Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux64/REPORT.md
T
Fr0zkaandClaude Opus 5.5 e157847509 research(L64, R33): regression bisect -- Core +16% worker time L14->L61, not 2x; four-strate plan baseline
Per-round bisect of a176d07..3202193 (56 builds; timings inflated by concurrent UE builds) plus a clean
quiet-machine endpoint re-measure: L14 12.62 vs L61 14.65 worker-s (+16%), LOD0 3.99 vs 4.99 ms/tile. The
owner's Windows x2 is therefore mostly on the Unreal side or not like-for-like. Four-strate U2 plan: all
strates generate and are deterministic; passage 0-1 lower mouth is out of FI's band, 0/3 interfaces pass,
FI/TN/VS use legacy evaluators. R33: ranked Unreal-side suspects and Windows experiments.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-26 18:38:27 +01:00

21 KiB
Raw Blame History

L64 Linux baseline

Empty top-strate result

SurfaceWorld is not empty. U2’s gate report says its sampled top window had 0 triangles, but does not give the camera/window coordinates. With the current owner file (SHA-256 f42f96adca58…), the correct SurfaceWorld windows mesh on Clang: the file viewer_start (0, 0, -496 m) gives 9/27 nonempty LOD0 tiles and 36,864 triangles; the resolved route arrival (38.75, 70.75, -374.121 m) gives 9/27 and 25,672; the departure mouth (-6.75, 25.25, -376.307 m) gives 11/27 and 33,582. The binding is SurfaceWorld-native; the 57 boundary probes had 0 dispatch disagreements. This points to the gate’s sample window placement, not an empty field or wrong generator binding. The precise U2 sample point cannot be confirmed from its report.

There is a separate owner-plan defect: strict layout validation rejects passage/0-1/0 because its lower mouth is at z = −822.06 m, 78.56 m below FloatingIslands’ content band (bottom −743.5 m; that point is in TunnelNetwork). The plan’s passage audit also fails; details are below.

Bisect

All 57 commits in a176d07^..3202193 were exported and built with Clang 19, -ffp-contract=off, -j12, taskset CPUs 0–11, a 6 GiB address-space limit, and TMPDIR under Saved/. 56 built. 3e443bf (L37/R1) fails to compile on an extra ); its following L37 fix 7f2d84d builds. The timing matrix has 402 successful runs: three shuffled/interleaved passes per recipe and data/commit configuration. Every run used the same view (−12, −118, −20 m), 12 workers, --force-generate, clip radius 3, max clip level 8, LOD8 step 0, hash off. The fixed data was 3202193 (7b095f9b…) for every commit; commits before L24 also ran the pre-L24 file (289cd755…).

Not Core. On fixed data, SurfaceWorld LOD0 worker time per tile is 10.05 ms at L14 and 12.78 ms at L61 (1.27×); whole-view is 4.31 s and 5.02 s (1.16×). The summed worker time across LOD0–8 is 31.2 s and 38.7 s (1.24×). This does not reproduce the Windows roughly 2× per-tile rise, so the cause is on the Unreal side: U1 or the adapter’s Core driving path. No Core fix was made; Source/ was untouched.

W/F/N/Σ/L0 below means whole-view seconds / first-tile seconds / near-ring seconds / summed worker-seconds across LOD0–8 / LOD0 worker-ms per tile. Each parenthesized percent is the delta against L14 for the same recipe and data file. L0 steps mark upward changes of at least 10% versus the previous successful commit in that dataset. For LOD0 p50/p95, builds before L55 lack tile-drain counters; later values are from separate --instrument-tile-drain passes on the same view/data (SW then FI, ms). Those opt-in passes have force-generation off, as required for the counters, and do not contribute to the timing medians.

