SolidCraft: an open-source parametric 3D CAD application built in pure Rust. Desktop, web, and automation through MCP.
Starter resource validation
Date: 2026-10-21. Parent catalog checkpoint:
10ada1b45a4eb285ce752ac72f9aaa968ff8b5c4.
This record covers authoring inputs, numerical invariants, format decoding and reproduction. It does not claim a rendered SolidCraft scene or application-ready library. The independent renderer implementation remains a separate delivery.
Submitted payload
| Collection | Result |
|---|---|
| Surface recipes | 48 original resources, with 240 PNG maps at 512 × 512 |
| Surface exports | 23,879,602 bytes, including 16-bit authoring height |
| Environment recipes | 24 original illumination-only resources |
| Environment exports | 24 RGBE HDR inputs at 1024 × 512; 2,568,767 bytes |
| HDR dynamic range | 22 fields exceed unit radiance; maximum decoded channel 66.5 |
| HDR comparison fields | One exact unit-uniform field and one exact black-vacuum field |
| Lighting recipes | 24 records, 38 calibrated area sources, one finite-angular Sun and eight passive cards |
| Lighting availability | 19 numerical/unbound; three calibration-blocked; two require an existing Sun |
Authoring was checked on Linux with Python 3.12.14 and on Windows/Pegasus with Python 3.14.7. Both used NumPy 2.3.5 and Pillow 12.3.0. The submitted PNG encoding is the Windows output. These are authoring tools, not application dependencies. The source and export manifests contain exact SHA-256 receipts. The optional independent HDR comparison executes the installed FFmpeg program without reading or shipping its source.
Checks completed
- 53 focused checks passed: 21 surface, 13 environment and 19 lighting tests.
- All 240 PNG exports decoded and reproduced with identical pixel arrays. Hashes, dimensions, channel depths, color/data roles, paths, positive normal hemisphere and height/normal physical-scale agreement were checked.
- All 240 PNG pixel hashes matched between Linux and Windows. Compressed bytes matched for 112 maps; the other 128 differ only in encoding. The Linux encoding occupied 23,993,945 bytes, versus the submitted 23,879,602 bytes. PNG file receipts identify the submitted bytes; decoded-pixel receipts are the cross-platform reproduction authority.
- Surface boundary comparisons passed for all admitted patterns. Analytic tests cover both tangent axes, derivative magnitude, scale changes, multiple image resolutions, sRGB anchors, zero-height normals and full 16-bit PNG preservation.
- Both plain weave and the separate 2/2 twill pattern were checked for their actual periodic crossing sequence. Wrong cycle parity is rejected.
- All 24 HDR files regenerated byte-for-byte and passed strict full-stream decoding. All decoded pixels exactly matched the separate FFmpeg floating-point RGBE decoder. The largest absolute quantization error was 0.46539607; each error also satisfied its own shared-exponent quantization-step bound.
- All 24 HDR file hashes also matched across Linux and Windows.
- HDR analytic checks cover cardinal orientation, exact latitude solid angle,
pole/seam continuity, uniform-field
E = pi L, unit directions and black vacuum. - Lighting data regenerated byte-for-byte and matched across Linux and Windows. The worst higher-order area-quadrature difference was 0.01749%, below the declared 0.15% convergence threshold. This is convergence evidence, not a rigorous mathematical error bound.
- Independent lighting checks cover six projected disks, a closed-form directional sphere response, inverse-square far field, area/range cancellation under uniform scene scaling, one-sided emission, solar-disk integration, EV ratios, colored emitters and passive cards. Source-color luminance is counted exactly once.
- Hostile data cases include duplicate JSON keys/IDs, nonfinite values, booleans, very large integers, unknown fields, malformed types, unsafe paths, oversized inputs, corrupted PNG/HDR streams, false semantic metadata and stale receipts.
