# Route 1 Static Resource Audit

Date: 2026-05-09

Checked paths:
- `/data/share/ud4scenesmith/static/hssd_scenesmith_official_preprocessed_20260509`
- `/data/share/ud4scenesmith/static/partnext_semantic_home_subset_20260509`
- `/data/share/ud4scenesmith/static/scenesmith_partaware_static_objaverse_20260509`
- `/data/250010098/SceneSmith_official`
- `/data/250010098/particulate`

SceneSmith official setup summary:
- HSSD models: clone `hssd/hssd-models` with Git LFS into `data/hssd-models`; this provides the required `objects/` GLB tree.
- HSSD preprocessed data: run `scripts/download_hssd_data.sh`; this downloads HSM `data.zip` for CLIP embeddings/index and `support-surfaces.zip`.
- Objaverse static flow: run `scripts/download_objaverse_data.sh`, then `python scripts/prepare_objaverse.py`.
- Articulated flow: use preprocessed ArtVIP assets or run `compute_articulated_embeddings.py` after SDF conversion.

Findings:
- HSSD preprocessed index exists: `clip_hssd_embeddings.npy`, `clip_hssd_embeddings_index.yaml`, `hssd_wnsynsetkey_index.json`, `object_categories.json`.
- HSSD support surfaces exist: 5187 top-level support-surface directories.
- HSSD full mesh objects are missing from the checked static root: `hssd-models/objects` does not exist.
- Therefore full `HssdRetriever.retrieve(...)` cannot load a mesh yet, but index/surface-only checks pass.
- Particulate local repo remote is official: `https://github.com/RuiningLi/particulate.git`, HEAD `9cf7f61`.
- Particulate worktree already had local modifications in `infer.py`, `particulate/models.py`, `particulate/evaluation_utils.py`, and `PartField/partfield/model_trainer_pvcnn_only_demo.py`; I did not edit them.

Small tests performed:
- Parsed HSSD `.npy` header without numpy: shape `(14099, 1024)`.
- Parsed HSSD JSON metadata: 481 WordNet categories, 10969 entries.
- Counted object category mappings: large 207, wall 70, ceiling 10, small 201 WordNet keys.
- Verified two PartNeXt shared GLBs and one Particulate example GLB have `glTF` magic headers and copied them into this portal page.

Artifacts:
- `route1_static_audit_results.json`
- `assets/partnext_000_162_5259596d645145e1ae62fc2b5d883101.glb`
- `assets/partnext_000_068_a1f4cd3fbcff48949116861a2ff6a9ba.glb`
- `assets/foldingchair.glb`

Blocking next step:
- Download/attach official HSSD `hssd-models/objects` before running a real SceneSmith HSSD mesh retrieval smoke.
- Install or repair a dedicated Particulate environment before running model inference; use `env -u LD_LIBRARY_PATH` because the global library path can break or stall conda Python startup.