Round Commit Data SW W/F/N/Σ/L0 (Δ%) FI W/F/N/Σ/L0 (Δ%) L0 step ≥10% LOD0 p50/p95 SW; FI (ms)
L14 a176d07 3202193 4.31 (+0%) / 1.71 (+0%) / 1.76 (+0%) / 31.2 (+0%) / 10.05 (+0%) 23.57 (+0%) / 3.21 (+0%) / 3.61 (+0%) / 236.6 (+0%) / 166.93 (+0%) — —
L14 a176d07 pre-L24 4.53 (+0%) / 1.61 (+0%) / 1.67 (+0%) / 34.6 (+0%) / 11.71 (+0%) 26.34 (+0%) / 3.91 (+0%) / 4.70 (+0%) / 291.3 (+0%) / 236.10 (+0%) — —
L15 a6cd7cf 3202193 4.26 (-1%) / 1.44 (-16%) / 1.49 (-15%) / 32.1 (+3%) / 10.50 (+5%) 23.88 (+1%) / 3.67 (+15%) / 4.30 (+19%) / 249.3 (+5%) / 179.96 (+8%) — —
L15 a6cd7cf pre-L24 2.96 (-35%) / 1.33 (-17%) / 1.36 (-18%) / 19.4 (-44%) / 5.30 (-55%) 27.50 (+4%) / 3.62 (-7%) / 4.33 (-8%) / 297.1 (+2%) / 247.25 (+5%) — —
L16 9a7c87e 3202193 3.59 (-17%) / 1.60 (-6%) / 1.65 (-6%) / 23.9 (-23%) / 8.47 (-16%) 23.94 (+2%) / 3.92 (+22%) / 4.26 (+18%) / 256.3 (+8%) / 182.06 (+9%) — —
L16 9a7c87e pre-L24 4.49 (-1%) / 1.75 (+9%) / 1.81 (+9%) / 32.7 (-5%) / 10.08 (-14%) 26.57 (+1%) / 3.66 (-6%) / 4.65 (-1%) / 271.7 (-7%) / 231.16 (-2%) SW ↑90.2% —
L17 8704129 3202193 3.95 (-8%) / 1.51 (-12%) / 1.56 (-11%) / 29.9 (-4%) / 9.06 (-10%) 24.71 (+5%) / 3.95 (+23%) / 4.67 (+29%) / 264.2 (+12%) / 176.04 (+5%) — —
L17 8704129 pre-L24 2.93 (-35%) / 0.87 (-46%) / 0.94 (-44%) / 26.0 (-25%) / 8.88 (-24%) 25.58 (-3%) / 2.40 (-38%) / 3.11 (-34%) / 278.4 (-4%) / 231.14 (-2%) — —
L18 dcf0763 3202193 4.08 (-5%) / 1.55 (-9%) / 1.60 (-9%) / 31.4 (+1%) / 10.92 (+9%) 29.75 (+26%) / 3.56 (+11%) / 4.42 (+22%) / 317.7 (+34%) / 257.73 (+54%) SW ↑20.5%, FI ↑46.4% —
L18 dcf0763 pre-L24 4.02 (-11%) / 1.65 (+3%) / 1.70 (+2%) / 28.4 (-18%) / 6.94 (-41%) 25.67 (-3%) / 4.14 (+6%) / 4.96 (+6%) / 258.9 (-11%) / 177.66 (-25%) — —
L19 811c4b7 3202193 4.79 (+11%) / 1.80 (+5%) / 1.85 (+5%) / 36.5 (+17%) / 12.45 (+24%) 26.15 (+11%) / 2.54 (-21%) / 3.15 (-13%) / 283.8 (+20%) / 235.00 (+41%) SW ↑14.1% —
