Skip to content

racing_milestone2.md

Latest commit

 

History

History
159 lines (125 loc) · 7.41 KB

racing_milestone2.md

File metadata and controls

159 lines (125 loc) · 7.41 KB
Raw

Racing Game Milestone 2 — authentic XNA pipeline/XNB proof

Status

Milestone 2 is complete as of 2026-09-02. The former Win7 filesystem block, authentic content build, public CNA load path, processor vertex-layout proof, unchanged-XNA structural comparison and meaningful four-model draw all pass.

No GLB, glTF, modern-repository material, modern FNA effect or handwritten content decoder participated. The input is the original XNA 4 sample and its real Microsoft XNA Game Studio 4.0 pipeline output.

Authentic Windows build

The original powered-off win7 VM was protected with snapshot pre-ntfs-repair-2026-09-02. Work happened in the temporary linked clone cna-win7-racing-repair; chkdsk repaired the cloned filesystem without modifying the protected original disk.

Inside the offline Windows 7 32-bit guest, Visual Studio/XNA content at C:\Users\vboxuser\Desktop\XNAGameStudio built the unchanged RacingGameWindows1\RacingGame\RacingGame\RacingGame.csproj as Debug|x86. The successful full-game rebuild log records Microsoft Build Engine 4.0, XNA Game Studio 4.0, three original HLSL pow warnings, zero errors and 44.75 seconds. The earlier direct RacingGameContent.contentproj /t:Rebuild experiment failed because TargetPlatform was absent; it is retained as negative evidence and was superseded by the correct game-project build.

Evidence root:

/rv/tmp/samples/SAMPLE-152-XNA-4-Racing-Game-Kit-master/evidence/xna4-authentic-build

The shared-folder export contains 358 files and 347.20 MiB:

  • 339 XNB files;
  • all 57 Content/Models/*.xnb products from the original .X inputs and RacingGameModelProcessor;
  • all ten original compiled effect XNBs;
  • 244 texture XNBs;
  • original XACT .xgs, .xsb and .xwb products;
  • track/combi/height raw files, RacingGame.exe and PDB;
  • SHA256SUMS-debug.txt, a 358-entry hash manifest;
  • successful racinggame-debug.log and robocopy-debug.log, plus the retained failed direct-content experiment msbuild-debug.log.

CNA public load proof

samples/RacingGameHarness accepts the authentic exported Content directory via --content-root and uses only ContentManager::Load<Model> for these original IDs:

Models/Car
Models/Windmill
Models/AlphaDeadTree
Models/Cube

On CNA_GRAPHICS_RENDERER=OPENGL33, CNA 8cab5f32a and sharp-runtimenext 9cc96cd5, the pre-draw harness reports 64/64 PASS. The measured model graph is:

Asset Bones Meshes Effect kind Vertex layout
Car 19 6 EffectMaterial processor tangent frame, 56 bytes
Windmill 7 2 EffectMaterial processor tangent frame, 56 bytes
AlphaDeadTree 7 2 EffectMaterial processor tangent frame, 56 bytes
Cube 2 1 BasicEffect position/normal, 24 bytes

Every part has non-null vertex/index/effect resources, positive vertex/primitive counts and finite bounds. The processor layout is exactly Position Vector3 at 0, Normal Vector3 at 12, TextureCoordinate Vector2 at 24, Tangent Vector3 at 32 and Binormal Vector3 at 44. This remains separate from the original game's runtime-generated 44-byte TangentVertex layout: Position at 0, texture coordinate at 12, normal at 20 and tangent at 32. The harness tests both independently.

Unchanged-XNA semantic comparison

An inspector built into the unchanged XNA game project emits the same ordered semantic report as the CNA harness. The committed scripts/compare-model-reports.py comparison is strict for record order, bone and mesh names, parent indices, part ranges, vertex layouts, effect type/technique and parameter reflection, texture dimensions, mip counts and surface formats. It allows only normal floating-point formatting differences for matrices/bounds and normalizes the representational difference between a nullable .NET name and CNA's non-null C++ string.

The authentic XNA report and the real OPENGL33 CNA run agree on all 531 records:

RACING_MODEL_REPORT_COMPARISON=PASS records=531 MODEL=4 BONE=35 MESH=11 PART=17 ELEMENT=82 PARAM=320 TEXTURE2D=30 TEXTURECUBE=31 TEXTURE=1

The comparison includes the processor-authored 56-byte tangent frames, all 320 effect parameter records and the original DXT formats for all 30 Texture2D and 31 TextureCube bindings. No XNB was modified and no asset was converted.

General CNA repairs

CNA commit 0eb5fc151 hardens reusable Effect Framework behavior exposed by these authentic XNBs:

  • accept repeated object records and bounded XNA type-1 auxiliary records;
  • preserve safe object/value bounds throughout the pinned MojoShader parser;
  • translate legacy Shader Model 1 TEXCRD for desktop GLSL and GLSL ES;
  • reject structurally invalid render-state records before applying a pass.

The first version incorrectly masked every render-state identifier with ~0xA0, which changed the legitimate XNA 4 shader pseudo-states 146/147 into 18/19. CNA commit 756096626 removes that invalid normalization, names the legacy PixelShaderConstant/SetSampler records 160/178 in MojoShader's enum and accepts exactly ordinary states 0..102 plus 146, 147, 160 and 178. Every other identifier is rejected before conversion to the C/C++ enum. Its regression applies an authentic Racing ShadowMap pass and would previously have failed as unsupported FogStart/FogEnd.

The committed 17,404-byte authentic ShadowMap EffectReader payload has SHA-256 0014e2286cdd67c96f7ab24c7bf311c533e1b12461d1cb9a5453a068c8842d5d. Qualification passed the direct Basic/Conformance/Racing MojoShader probe, a deterministic 2,000-iteration mutation campaign with seed 0xffffffffffffffff, and all 24 EffectTest.* tests in both Debug and ASan/UBSan configurations on a real Mesa OpenGL 4.5 core context. The pinned patch stack also passes clean apply and reverse/idempotence checks.

Meaningful four-model draw

The harness draws all 17 parts from the four public ContentManager model results. It uses the original objectMatrix, bone transforms, processor-authored mesh suffix technique selection and authentic effect/material parameters. It neither decodes nor modifies an XNB and does not use a renderer helper. The observed technique bindings include ReflectionSpecular20 for car glass, the expected SpecularWithReflection20/Specular20 car-part split, DiffuseSpecular20 for Windmill and AlphaDeadTree, and BasicEffect for Cube.

The complete isolated-Xvfb run against CNA 756096626 reports 69/69 PASS with three completed update/draw/present cycles. Each 160x90 proof viewport differs from the clear colour and has non-flat luminance:

Asset Changed pixels Luminance range
Models/Car 605 483
Models/Windmill 79 422
Models/AlphaDeadTree 416 209
Models/Cube 613 308

Evidence is retained under evidence/cna-opengl33/milestone2/:

  • authentic-four-model-draw.log — complete 69/69 run and all 17 submissions;
  • authentic-four-model-draw.ppm — exact 320x180 readback, SHA-256 884d8089f8f4862476aeff3d7629b0721161641e303b2b610086bc94314151bd;
  • authentic-four-model-draw.png — review copy, SHA-256 99a9913c879e7ed39f4a62cdf9dfcc13409270b2cce11c7ab4888130c0f6697f.

Exit and next milestone

The four proof assets render with their authentic transforms and material assignments, so Milestone 2's exit gate is satisfied. Milestone 3 may now qualify one representative normal/specular Effect XNB and one multi-pass post Effect XNB before gameplay translation begins.