mech-ai/tools/model-diagnostics/SectionBrepExtractor

SectionBrepExtractor

Temporary extractor for true section B-rep evidence.

Goal:

1. Use a chosen section plane.
2. Intersect SolidWorks bodies with that plane using official API hooks.
3. Convert the section result into a 2D B-rep-style evidence object: faces, edges, loops, and component provenance.

This is the layer that will eventually compare a textbook section image's lines against a model's section B-rep.

Run:

dotnet run --project tools\model-diagnostics\SectionBrepExtractor\SectionBrepExtractor.csproj -c Release -- "D:\path\model.SLDASM"

Optionally export model-view JPG images during the same run:

dotnet run --project tools\model-diagnostics\SectionBrepExtractor\SectionBrepExtractor.csproj -c Release -- "D:\path\model.SLDASM" --output-dir runtime\section_brep\trial1 --export-images

When --image-views is omitted, the tool exports seven baseline views: front,back,top,bottom,left,right,isometric.

Section B-rep and section image export are disabled in the current workflow. Use assembly/component B-rep, component highlight images, part-file images, and internal component group context images instead.

Image export is implemented in ModelImageExporter.cs, separate from the B-rep extraction logic in Program.cs.

Image evidence families are hard workflow boundaries:

• assembly_component_position: component position and whole-component function only. It has two visibility variants:
  • unobstructed (assembly_component_highlight_view): the assembly's existing visibility state is preserved and the target instance is highlighted in place. The exporter hides no component in this mode.
  • occluded_context (assembly_component_context_isolate_view): only geometry-derived occlusion blockers are hidden. Direct contact or mate neighbors are always preserved, and unrelated non-blocking components keep their original visibility. These are complementary renderings of the same evidence family, not component evidence versus interface evidence. Use the first whenever the target is already readable in the assembly; use the second only for an actually occluded internal target.
• assembly_physical_interface (assembly_physical_interface_view): physical contact/fit function only. Images isolate exactly two participating components and highlight the grouped interface faces. They must not be used as component-position or per-face part evidence. Planar interface facts are retained in the B-rep report, but ordinary single-plane contacts are not exported as physical_interfaces images because they are not fit-tolerance evidence. Planar interface images are only for multi-face/profile fits such as slots, keys, or prismatic locating contacts. Cylindrical interface images first require analytic cylinder axis, radius, and axial-range compatibility. The common axial range is checked in 2 mm slices; for each slice the actual angular intervals are obtained from the SolidWorks trimmed Face and its Loop/Edge boundaries, without face tessellation or triangle-grid overlap. One slice with a positive angular interval intersection is sufficient; no minimum overlap-area threshold is applied. Opposing face normals are still required so two exterior faces on the same theoretical cylinder are not treated as a hole/shaft fit.
• part_all_face_surface (functional_group_all_face_highlight): post-functional-group part and surface diagnosis only. These images must not enter component-function or interface-function analysis.
• part_feature_surface (part_feature_highlight_view): direct-part feature evidence only.

ImageKind remains the concrete renderer/output type for compatibility. EvidenceFamily controls which AI stage may consume an image. Do not route images by filename alone.

Outputs:

• section_brep_report.json: assembly/component B-rep facts, face/contact facts, FeatureGraph, SemanticFusionContract, optional ImageExport, and empty legacy section fields.
• section_brep_report.md: readable summary with exported image paths.
• model_images\*.jpg: SolidWorks standard-view and section-view screenshots when image export is enabled.

Supported image views:

• front
• back
• left
• right
• top
• bottom
• isometric
• trimetric
• dimetric

Supported section planes:

• front
• top
• right

Feature graph:

The extractor also emits FeatureGraph as the B-rep side of the image/B-rep fusion layer. The first version groups raw faces into:

• hole_group / hole_feature
• main_axis_cylindrical_surface_group
• cylindrical_surface_group
• planar_pad_step_or_face_candidate
• assembly contact features when contacts exist

Each feature keeps source face refs and lists VisualObservationNeeds plus RequiredBrepChecks so AI diagnostics can combine image cues with measurable B-rep facts instead of treating either source as absolute.

Semantic fusion contract:

The extractor emits SemanticFusionContract as the machine-readable instruction boundary for AI diagnosis. It does not generate final mechanical semantics in code. Instead, it defines:

• the required pipeline from B-rep facts and images to a global mechanical semantic graph, then to local semantic units;
• supported B-rep feature types;
• semantic assertion types such as part_identity, structural_semantic, functional_semantic, manufacturing_semantic, assembly_semantic, and risk_semantic;
• retrieval views, local semantic unit schema, B-rep fact bundle schema, confidence/status policy, missing-check fields, retrieval policy, and dedup policy.

At diagnosis time, the AI uses this contract plus FeatureGraph, raw B-rep facts, standard images, and section images to generate the actual mechanical semantic graph. Then it generates knowledge-blind LocalSemanticUnit objects: one neutral retrieval sentence plus concrete B-rep/image facts. Rule ids and final judgements are added only after knowledge retrieval and detailed rule verification.

Assembly diagnosis scope:

• For SLDASM input, diagnose both assembly-level relations and part-level structure.
• Standard or purchased components are marked review_scope=assembly_context_only; they are used for mating context, accessibility, and standard-part selection/calculation checks, but not for nonstandard part-structure taboo checks.
• Nonstandard or self-designed components are marked review_scope=part_design_required; AI must generate local semantic units for their local shape, surfaces, holes/slots/bosses/ribs, machining accessibility, stiffness, and other part-design views.

For assembly input, the report also emits two alignment structures:

• ComponentEvidencePackages: one package per SolidWorks component instance. It partitions assembly-derived B-rep faces, feature ids, and contact ids by component so evidence from different parts is not mixed.
• PartImagePlan: image-export plan for components with review_scope=part_design_required. These images may come from the component part file, but their geometry facts must still use the component's B-rep subset extracted from the assembly. Do not re-extract part B-rep independently and merge it into the assembly report.

Diagnosis order for assemblies:

1. Analyze assembly-level design issues first: contacts, fits, locating, support, disassembly, accessibility, motion/interference, sealing, and lubrication.
2. Then analyze each nonstandard/self-designed component's part-design issues using its ComponentEvidencePackage plus the matching PartImagePlan images.
3. Standard or purchased components remain assembly context unless the specific rule is about standard-part selection or calculation.