L19 811c4b7 pre-L24 3.51 (-23%) / 0.80 (-50%) / 0.86 (-49%) / 32.4 (-6%) / 9.34 (-20%) 25.38 (-4%) / 3.94 (+1%) / 4.64 (-1%) / 259.8 (-11%) / 213.94 (-9%) SW ↑34.6%, FI ↑20.4% —
L20 6f9c7bd 3202193 4.52 (+5%) / 1.74 (+2%) / 1.78 (+1%) / 33.3 (+7%) / 12.08 (+20%) 23.30 (-1%) / 2.36 (-27%) / 3.22 (-11%) / 253.6 (+7%) / 173.02 (+4%) — —
L20 6f9c7bd pre-L24 4.32 (-5%) / 1.65 (+2%) / 1.69 (+1%) / 31.9 (-8%) / 10.91 (-7%) 22.85 (-13%) / 2.18 (-44%) / 3.41 (-27%) / 229.8 (-21%) / 170.36 (-28%) SW ↑16.8% —
L20 review 5d5447d 3202193 4.76 (+11%) / 1.63 (-4%) / 1.70 (-3%) / 35.1 (+13%) / 11.85 (+18%) 25.93 (+10%) / 3.91 (+22%) / 4.78 (+33%) / 275.4 (+16%) / 227.96 (+37%) FI ↑31.8% —
L20 review 5d5447d pre-L24 4.50 (-1%) / 1.65 (+3%) / 1.72 (+3%) / 34.5 (-0%) / 11.83 (+1%) 27.01 (+3%) / 3.09 (-21%) / 3.87 (-18%) / 284.0 (-3%) / 225.37 (-5%) FI ↑32.3% —
L21 8753f6d 3202193 4.26 (-1%) / 1.50 (-12%) / 1.56 (-11%) / 33.0 (+6%) / 11.30 (+12%) 26.21 (+11%) / 3.80 (+19%) / 5.09 (+41%) / 268.6 (+14%) / 192.22 (+15%) — —
L21 8753f6d pre-L24 3.78 (-17%) / 1.02 (-37%) / 1.12 (-33%) / 33.1 (-4%) / 10.18 (-13%) 23.71 (-10%) / 3.73 (-5%) / 4.60 (-2%) / 260.3 (-11%) / 196.70 (-17%) — —
L22 55c26a0 3202193 4.01 (-7%) / 1.48 (-13%) / 1.55 (-12%) / 28.9 (-7%) / 10.23 (+2%) 35.34 (+50%) / 4.74 (+48%) / 5.49 (+52%) / 369.9 (+56%) / 240.61 (+44%) FI ↑25.2% —
L22 55c26a0 pre-L24 4.13 (-9%) / 1.58 (-2%) / 1.64 (-2%) / 30.4 (-12%) / 11.32 (-3%) 33.67 (+28%) / 3.50 (-11%) / 4.07 (-13%) / 360.8 (+24%) / 273.30 (+16%) SW ↑11.2%, FI ↑38.9% —
L23 71d09a8 3202193 3.38 (-21%) / 1.54 (-10%) / 1.60 (-9%) / 22.0 (-29%) / 9.67 (-4%) 33.92 (+44%) / 3.21 (+0%) / 3.77 (+4%) / 359.9 (+52%) / 234.50 (+40%) — —
L23 71d09a8 pre-L24 4.34 (-4%) / 1.70 (+6%) / 1.76 (+5%) / 31.5 (-9%) / 10.93 (-7%) 30.60 (+16%) / 4.29 (+10%) / 5.16 (+10%) / 316.1 (+9%) / 222.51 (-6%) — —
L24 cb9236c 3202193 2.83 (-34%) / 0.95 (-44%) / 1.03 (-42%) / 17.5 (-44%) / 6.50 (-35%) 36.87 (+56%) / 4.38 (+37%) / 5.18 (+44%) / 404.1 (+71%) / 298.88 (+79%) FI ↑27.5% —
