unreal: parallel sign-edge repair + cheap SW route prefilter (L94); SW startup 2.75 -> 0.43 s

- Unreal requests Core's deterministic parallel sign-edge repair (UE ParallelFor, capped at
  10 workers); headless defaults to parallel (--serial-sign-edge-repair for A/B). SW
  sign_edge_distances 1,942 -> 37 ms, ShapeStage 2,156 -> 198 ms.
- SW fallback route planner checks the exact mouth-radius endpoint constraints before sampling
  terrain (58,480 of 78,408 candidates skip heightfield work); same 35 eligible, same chosen route.
  SerializeResolvedWorld 468 -> 115 ms; BeginPlay block 0.58 -> 0.22 s.
- Unreal SW: first admission 2.75 -> 0.43 s, LOD0-2 settled 5.20 -> 2.81 s; FI 3.40 -> 3.12 s.
- World identical: assembly/recipe/plan hashes unchanged; serial vs parallel per-tile hashes equal.
  Fast tier PASS; Linux UE host builds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1

Standalone/Research/WorldAssemblerM3/Linux94/REPORT.md

@@ -0,0 +1,45 @@
+# Linux 94 — Unreal startup
+
+## Before and after
+
+Medians are from L92's final serial-repair `after-direct-json` traces and L94's parallel-repair traces. Times start at `BeginPlay`, except export duration and `core_compile_start` → `bReady`. LOD0–2 means every currently desired key was settled before the scripted movement at 20 s.
+
+| Arrival | Measure | Before | After | Runs |
+|---|---|---:|---:|---:|
+| SW | `SerializeResolvedWorld` | 468.0 ms | 115.2 ms | 3 / 4 |
+| SW | Core compile → `bReady` | 2.157 s | 0.198 s | 3 / 4 |
+| SW | `bReady` | 2.738 s | 0.421 s | 3 / 4 |
+| SW | First admission | 2.747 s | 0.434 s | 3 / 4 |
+| SW | LOD0–2 settled | 5.198 s | 2.813 s | 3 / 3 of 4 |
+| SW | `BeginPlay` game-thread block | 0.582 s | 0.224 s | 3 / 4 |
+| FI | `SerializeResolvedWorld` | 465.6 ms | 115.4 ms | 3 / 3 |
+| FI | Core compile → `bReady` | 0.184 s | 0.180 s | 3 / 3 |
+| FI | `bReady` | 0.750 s | 0.408 s | 3 / 3 |
+| FI | First admission | 0.762 s | 0.422 s | 3 / 3 |
+| FI | LOD0–2 settled | 3.396 s | 3.119 s | 3 / 3 |
+| FI | `BeginPlay` game-thread block | 0.570 s | 0.229 s | 3 / 3 |
+
+Export, compile, `bReady`, admission and residency values are direct trace measurements. The SW route candidate log window fell from a 451 ms median to 96 ms (timestamps rounded to 1 ms); the enclosing export timer fell by 353 ms. The route window is a boundary measurement, not an internal function timer.
+
+## Changes and hashes
+
+- Unreal now requests parallel sign-edge repair through its existing Unreal `ParallelFor` scheduler. The request remains capped at 10 workers on this 12-logical-core host; it creates no separate worker pool. Traces show compile launch queue delays of 0.04–0.18 ms. The first two seconds' sampled frame maximum was 28.7 ms in SW; FI had one 45.9 ms sample and two samples below 27.2 ms. The direct synchronous `BeginPlay` block is measured above.
+- The SW fallback route planner checks its exact mouth-radius endpoint constraints before allocating route points or sampling terrain. It still visits 78,408 candidates, but skips heightfield work for the 58,480 placement rejects. Eligible candidates remain 35; the selected route remains center `(32,80)`, direction 6, half-length 16, lift 128, with three sweep attempts.
+- Headless sign-edge repair now defaults to parallel, with `--serial-sign-edge-repair` retained for A/B runs. The passage composer was not changed.
+- Unreal `world_assembly_hash=1ecc162755ab83ac`, `recipe_hash=52b010081ad4b550`, and `plan_hash=a06d76fd11d7a08e` match L92 in every FI/SW run.
+- Unreal SW `sign_edge_distances` fell from 1,942.2 ms to 36.5 ms median; `ShapeStage` fell from 2,156.3 ms to 198.1 ms. FI stayed effectively level: 8.8 → 9.5 ms and 183.5 → 179.8 ms.
+- A serial/parallel Core hash A/B at the SW and FI spawn coordinates produced identical per-key tile hashes, full-detail hashes, whole-view hashes and per-LOD hashes. SW matched for 1,408 tiles (whole-view density/mesh/payload/topology: `ba04e6da1010c52f / 93555a2b969b16a5 / 339eba0fad4f79a5 / c2edafbec2fa9305`); FI matched for 1,216 tiles (`3175bceb00bfc76c / d1dabfba70a95d8a / b23900f97cfbb2cf / 3967a02b7c1ae1c8`). This focused headless A/B used the current owner-export manifest `bcc0d0d33ea39435`; the separate Unreal transient manifest is the `1ecc…` hash above.
+- In that headless A/B, SW `sign_edge_distances` was 1,753.9 ms serial and 86.4 ms parallel. The FI input had zero endpoint corrections in both modes.
+
+## Verification and failed runs
+
+- Fast tier: **PASS**, 274 checks (269 PASS, 5 diagnostics), 297.1 s; source stamp `03a8fce1a9067edb8a83461a836292686d74abd8ea84b8014d8006296da8428d`. Headless build and tier were pinned to CPUs 0–11 and capped at 6 GiB. Unreal Linux build succeeded with `-NoUBA -MaxParallelActions=4`, pinned to CPUs 0–11, without `--as`; it was not run during timing. The build emitted three existing RealtimeMesh deprecation warnings.
+- SW repeat 3 exited Unreal successfully but missed three desired LOD0–2 tiles at the 20 s stationary boundary. They settled after movement began, at 25.36 s, so this run is excluded from the LOD median; repeat 4 completed the three-run residency sample.
+- An auxiliary full `CoreWorldTest` without trace-only mode stopped before hashing: the host project's staged first-person Blueprint failed its `MaxWalkSpeed` pin preflight. The trace-only startup runs and the separate headless hash A/B completed.
+- Logs and traces are under `Saved/Linux/l94/`; baseline traces are under `Saved/Linux/l92/unreal-arrival/`. `RunL94UnrealArrival.py` reproduces an individual Unreal trace run. No run exceeded 30 minutes. No commit, push or stash was made.
+
+## Source files
+
+- `Source/VoxelForge/Private/VoxelWorld.cpp`
+- `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp`
+- `Standalone/Source/StreamBenchmarkHeadless.cpp`