VoxelForge/Standalone/Research/WorldAssemblerM3/Linux90/R41-unreal-gap-notes.md

## Ranked causes for L85

1. **The first tile request starts about 0.7 s later than headless.** [TRACE] The Unreal clock starts at `BeginPlay` (`0.000 s`), which ends at `0.598 s`. Desired-set construction takes only 3.7 ms at `0.789–0.793 s`; the first tile request is at `0.793 s`. The scheduler still reports `core_not_ready` at `0.772 s`. [CODE] `BeginPlay` synchronously exports and parses the Core snapshot, then launches shape compilation on a `BackgroundNormal` task; `UpdateChunksAroundPosition` returns while Core is not ready. ([trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:18), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:48), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:52), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:7646), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:7782), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelWorld.cpp:11782).)

   [CODE] Strata/passage work is measured at roughly 76 ms. Snapshot serialization and `FromJson` run synchronously on the `BeginPlay` caller; shape compilation runs asynchronously and overlaps the rest of `BeginPlay`. The trace does not split the remaining `BeginPlay` time. Package or asset loading before `BeginPlay` is outside this trace; the recorded settings asset is transient. [INFERENCE] Headless starts its desired request around 60–80 ms from cold start, while Unreal’s first tile request is at 793 ms. That accounts for roughly 0.7 s of the gap as later pre-request work or a different clock boundary, but does not identify which part is avoidable. ([Linux87 report]( /home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Research/WorldAssemblerM3/Linux87/REPORT.md:9).)

2. **The remaining difference is in Unreal’s post-request chain, especially the staged LOD2 tail.** [TRACE] Unreal’s LOD0 ring is resident at `2.391 s`; at `3.134 s`, LOD0 and LOD1 are loaded but 240 LOD2 tiles remain missing. All 1,216 LOD0–2 tiles are resident at `3.650 s`. That is 2.86 s from first request to full residency, versus about 2.2 s for the headless run. ([trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:1455), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:2438), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:3135), [Linux85 report]( /home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Research/WorldAssemblerM3/Linux85/REPORT.md:21).)

   [CODE] Admission is frame-paced: the actor calls `UpdateChunksAroundPosition` once per tick. The trace shows a task cap of 8, ready queue depth 8, and total admission cap 16. At `3.134 s` all 16 slots are occupied and all 8 workers are running. So the cap is active, but this sample does not show idle workers waiting for admission. [INFERENCE] Frame pacing and LOD2 staging plausibly explain some of the post-request excess; the trace does not isolate their exact shares. ([VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:7074), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:13449), [Linux87 report]( /home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Research/WorldAssemblerM3/Linux87/REPORT.md:26).)

3. **The game-thread apply cap may add some tail latency, but it is not the main measured cost.** [TRACE] The startup boost ends at `3.157 s` because the desired epoch changed; that frame reports the normal limit of four mesh applies and hits the count limit. [CODE] Any desired-epoch change ends the boost. [INFERENCE] Keeping the boost active through the initial epoch churn is a small candidate to test, but the trace does not show that it accounts for a large fraction of 1.4 s. ([trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:2458), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:2466), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:9505).)

4. **Worker count, publication, and collision cooking are weak explanations for this gap.** [TRACE] `streaming_background_worker_cap=1` yields 8 effective tile slots, matching the 8 TaskGraph background threads; the setting is a clamp toggle, not a one-worker limit. Headless FI readiness changes by only about 36 ms between W8 and W12 in L87 (124 ms in the L85 matrix), far below this gap. ([trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:5), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:11592), [Linux87 report]( /home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Research/WorldAssemblerM3/Linux87/REPORT.md:38).)

   [TRACE] Unreal’s per-tile summaries show small apply times for LOD0–2; they cover the full trace, so they are indicative rather than an isolated startup breakdown. [CODE] The tile event records worker completion and game-thread application; application does not mean collision-ready. L85 measures residency through `LoadedTiles`, and its later milestone shows collision readiness is still a separate, incomplete count. RMC cooking therefore cannot explain the 3.650 s residency timestamp. ([trace summary]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:12896), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:9908), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l85/unreal-fi/fi-arrival-l85.trace.jsonl:5843).)

## L80’s four missing tiles

[TRACE] The four are LOD0 `(47,-2,-99)`, `(47,-2,-100)`, `(47,-2,-101)`, and `(47,-2,-102)`. At `45.375 s`, desired LOD0 is 516, resident is 512, pending is zero, and all 516 are visually covered. The trace repeatedly records empty results for those keys and then requests them again. ([trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l80/movement/unreal/fast-L80.trace.jsonl:21823), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l80/movement/unreal/fast-L80.trace.jsonl:21828), [trace sample]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l80/movement/unreal/fast-L80.trace.jsonl:21843).)

[INFERENCE] They most likely form the pawn collision gate’s four-tile downward support prefetch: the code requests the support tile and three below it. Empty results are marked both loaded and `CollisionNotRequired`; a later desired-set rebuild removes no-collision gate keys, while unload clears both that marker and loaded bookkeeping. The next gate check can then add the same keys again. This fits the repeated four-tile region compiles at the same center, though the trace does not record per-key collision-only membership or unload reasons. ([VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:8277), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:10653), [VoxelWorld.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:14330), [trace]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l80/movement/unreal/fast-L80.trace.jsonl:21841).)

This is **not a stranded worker task**: the empty tile work completes, and the commandlet ends with pending, cooks, and active tasks all zero. There is also a **readiness-check mismatch**: the commandlet requires every desired key to be in `LoadedTiles`, while its collision-set check accepts `CollisionNotRequiredTiles`. The short resident count is real at that instant, but the coverage count says it is not a visual hole. ([ExploreCommandlet.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForgeEditor/Private/VoxelForgeExploreCommandlet.cpp:18132), [ExploreCommandlet.cpp]( /home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForgeEditor/Private/VoxelForgeExploreCommandlet.cpp:18142), [L80 log]( /home/jahni/VoxelM/Plugins/VoxelForge/Saved/Linux/l80/movement/unreal/fast-L80.log:3542).)

## Cheapest confirmation and smallest fixes

- **For the L85 timing:** add trace timestamps around snapshot serialization, `FromJson`, compile launch, compile worker start, and `bReady` publication. The existing trace already records first request, tile worker times, apply, and residency. One matched run would show whether the unexplained pre-request interval is snapshot work or shape compilation.
- **Smallest latency change to A/B:** keep the startup apply boost through initial desired-epoch changes, ending it on initial readiness or timeout. The trace shows the current early reset and a count-limited apply frame. It may shorten the LOD2 tail; it will not explain or remove the pre-request delay.
- **Smallest L80 logic fix:** retain known-empty collision classification across unloads for the current modification epoch, invalidating it when tile data changes. That should stop re-requesting those support keys; confirm their collision-only role with per-key logging before changing the readiness predicate.

No files were changed; no builds, tests, or project binaries were run.