L25 ef789aa 3202193 4.29 (-0%) / 1.72 (+1%) / 1.78 (+1%) / 30.8 (-1%) / 10.86 (+8%) 36.36 (+54%) / 2.18 (-32%) / 2.88 (-20%) / 387.8 (+64%) / 286.83 (+72%) SW ↑67.0% —
L26 37a2230 3202193 4.06 (-6%) / 1.69 (-1%) / 1.76 (-0%) / 32.4 (+4%) / 7.90 (-21%) 43.11 (+83%) / 4.54 (+42%) / 5.28 (+46%) / 456.7 (+93%) / 325.35 (+95%) FI ↑13.4% —
L27 07b3fe2 3202193 4.10 (-5%) / 1.44 (-16%) / 1.49 (-16%) / 32.2 (+3%) / 9.12 (-9%) 40.19 (+71%) / 3.68 (+15%) / 4.79 (+33%) / 441.3 (+87%) / 255.16 (+53%) SW ↑15.4% —
L28 f23ab7b 3202193 4.44 (+3%) / 1.72 (+1%) / 1.78 (+1%) / 32.1 (+3%) / 10.38 (+3%) 32.25 (+37%) / 3.55 (+11%) / 4.26 (+18%) / 339.3 (+43%) / 248.36 (+49%) SW ↑13.9% —
L29 63c7e56 3202193 2.55 (-41%) / 0.72 (-58%) / 0.75 (-58%) / 21.5 (-31%) / 6.19 (-38%) 34.47 (+46%) / 2.83 (-12%) / 3.64 (+1%) / 376.1 (+59%) / 265.76 (+59%) — —
L30 103d8b4 3202193 4.41 (+2%) / 1.59 (-7%) / 1.65 (-6%) / 32.1 (+3%) / 10.65 (+6%) 36.40 (+54%) / 4.25 (+33%) / 4.96 (+37%) / 387.8 (+64%) / 303.13 (+82%) SW ↑72.0%, FI ↑14.1% —
L31 28069f0 3202193 3.12 (-27%) / 1.55 (-9%) / 1.61 (-8%) / 18.8 (-40%) / 7.23 (-28%) 37.60 (+60%) / 4.14 (+29%) / 5.17 (+43%) / 407.7 (+72%) / 296.35 (+78%) — —
L32 d0158c6 3202193 4.88 (+13%) / 1.57 (-8%) / 1.71 (-3%) / 37.2 (+19%) / 12.07 (+20%) 34.29 (+46%) / 2.85 (-11%) / 3.31 (-8%) / 362.5 (+53%) / 229.88 (+38%) SW ↑66.9% —
L33 b7a47be 3202193 2.65 (-38%) / 0.86 (-49%) / 0.89 (-50%) / 21.3 (-32%) / 9.38 (-7%) 31.98 (+36%) / 2.91 (-9%) / 3.84 (+7%) / 350.3 (+48%) / 257.32 (+54%) FI ↑11.9% —
L34 49e29fc 3202193 4.40 (+2%) / 1.53 (-11%) / 1.58 (-10%) / 34.5 (+11%) / 10.81 (+8%) 36.15 (+53%) / 4.22 (+31%) / 5.02 (+39%) / 378.7 (+60%) / 252.48 (+51%) SW ↑15.3% —
L35 ba01707 3202193 4.80 (+11%) / 1.77 (+4%) / 1.84 (+4%) / 36.0 (+15%) / 7.08 (-30%) 35.01 (+49%) / 4.17 (+30%) / 4.96 (+37%) / 390.0 (+65%) / 289.50 (+73%) FI ↑14.7% —