- Original catalog validator: 0 errors, 0 warnings. Prepared SELENE validator:
0 errors, with
awaiting_native_sourceandruntime_ready: falsepreserved. - Exact asset attribution and whitespace are checked with the pack validator and the repository checks. No review media, bytecode, cache or native model is part of the resource collection.
The unlit material sheet, HDR field sheet and numerical Lambert-sphere sheet were
inspected from ignored target/render-starter-review/. These are diagnostic
visualizations, not SolidCraft renderer output. The wood, cork and regolith
resources remain visibly procedural approximations; no photorealism or calibrated
species/location identification is inferred from those images.
Issues found and corrected during review
Independent review corrected periodic weave admission, separate twill semantics, invalid object/type handling, false map metadata acceptance, seam-evidence checks, oversized integer conversion and a colored-light test that counted color twice. The submitted tests retain coverage of those cases. All 53 checks also passed in the separate Windows environment.
Windows Git was observed converting hashed text sources to CRLF. Scoped
.gitattributes now fixes LF for the render authoring sources and catalog data,
and explicitly treats PNG/HDR as binary. JSON writers emit canonical UTF-8/LF
bytes. No global Git setting or other worktree was changed.
Eight convergence-diagnostic values differed between Linux and Windows by at
most 1.01e-15. Their export now reports 12 decimal places, while convergence
acceptance continues to use the unrounded result. A threshold-boundary regression
checks that reporting precision cannot let a failing calibration pass. Source
gains and transforms retain their existing precision.
Two LED/aperture recipes place their source inside/intersecting the opaque
0.25B calibration sphere. STU-018 and STU-019 therefore retain null unresolved
gains and an explicit requirement for an aperture/diffuser fixture. STU-024
retains its dependency on the actual existing Sun's response. The generator
does not assign invented numbers to make these records look ready.
Repository and runtime gates
Additional Pegasus verification on 2026-10-22 used a separate Windows worktree based on the frozen catalog commit, without changing the renderer implementation or other branches:
- 53/53 focused Python authoring tests passed again, along with the combined
manifest/PNG/HDR/lighting pack validator. The lighting numerical export hash
remained
52cac74c3b9ca2d66f0c36734336f498a4b8992d203d3305a9113a7753956617. - The repository attribution checker reported 725 attributed assets; Rust formatting and the targeted icon standards were also checked on Pegasus.
- The first
cargo xtask ciattempt reached and passed workspace Clippy, WASM checking and many test groups. The icon audit did fail on Windows because the existingrequested_icons.txthad been checked out as CRLF whereas the runtime snapshot compared the canonical LF bytes. The entries were identical after line-ending normalization; no artwork or assignments changed. - Added a path-scoped policy in the root
.gitattributesto pin the existing icon manifest to LF, restored its exact Git blob, and reran the failingrequested_menu_icons_have_artwork_and_match_manifesttest: PASS (1/1). The policy is recorded in root attribution and kept outside the closed icon-artwork directory, as required bycargo xtask icons. The complete aggregatecargo xtask cihas not been rerun after that fix; this draft must not be represented as having an end-to-end green repository CI. - FFmpeg was unavailable on Pegasus, so independent HDR decoding was not repeated on Windows. The separate original Linux independent-decoder and reproduction checks are the evidence for that claim.
These resource-specific tests are not automatically invoked by
cargo xtask ci. Both the portable authoring suite and the full repository
gate are required before a production merge.
Still required for an application release:
- Full repository gates on the integrated tree, including exact attribution.
- Typed resource binding to the actual renderer revision and its memory budget.
- Native CPU/GPU reference renders, common-exposure comparisons, normal handedness and filtering evidence, durable resource ownership and save/reopen checks.
- Reviewed near-field LED calibration and existing-Sun binding for blocked rigs.
- Actual scene geometry for environments that need parallax or terrain.
- Explicit supported optical/layered material variants and their acceptance tests.
- Improved SELENE native-source admission before any complete rover demonstration.
The authoring slice can be reviewed and reproduced now. Runtime availability remains false throughout the manifests.