Fr0zka/VoxelForge
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tooling(test): one command to test VoxelForge, with a commandlet-vs-game parity check

Browse Source 
Tools/VoxelForgeTest.ps1 runs scenarios (canonical, owner, probe, perf, tests, parity) with the known-good flag sets, isolated user/config dirs, serialized launches, an owner-process guard, and a compact result.json + summary.txt instead of 300 MB logs that Codex has to stream. Answers the owner's question 'can the commandlet reflect in-game behaviour': -gameconfig=1 makes the commandlet use the authored settings, and the parity scenario compares per-tile geometry hashes between the commandlet export and the game's streaming path. Parity PASS: 64 commandlet tiles vs 810 game records, no differing tile.

First run of the tests scenario: 44 tests, 28 pass, 16 fail (8 composer, 6 op-stack equivalence, PassageLandsInOpenSpace, StrateParamBlendCoverage). Triage is the next round: stale expectations from the floor removal and the degrees change, versus real defects, especially the legacy-vs-op-stack equivalence ones.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
This commit is contained in:
Fr0zkaandClaude Opus 5 committed 2026-09-18 05:13:16 +02:00
1 parent 1c642fd182
commit a7b8d10de6
10 files changed
+1484 -52
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PERF-SAMPLED.md
+17
View File
@@ -6,6 +6,23 @@ Measurement-only report for the in-process stack sampler, 2026-09-13. No generat
The sampler is useful: it samples registered generation workers without adding scopes to the density hot path, tags every capture with LOD, finds the planted cost, and exposes the game path's dominant work. It is not yet an acceptance-clean low-overhead instrument at the requested default interval. Static timing and worker CPU move beyond the two-run off baseline; moving worker CPU is within noise, but moving request/generation timing is not.
## Current command entry point
Use the guarded harness at `E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1` for new measurements. It owns the staged host under `Saved\BuildHost`, serializes Unreal launches, uses an isolated `-userdir`, Zen/DDC directory, and log directory under the requested run directory, waits by PID, records crash/`UECC-*` evidence, and writes compact `result.json` plus a ≤40-line `summary.txt`; full logs remain beside them.
All paths passed to `-Out` must be absolute and under the plugin's `Saved` directory. The normal entry forms are:
```powershell
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario canonical -Build -Assets default -Label canonical
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario owner -Assets owner -Label owner
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario probe -Assets owner -Label probe
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario perf -Label perf
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario tests -Label tests
pwsh -NoLogo -NoProfile -File "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Tools\VoxelForgeTest.ps1" -Scenario parity -Build -Assets owner -Label parity
```
Optional controls are `-Cvars @{...}`, `-Out "E:\Projet Unreal\VoxelM\Plugins\VoxelForge\Saved\TestHarness\named"`, and `-Label name`. `-Assets owner` stages `DA_Strate3` and `DA_Settings` and records both SHA-256 values. `parity` exports 64 level-0 tiles through `VoxelForgeExplore -gameconfig=1`, runs the headless game's streaming path with the same owner settings, dumps per-tile hashes, and fails on the first missing or differing tile. Do not invent a direct `UnrealEditor` command line: the harness's launch ledger is the reproducible command record.
The headline game result is that LOD0 is not dominated by candidate AABB rejection. It is dominated by the evaluator and its downstream tunnel/floor/core and structural work. Coarse LOD1+ samples are dominated by room-cache/player-fit/room-landing work. The profile supports the suspected swept-tunnel/core duplication, but does not isolate the parameter-copy cost.
## Instrument and build