L36 8be1c4d 3202193 3.67 (-15%) / 1.03 (-40%) / 1.08 (-39%) / 34.2 (+10%) / 10.92 (+9%) 34.93 (+48%) / 2.72 (-15%) / 3.21 (-11%) / 384.8 (+63%) / 266.69 (+60%) SW ↑54.2% —
L37/R1 3e443bf 3202193 — — compile fail —
L37 fix/R2 7f2d84d 3202193 4.42 (+3%) / 1.73 (+1%) / 1.78 (+1%) / 32.2 (+3%) / 10.07 (+0%) 38.96 (+65%) / 4.48 (+40%) / 5.25 (+46%) / 417.3 (+76%) / 259.39 (+55%) — —
R3 945225d 3202193 4.86 (+13%) / 1.85 (+8%) / 1.91 (+9%) / 35.3 (+13%) / 10.69 (+6%) 33.96 (+44%) / 2.68 (-17%) / 3.43 (-5%) / 374.9 (+58%) / 266.08 (+59%) — —
L38/R4 cf57c87 3202193 3.14 (-27%) / 1.34 (-22%) / 1.36 (-23%) / 21.5 (-31%) / 5.96 (-41%) 37.38 (+59%) / 4.72 (+47%) / 5.58 (+55%) / 396.3 (+67%) / 252.07 (+51%) — —
R5 88c54ea 3202193 3.37 (-22%) / 1.46 (-15%) / 1.56 (-11%) / 30.5 (-2%) / 11.09 (+10%) 39.45 (+67%) / 3.52 (+10%) / 4.89 (+36%) / 416.6 (+76%) / 265.16 (+59%) SW ↑85.8% —
L39/R6 753cc9c 3202193 3.32 (-23%) / 0.86 (-50%) / 0.88 (-50%) / 21.7 (-31%) / 9.49 (-5%) 37.03 (+57%) / 4.41 (+37%) / 5.56 (+54%) / 389.5 (+65%) / 232.40 (+39%) — —
L40/R7 efa9e89 3202193 4.57 (+6%) / 1.62 (-5%) / 1.68 (-4%) / 35.3 (+13%) / 10.78 (+7%) 38.63 (+64%) / 3.39 (+6%) / 3.89 (+8%) / 412.4 (+74%) / 306.46 (+84%) SW ↑13.6%, FI ↑31.9% —
L41/R8 1abee91 3202193 4.30 (-0%) / 1.71 (+0%) / 1.81 (+3%) / 26.8 (-14%) / 8.08 (-20%) 38.25 (+62%) / 2.88 (-10%) / 3.42 (-5%) / 413.8 (+75%) / 324.95 (+95%) — —
L42/R9 2566f53 3202193 4.37 (+1%) / 1.51 (-11%) / 1.56 (-11%) / 34.1 (+9%) / 10.05 (+0%) 37.43 (+59%) / 2.15 (-33%) / 2.48 (-31%) / 427.7 (+81%) / 306.73 (+84%) SW ↑24.4% —
R9 497e974 3202193 4.55 (+6%) / 1.75 (+2%) / 1.79 (+2%) / 33.5 (+7%) / 10.31 (+3%) 35.69 (+51%) / 4.87 (+52%) / 5.68 (+57%) / 372.0 (+57%) / 257.01 (+54%) — —
L43/R10 422dd77 3202193 3.54 (-18%) / 1.60 (-6%) / 1.69 (-4%) / 23.7 (-24%) / 8.14 (-19%) 37.50 (+59%) / 2.92 (-9%) / 3.37 (-7%) / 401.3 (+70%) / 320.95 (+92%) FI ↑24.9% —
