Fr0zkaandClaude Opus 5.5
b09728354d
Five conflicts resolved (CURRENT-STATE, ResolvedLayoutInternal.h, StreamBenchmarkHeadless.cpp, StreamBenchmarkPipeline.h, WorldAssemblyPrototype.h) keeping liminal + cave planner; L156 band fit only; evaluator v239 with v239 baselines (18/18 Clang/GCC W1/W8/W12). Fast tier FAILS (FI upper arrival: 40 unsupported chamber samples, 10 chamber-floor shell violations); TN and VS composed sweeps 0/16 (passage, seam, shell gates). Living plane loads at PostConfigInit (UE smoke passes). Source/: VoxelForgeCoreAdapter.cpp; VoxelForge.uplugin. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
3.1 KiB
Linux136 — cave planner speed
Source stamp ebe6b32a7cd96c5aa69c6dc64ae6b9c6ff4bc18bc484efc5148611c48723c28f (v239; v236 is commit 2d89023).
Findings
The tile query now keeps sorted passage IDs with route candidates, so the composer skips unrelated routes. A route-free tile returns before per-sample coordinate and scale math. Native cave evaluation also returns the exact host-rock value when its conservative tile feature list is empty. These paths preserve v236 output; no version bump was needed.
The LOD0–2 census shows the shared composer does substantial work outside Tunnel Network. Surface World has zero native cave samples at LOD0–3, yet near-ready passage visits are 19.61M on both assembler/world and v236. The route-ID path reduces that count to 3.20M, but cumulative near-ready composer worker time stays around 3.5 s (3.42 s baseline, 3.51 s v236, 3.52 s v239). Whole-view classification worker time is 584 ms / 641 ms / 623 ms. Planning is about 4 ms. This rules out cave evaluation and planning as the Surface World regression, but the census does not isolate mouth-profile checks from the other shared composer/classifier work.
Floating Islands and Vertical Shafts have unrelated-route churn: v236 adds 7.79M and 8.43M passage visits over baseline. v239 brings those totals to 7.89M and 2.54M, and reduces composer worker time by about 9% and 8% versus v236. Tunnel Network falls from 34.63M visits in v236 to 7.47M; its composer worker time returns to baseline. The views still process about 4.7–5.4M native cave samples. The attempted LOD4–8 cave tile proofs produced no skips: every cave-view tile at those LODs remained unclassified.
L80 timing
Five interleaved measured runs per version and cell, one warmup per version. The pre-run gate was load 0.74 and 99.3% CPU idle. Cells show near-ready and whole-view median deltas versus assembler/world, in W8/W12 order; the last column shows near-ready deltas versus v236.
| Stratum | Near-ready Δ | Whole-view Δ | Near-ready Δ vs v236 |
|---|---|---|---|
| Surface World (0) | +4.29% / +5.78% | +2.45% / −0.17% | +0.52% / −1.99% |
| Floating Islands (1) | +8.88% / +5.48% | +1.31% / −0.48% | −2.40% / −4.29% |
| Tunnel Network (2) | +8.23% / +8.26% | +0.35% / +0.88% | −3.51% / −3.97% |
| Vertical Shafts (3) | +13.29% / +6.36% | +1.26% / −1.97% | −4.22% / −3.81% |
v239 improves on v236 in seven of eight near-ready cells, but misses the requested ≤+3% target in all eight cells versus assembler/world. Whole-view deltas stay within +2.45%/−1.97%.
Gates
- Fast tier: PASS, 289.9 s.
- Composed passage, seam, shell, mouth, ripple, naturalness, and blocked-route audits: 16/16 seeds PASS.
- Clang and GCC at W1/W8/W12: all six runs match 13 hash lines.
- Linux VFHostEditor Development: PASS, 5/5 actions in 41.33 s with
-NoUBA -MaxParallelActions=4; noprlimit.
Timing, census, seed, determinism, and build logs are under Saved/Linux/l136/. The completed timing matrix is Saved/Linux/l136/timing/summary.json; the interrupted partial matrix is excluded. No commit or push was made.