cdsl-cad/json_to_cdsl

JSON to CDSL

This program converts SolidWorks export JSON into parameterized Learning IR (*.cdsl.json).

Conversion responsibilities:

• normalize SolidWorks JSON into an intermediate representation;
• classify sketches into named parameterized profiles;
• reuse equivalent profiles through semantic references;
• derive deterministic geometry through the sketch solver;
• remove avoidable floating-point noise;
• report unsupported or non-self-sufficient shapes.

The converter must not hide missing geometry behind an external context file when producing a training-ready CDSL artifact.

Evidence v2 batch converter

evidence_v2_to_cdsl.py converts the newer solidworks.cad_evidence.v2 extraction format into semantic CDSL v1.1. It does not modify the current rebuild engine: features without a runtime implementation are emitted with execution_status: "deferred" instead.

PYTHONPATH=backend/engine python json_to_cdsl/evidence_v2_to_cdsl.py \
  json_to_cdsl/input/evidence_v2 \
  --truth-dir json_to_cdsl/input/evidence_v2/truth \
  --out json_to_cdsl/output \
  --workers 1

For every *.solidworks_evidence_v2.json input it writes:

• <part_id>.cdsl.json: self-contained semantic CDSL in millimetres;
• <part_id>.diagnostic.json: source blockers, deferred features, optional inferred_from_step selectors, semantic validation, and STEP-versus-source bounding-box/volume/area/topology metrics;
• manifest.json: batch-level counts and paths.

The optional STEP directory is used only to infer a host face when exactly one candidate is compatible with a Hole Wizard's captured locations. Ambiguous geometry stays in unresolved; the converter never invents a selector.

Source records are discovered recursively. Output remains flat and is named by the normalized full source filename. Files that normalize to the same part ID receive a deterministic relative-path hash suffix, so every input record maps to a distinct output. When a matching STEP file exists, it is resolved through the same relative subdirectory as its source record.