This package produces an auditable CDSL dataset from official public Onshape Part Studio URL mappings. It never treats a feature tree as sufficient proof: each accepted record has its complete raw API evidence, original AP242 STEP, CDSL-only rebuilt STEP, and an independent strict geometry report.

scan -> select -> download -> convert -> validate -> dataset/<id>.cdsl.json

Run from the repository root. The only runtime path setup needed is the engine and this source package; no new packaging tool is introduced.

export ONSHAPE_ACCESS_KEY='...'
export ONSHAPE_SECRET_KEY='...'
export PYTHONPATH=backend/engine:onshape_to_cdsl/src

python -m onshape_to_cdsl scan --url-file /path/to/objects.yml --limit 500 --workers 1
python -m onshape_to_cdsl select --count 100 --seed 0
python -m onshape_to_cdsl download
python -m onshape_to_cdsl convert
python -m onshape_to_cdsl validate

To find more than one URL range yields, scan non-overlapping batches and merge them before deterministic selection:

python -m onshape_to_cdsl scan --url-file objects_0000.yml --offset 0 --limit 500 --out data/scan/00000.json
python -m onshape_to_cdsl scan --url-file objects_0000.yml --offset 500 --limit 500 --out data/scan/00500.json
python -m onshape_to_cdsl merge-scans --scan data/scan/00000.json --scan data/scan/00500.json --out data/scan/merged.json
python -m onshape_to_cdsl select --scan data/scan/merged.json --count 100 --seed 0

For hosts that limit foreground process duration, download the deterministic selection in resumable batches. Existing artifacts are reused by SHA-256:

python -m onshape_to_cdsl download --offset 0 --limit 3
python -m onshape_to_cdsl download --offset 3 --limit 3

Or run all stages in sequence:

python -m onshape_to_cdsl run --url-file /path/to/objects.yml --scan-limit 500 --scan-workers 1 --count 100 --seed 0

--url-file accepts id URL, id: URL, and the same-line ID/URL records in ABC objects_*.yml. URLs are deliberately not hard-coded in the repository. Credentials are read from ONSHAPE_ACCESS_KEY/ONSHAPE_SECRET_KEY or hidden terminal prompts, and are never written to files or logs.

The official ABC mapping archive is published by NYU as abc_objects_00-49.7z and abc_objects_50-99.7z. Each contained objects_*.yml file has 10,000 URLs. Start with --scan-limit 500; the pipeline stratifies those API-inspected candidates and selects 100 complete downloads. The client honors 429 Retry-After responses; use one scan worker unless the API quota explicitly permits higher concurrency.

All generated data lives in ignored onshape_to_cdsl/data/:

scan/manifest.json          feature classifications for all source URLs
selection/manifest.json     deterministic, stratified selected candidates
raw/<id>/                   complete public API responses, meshes, STEP and hashes
converted/<id>/             candidate CDSL, rebuild STEP, conversion/validation reports
dataset/<id>.cdsl.json      only strict-validation accepted CDSL
rejected/<id>/              candidate/rebuild/report evidence for every failure
reconstruction_issues.md    generated rejection and engine capability work queue
reconstruction_issues.json  machine-readable version of the same register

The v1 executable converter intentionally supports only solved newSketch geometry composed of lines, arcs and circles plus standard, one-direction, no-draft BLIND extrudes (NEW, ADD, REMOVE). The registry audits all other feature types. Unsupported standard operations are deferred; custom FeatureScript, imports and external/derived dependencies are not_admissible. Their raw feature trees, FeatureScript payload, B-rep and STEP remain in raw/, but they cannot enter dataset/.

Admission uses a separate comparator, not the engine's preview comparator. It requires bidirectional B-rep surface sample maximum and P99 distances at most 0.01 mm, six-coordinate bounding-box error at most 0.01 mm, volume and surface-area relative error at most 1e-5, and equal solid counts. Topology counts/types are report-only diagnostics because kernel exports may split equivalent faces differently.

validate regenerates reconstruction_issues.md on every run. It classifies each failure as a source dependency limitation, converter gap, engine capability gap, engine execution failure, or strict geometric regression, and adds engine-specific blockers to an explicit implementation work queue.