Fr0zkaandClaude Opus 5.5
05e78331fe
Identical builds differ <=0.6% median; residual near-ready vs assembler/world is real: Surface World +2.9..+3.6%, Vertical Shafts +4.0..+4.3%, Tunnel Network -3.6/+6.8%. Profile points at column-cache reads and the sweep lookup. Removed an unused sample field and per-lookup shared_ptr refcounting. 16/16 seeds, fast tier, Clang/GCC green. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
1.9 KiB
Linux137 report
Measurement
Two assembler/world binaries had identical SHA-256. On Surface World, their L80 near-ready medians differed by +0.52% at W8 and −0.59% at W12. The no-op v228 control differed by −0.37% and −0.05%; its stream hashes matched v227. Use ±1% as the median noise floor. Individual paired controls varied by up to about 2.5% (3.6% for one no-op pair).
The three-run L77 profile sampled LOD 0–8 with no lost samples. Function share changes, HEAD versus assembler/world: CachedColumn +0.93 pp, shared column TryRead +0.58 pp, prepareSweptTubeBase +0.38 pp, and sweptSample +0.17 pp. ApplyResolvedStratePassageCarves fell by 0.97 pp and FindBoundStratumForManagerZ by 0.39 pp. The profile points more toward column-cache work and the sweep lookup than mouth-profile work; it does not isolate one cause for every world.
Changes retained: removed the unused floorShellSignedDistance sample field; removed per-lookup shared-pointer refcounting and batched diagnostic counters per query scope.
L80 near-ready medians
Each delta is candidate versus assembler/world; the A/A median noise floor is ±1%.
| World | W8 delta | W12 delta |
|---|---|---|
| Surface World | +3.58% (±1%) | +2.90% (±1%) |
| Floating Islands | +0.19% (±1%) | +2.20% (±1%) |
| Tunnel Network | −3.60% (±1%) | +6.84% (±1%) |
| Vertical Shafts | +4.31% (±1%) | +3.96% (±1%) |
The requested targets are not met except Floating Islands W8 and Tunnel Network W8. Whole-view medians were within −8.39% to +0.27% of baseline.
Verification
Fast tier passed (273 checks); the composed sweep passed 16/16 seeds with seams and blocked controls; Clang/GCC W1/W8/W12 matched all 13 hashes. Linux UE host build passed with -NoUBA -MaxParallelActions=4 and no prlimit. Output audits stayed green, so version and baselines were left unchanged. No commit, push, or stash.