0-first order. Explicit CVar parsing fixed the A/B. An ungated coarse-first run then delayed the walkable gate to 1.653 s; admission was changed to wait for 9/9 and rerun.
• FloatingIslands at 25 cm hit its 120 s test timeout; the bounded 600 s retry passed in 192 s.
• The first --stream-benchmark and --stream-pipeline sprint commands emitted no Flight markers. The corrected viewer command above produced valid crossing deltas, with the latency failure reported rather than hidden.
• Unsupported -RunName, a fresh-shell type reuse, and one PowerShell metric-parser expression failed before/after measurement without changing an Unreal run. The corrected command/parser produced the data above.

Every Unreal launch was sequential, with more than 17 GiB free memory at the final launch checks; actor-harness peak was about 3.5 GiB and all measured processes stayed under 8 GiB. No run returned 0xC0000142. Crash audit found no new UECC folder; the only folder is the pre-existing UECC-Windows-F1A3EFD5469F5DD62E68EBBACAC84376_0000 dated 2026-09-17. Test stdout and JSONL traces remain under Saved/BuildHost/...; none were copied into Research/.

6. Files and compile spots

No new C++ source file was added. Modified source files and likely compile spots:

• Source/VoxelForge/Private/VoxelWorld.cpp, Source/VoxelForge/Public/VoxelWorld.h, Source/VoxelForge/Private/VoxelStartupTrace.cpp, Source/VoxelForge/Public/VoxelStartupTrace.h — VoxelForge runtime module; BuildHost UBT.
• Source/VoxelForgeEditor/Private/VoxelForgeExploreCommandlet.cpp — editor module; BuildHost UBT.
• Standalone/Source/StandaloneViewer.cpp — Standalone/Build.ps1 -Target All.
• Tools/TestVoxelForgeCoreWorldHost.ps1 — harness argument/default changes; no C++ compile.
• CODEMAP.md §3, ARCHITECTURE.md §8.27, this report, and Standalone/Research/OwnerShapeWorld/gpu_field_port/LUNA_DIALOGUE.md — documentation.

No owner file or asset was touched; no commit, push or stash was made.