0 Fork 0

Files

T

Fr0zkaandClaude Opus 5.5 efa9e89468 refactor(core): bind generator fields once per resolved strate (L40, migration step 2); research R7

SurfaceWorld, FloatingIslands and legacy evaluators sit behind one immutable densityAt adapter held in
a variant (no virtual call per sample); BindResolvedStratumFields builds a sorted strate table at world
setup, so DensityAt, native SW/FI sampling and ClassifyBoxDetailed no longer ask the strate manager per
sample. No arithmetic change: owner hashes identical to L37/L39 at W1/W12. Cost vs L39: FI -0.66%,
SW -1.31% (noise). verify fast/medium/full PASS (full re-run alone: 407 gates).
Research R7: disk cache design for UE5/Windows/multiplayer and a zlib probe (about 1/3 size, 0.1 ms/tile).

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

2026-10-23 17:15:46 +02:00

4.6 KiB

Raw Blame History

Linux40 — immutable field binding

Change

ShapeWorldInternal.h now has concrete immutable SurfaceWorld, FloatingIslands and legacy adapters with a common densityAt(x,y,z,out) call shape. FGeneratorFieldAdapterV1 stores them in a variant, so hot sampling uses concrete dispatch without a virtual call. FI retains its fixed-point sampler and captures the legacy boundary parameters while binding.

BindResolvedStratumFields creates a sorted table of resolved strate bounds, plan references and evaluator objects during world construction/initialization. Native SurfaceWorld and FI sampling, boundary composition lookup, legacy SurfaceWorld bridging and ClassifyBoxDetailed select from that table. The manager layout width is captured once for the legacy-compatible chunk projection.

Lookup boundary and identity

The direct manager selection lookups in native SurfaceWorld/FI evaluation and box classification are removed. DensityAt now selects native fields through the bound table. Legacy UVoxelGenerator::GetDensityAt still performs its own manager routing when reached through the legacy bridge or an out-of-plan fallback. Manager-derived layout width/FI compatibility parameters are captured at bind time; manager access also remains in build/planning helpers. Generator-specific boundary and passage policy still lives in FStandaloneShapeWorld.

No density arithmetic, fixed-point conversion order, evaluator version or Unreal adapter changed. The canonical owner FI and SurfaceWorld hashes are identical to L37/L39 for W1 and W12. The owner layout IDs also match L39: FI b7a7d9ecc48d8f26, SurfaceWorld 8a52e84c26ed92d2. Fast logs and hash summaries: Saved/Linux/l40/fast-trial1.log and Saved/Linux/l40/fast-summary.json.

Cost

Five interleaved pairs against the L39 Clang binary, using the canonical owner recipes:

Recipe	L39 median	L40 median	Median change	Paired changes (L40 vs L39)
FloatingIslands	17.556 s	17.440 s	−0.66%	+2.26%, −3.99%, −0.66%, −0.87%, −0.47%
SurfaceWorld	1.829 s	1.805 s	−1.31%	−2.64%, −0.22%, +0.55%, −0.44%, −1.48%

Both deltas are within observed pair noise; no per-voxel regression was measured. Raw output is in Saved/Linux/l40/cost-l40-vs-l39-run.log and cost-l40-vs-l39-summary.json.

Verification

Tier	Result	Wall time	Evidence
Fast	PASS	162.0 s	Canonical owner hashes match L37 for both recipes at W1/W12; all controls pass.
Medium	PASS	468.6 s	Transition, route/proof, LOD0–8 ring, owner shape and capture matrix pass.
Full	TIMED OUT	1740 s (29 min)	GCC build and supplement passed; L25 and crossing24 determinism passed for both compilers, and Clang upper-SW style0 W1/W8/W12 passed. The run timed out during the GCC upper-SW style0 W1 cell; no full summary or cost-baseline/pair results were produced.

Clang and GCC candidate builds succeeded with source stamp f35bc59c6172370e98dc83ec8ee5ba51dfab83a3faf075c26189fc83172f4ac8. The full run used the requested CPU/memory limits and a 1740-second timeout. At the stop, host load was 20.5. The partial log is Saved/Linux/l40/full-final.log; a full-summary.json was not written. Other tier and build logs are under Saved/Linux/l40/.

Failed attempts

The first Clang build passed unique_ptr objects to adapter constructors expecting raw observer pointers; corrected with .get(), then the build passed.
Two early cost-harness launches failed before starting measurements: one shell quoting error and two Python -c quoting errors. The later logged five-pair run completed successfully.
The single full-tier run reached its 1740-second timeout while the GCC upper-SW fixture style0 W1 determinism cell was still running. The cell has no completion record; the run returned timeout status 124. No full-tier L36 baseline build or full-tier cost pairs ran.

No Unreal compile was run; no Unreal adapter or VoxelWorld.cpp source change was required.

Orchestrator review (Claude)

Accepted. The full tier had timed out here under load (two Lunas plus host load of about 20). I re-ran it alone on the final source (stamp f35bc59c…): PASS, 407 gates, 2857 s. Its built-in cost comparison is against L36 and shows FI +3.03% (all five pairs positive) and SW +0.38%. That FI delta is L37's mouth geometry; this round's own L39→L40 pairs are −0.66% FI and −1.31% SW. Two follow-ups for the tool:

the full tier's cost baseline is hard-coded to L36; it should follow the previous commit;
the full tier now needs about 48 minutes including builds, so its timeout must be above 29 minutes.

4.6 KiB Raw Blame History Unescape Escape