L44 036bd53 3202193 3.48 (-19%) / 0.95 (-44%) / 1.03 (-41%) / 30.2 (-3%) / 10.30 (+3%) 36.78 (+56%) / 4.32 (+35%) / 5.20 (+44%) / 390.7 (+65%) / 239.12 (+43%) SW ↑26.6% —
L45/R12 f13b145 3202193 4.50 (+5%) / 1.67 (-2%) / 1.75 (-1%) / 33.9 (+9%) / 10.38 (+3%) 38.59 (+64%) / 4.50 (+40%) / 5.26 (+46%) / 410.7 (+74%) / 300.28 (+80%) FI ↑25.6% —
L46/R13 e016f6a 3202193 4.15 (-4%) / 1.93 (+13%) / 2.01 (+14%) / 31.3 (+1%) / 9.64 (-4%) 38.45 (+63%) / 4.10 (+28%) / 4.99 (+38%) / 406.5 (+72%) / 296.82 (+78%) — —
L47/R14 c091e44 3202193 4.70 (+9%) / 1.84 (+8%) / 1.90 (+8%) / 35.6 (+14%) / 10.76 (+7%) 39.16 (+66%) / 4.03 (+26%) / 4.72 (+31%) / 420.2 (+78%) / 274.84 (+65%) SW ↑11.6% —
L48/R15 3073d1b 3202193 4.66 (+8%) / 1.54 (-10%) / 1.65 (-7%) / 37.4 (+20%) / 11.89 (+18%) 37.86 (+61%) / 3.98 (+24%) / 4.82 (+33%) / 406.9 (+72%) / 289.92 (+74%) SW ↑10.5% —
R16 6110f11 3202193 4.72 (+10%) / 1.47 (-14%) / 1.59 (-10%) / 33.1 (+6%) / 10.39 (+3%) 35.86 (+52%) / 2.42 (-25%) / 3.66 (+1%) / 400.1 (+69%) / 284.96 (+71%) — —
R17 b9ccff4 3202193 3.78 (-12%) / 1.33 (-22%) / 1.36 (-22%) / 34.4 (+10%) / 10.78 (+7%) 38.70 (+64%) / 4.79 (+49%) / 5.58 (+55%) / 409.9 (+73%) / 282.19 (+69%) — —
L50/R18-R19 ca85e87 3202193 4.95 (+15%) / 1.76 (+3%) / 1.84 (+4%) / 38.2 (+22%) / 11.85 (+18%) 36.53 (+55%) / 4.80 (+50%) / 5.20 (+44%) / 381.1 (+61%) / 251.21 (+50%) — —
L51/R20 7555025 3202193 4.44 (+3%) / 1.61 (-6%) / 1.67 (-5%) / 33.8 (+8%) / 10.63 (+6%) 18.33 (-22%) / 2.32 (-28%) / 2.80 (-22%) / 174.2 (-26%) / 180.36 (+8%) — —
L51 c42c785 3202193 4.47 (+4%) / 1.61 (-6%) / 1.67 (-5%) / 32.2 (+3%) / 9.50 (-5%) 14.36 (-39%) / 2.78 (-13%) / 3.17 (-12%) / 140.5 (-41%) / 131.17 (-21%) — —
L52/R21 408c14b 3202193 4.49 (+4%) / 1.58 (-7%) / 1.65 (-6%) / 35.5 (+14%) / 11.05 (+10%) 12.50 (-47%) / 4.75 (+48%) / 5.20 (+44%) / 102.9 (-57%) / 125.17 (-25%) SW ↑16.3% —
L53/R22 20f6c98 3202193 4.42 (+3%) / 1.66 (-3%) / 1.83 (+4%) / 29.4 (-6%) / 9.54 (-5%) 11.57 (-51%) / 3.77 (+17%) / 4.18 (+16%) / 112.4 (-52%) / 134.18 (-20%) — —
