docs/verification.md

Verification status

September 17, 2026. The Space is public at https://huggingface.co/spaces/mantrakp/component-studio and both gated models are authorized. The container builds, native CUDA/PartField imports pass, model loading succeeds, and ZeroGPU startup reaches RUNNING. The first actual image-only request exposed a Transformers 5.17 DINOv3 encoder nesting change. The adapter patch preserves the upstream feature normalization and passes a CPU encoder equivalence check. End-to-end generation is still under verification; RUNNING is not proof of inference. The local development server remains stopped.

HF installs requirements before copying the application, so requirements.txt contains dependencies directly, excludes the local editable package, and includes HF's Gradio OAuth/MCP extras.

Verified locally:

• 16 tests cover input modes, ordered stages, face/UV/normal preservation, PBR channel retention, textured GLB round trips, refinement validation, credential redaction, failure artifacts, resource admission container resource limits, and explicit texture transfer across ZeroGPU process boundaries. Neural generation is explicitly replaced with test fixtures in orchestration tests, not in production.
• The portable uv lock is current and Ruff passes. Linux x86_64/Python 3.12 CUDA dependency resolution succeeds; requirements-hf.lock captures that environment. Dependency resolution does not prove that native CUDA extensions run on ZeroGPU.
• Browser checks cover all three clickable examples and the explicit unavailable-runtime error on macOS. No sample mesh is presented as a generated result.
• Blender rendered and reimported an actual textured test GLB with two components.
• The requested free Ling route was present in OpenRouter's live catalog. A real Ling vision request inspected those renders. fx built from the pinned source returned a schema-valid repair plan through that exact OpenRouter model. The repaired GLB reopened, and a subsequent Ling vision comparison accepted the test candidate. Evidence is in outputs/verification-refinement-v2/; it is a technical cube fixture, not model generation.
• fx's actual permission report confirms ask mode and a persistent wildcard deny rule in its dedicated workspace. No arbitrary agent tools are used by refinement.
• The actual Real-ESRGAN checkpoint produced finite 64×64 output from a 16×16 tensor on CPU. The native PartField scatter adapter matched independently specified grouped means.

Still requires the target GPU:

• TRELLIS.2 inference, PartField checkpoint loading/inference and concurrent CUDA SR.
• End-to-end ZeroGPU leases, actual VRAM profiling/OOM behavior and runtime duration.
• Visual quality on real assets across categories. No claim of benchmark superiority.

After uploading to a ZeroGPU Space with its secrets and gated-model access configured:

uv run scripts/verify_space.py --space OWNER/SPACE --image assets/stool.png
uv run scripts/verify_space.py --space OWNER/SPACE --prompt "A brass desk lamp with a green glass shade"
uv run scripts/verify_space.py --space OWNER/SPACE --image assets/stool.png --prompt "Keep the shape; use a walnut seat and black metal legs"

These commands use real GPU quota and the requested Ling endpoint. They download and validate the final GLB and archive, rather than interpreting a queued job as success.