L54/R23 9458217 3202193 4.62 (+7%) / 1.50 (-12%) / 1.56 (-12%) / 35.6 (+14%) / 11.17 (+11%) 14.71 (-38%) / 4.59 (+43%) / 5.11 (+42%) / 123.4 (-48%) / 137.96 (-17%) SW ↑17.1% —
L54 e35a77f 3202193 2.66 (-38%) / 0.77 (-55%) / 0.81 (-54%) / 22.6 (-27%) / 8.05 (-20%) 12.72 (-46%) / 2.50 (-22%) / 3.18 (-12%) / 94.6 (-60%) / 92.67 (-44%) — —
L55/R24 aee1487 3202193 4.16 (-3%) / 1.73 (+1%) / 1.79 (+2%) / 27.7 (-11%) / 8.99 (-11%) 10.74 (-54%) / 1.45 (-55%) / 1.65 (-54%) / 111.8 (-53%) / 146.17 (-12%) SW ↑11.7%, FI ↑57.7% 4.65/11.20; 94.70/122.83
L56/R25 4e64b75 3202193 3.11 (-28%) / 0.76 (-55%) / 0.79 (-55%) / 28.0 (-10%) / 10.12 (+1%) 9.68 (-59%) / 2.65 (-17%) / 3.23 (-10%) / 87.3 (-63%) / 95.93 (-43%) SW ↑12.5% 4.61/11.62; 97.07/124.52
L57/R26 1e3e04b 3202193 5.11 (+19%) / 1.59 (-7%) / 1.67 (-5%) / 43.7 (+40%) / 14.12 (+41%) 12.83 (-46%) / 2.48 (-23%) / 2.83 (-21%) / 125.4 (-47%) / 130.60 (-22%) SW ↑39.6%, FI ↑36.1% 4.77/11.47; 86.41/114.69
L58/R27 6c6cc37 3202193 3.72 (-14%) / 0.93 (-46%) / 0.98 (-44%) / 33.5 (+7%) / 10.09 (+0%) 9.00 (-62%) / 1.93 (-40%) / 2.13 (-41%) / 85.1 (-64%) / 95.95 (-43%) — 4.62/11.10; 86.08/114.90
R28 fbae82e 3202193 4.85 (+13%) / 1.85 (+8%) / 1.90 (+8%) / 35.9 (+15%) / 11.24 (+12%) 10.47 (-56%) / 2.18 (-32%) / 2.42 (-33%) / 99.6 (-58%) / 109.15 (-35%) SW ↑11.5%, FI ↑13.8% 4.71/11.40; 85.49/114.11
L60 c0c204b 3202193 3.22 (-25%) / 1.46 (-14%) / 1.49 (-16%) / 21.0 (-33%) / 8.59 (-15%) 11.36 (-52%) / 2.61 (-19%) / 2.99 (-17%) / 101.0 (-57%) / 96.22 (-42%) — 5.21/11.99; 85.26/114.07
L61/R31 1359578 3202193 5.02 (+16%) / 1.79 (+5%) / 1.86 (+6%) / 38.7 (+24%) / 12.78 (+27%) 13.38 (-43%) / 2.92 (-9%) / 3.26 (-10%) / 126.2 (-47%) / 131.47 (-21%) SW ↑48.8%, FI ↑36.6% 5.01/11.65; 85.11/113.95
R32 3202193 3202193 3.69 (-14%) / 1.61 (-6%) / 1.71 (-3%) / 27.2 (-13%) / 11.06 (+10%) 12.73 (-46%) / 2.60 (-19%) / 2.94 (-18%) / 124.2 (-48%) / 131.70 (-21%) — 4.89/11.24; 86.41/114.52

Four-strate U2 baseline

The local mesh windows are all nonempty and deterministic at W1/W8/W12 on Clang and GCC. Each hash vector matched across all six compiler/worker configurations. The cost numbers are medians of three W12 full-view runs using the comparable clip-radius-3 speed-matrix configuration, with each camera in the named stratum. Because the owner plan fails strict ownership and passage validation, these full-view cost runs use a temporary diagnostic binary that continues after those validators; no in-tree Core or Unreal Source/ file was changed; the bypass is confined to Saved/Linux/l64/u2cost/export. The cost cameras were SW viewer_start (0, 0, -496 m), FI passage-1 upper mouth (19.166, -14.200, -732.201 m), TN passage-1 lower mouth (19.166, -14.200, -753.374 m), and VS passage-2 lower mouth (−20.063, −5.068, −900.354 m). The SW local mesh window above is the resolved route arrival, not its cost camera.

Strate (chunks; content band) Core generator path Local window (nonempty/27; triangles) Determinism Whole-view W/F/N/Σ/L0 Arrival→departure audit
SW (0…−63; −503.5…7.5 m) SurfaceWorld-native route spawn: 9/27; 25,672 W1/W8/W12 Clang/GCC: PASS 6.34/0.63/0.80/68.8/60.39 SW route fails: 24/33 points unsupported; SW→FI interface fails
FI (−64…−93; −743.5…−504.5 m) compiled-analytic/FloatingIslands (legacy) passage 1 upper mouth: 21/27; 93,149 W1/W8/W12 Clang/GCC: PASS 8.41/0.53/0.97/95.1/74.21 SW→FI lower mouth out of band; FI→TN interface fails
TN (−94…−109; −871.5…−744.5 m) compiled-analytic/TunnelNetwork (legacy) passage 1 lower mouth: 13/27; 65,248 W1/W8/W12 Clang/GCC: PASS 6.50/0.53/0.79/71.6/56.33 FI→TN and TN→VS interfaces fail
VS (−110…−125; −999.5…−872.5 m) compiled-analytic/VerticalShafts (legacy) passage 2 lower mouth: 5/27; 4,761 W1/W8/W12 Clang/GCC: PASS 3.39/0.52/0.54/34.6/14.51 TN→VS: 3 unsupported; floor ripple 3.57 cm > 2.5 cm

Across the three interface audits, 0/3 passages passed. In addition to the out-of-band 0–1 mouth, passage 1–2 has floor ripple 4.78 cm (2.5 cm limit), and 2–3 has 3.57 cm ripple; unsupported/blocked samples also remain. Native caves and lattice are not enabled in this plan; TN and VS use their legacy analytic evaluators.

Failed runs

  • Historical build 3e443bf (L37/R1) failed at ShapeWorldInternal.h:1554 with an extra ); the next L37 fix built and was measured. The corrected bisect has 402/402 successful timing runs. An earlier discarded invocation used unsupported --data= syntax and was replaced by the positional data path.
  • verify.py --tier fast --out Saved/Linux/l64 in the copied legacy-plan shadow reached the motif fixture and exited 1: motif-slice1-capsule-reachability-clang rejected its motif override because it did not resolve to a SurfaceWorld strate. Its hard-coded two-strata style map does not validate U2; U2 boundary, determinism, and passage audits are reported above. A first stale-source-stamp refusal was corrected by copying the L37 audit script into the shadow.
  • The first U2 whole-view SW pass exited at the passage walkability gate (65/7,729 samples solid). The cost table reruns use the same diagnostic binary for all four strata with route validation bypassed, and retain the audit failure. A first build attempt also used a nonexistent TMPDIR; it was corrected and the diagnostic build passed.

Evidence is under Saved/Linux/l64: runs-spawn-correct.json and runs-spawn-correct/, profiles-spawn.json, u2probe/determinism.json, u2probe/passage-audit-clang.log, u2probe/boundary-dispatch.log, u2cost/speed-cost.json, and verify-fast.log.

Orchestrator review (Claude)

The bisect timings are inflated by machine contention. I was building Unreal Engine on cores 12–19 during this round (memory bandwidth and disk I/O), and adjacent commits swing ±40%. A clean re-measure on a quiet machine (6 interleaved runs each, same binaries, data owner_world-3202193.json, same command and camera):

Build Whole view (median of 6) Worker-s, LOD0–8 LOD0 ms/tile
L14 a176d07 ≈1.55 s 12.62 3.99
L61 1359578 ≈1.72 s 14.65 4.99

Core regressed about +16% in worker time and +25% per LOD0 tile, not ×2. The owner's Windows trace shows about 33 ms per LOD0 tile p95, roughly 6× Core's per-tile cost here, so most of the Windows regression is on the Unreal side or isn't a like-for-like comparison (R33: U1's trace was not a controlled A/B, at 2,578 tiles). Core's own +16% is plausibly R33 suspect #1: per-sample ApplyResolvedStrateComposition (L23) plus the per-sample passage loop (L37). The new four-strate plan has real defects:

  • passage 0–1's lower mouth is out of FI's band;
  • all three interfaces fail their audits;
  • FI, TN and VS run on legacy evaluators;
  • the full view costs 3.4–8.4 s per strate.

Next: get the Linux UE build to compile, then trace the Unreal path itself for L14 vs HEAD here.