Normalize every reference view into a common axis contract:
- front: width X, height Z;
- side: depth Y, height Z;
- top: width X, depth Y;
- back: width X, height Z.
Measure visible body bboxes, not labels/background/aura unless they are part of the model contract.
Compare shared axes:
- front width vs top width;
- front/back height vs side height;
- side depth vs top depth.
If ratios differ beyond tolerance, write a conflict report and select a policy:
- front_side_canonical: front brand read and side depth/height are authoritative;
- front_top_canonical: front brand read and top footprint are authoritative;
- corrected_templates_required: no rigid rebuild should claim final fit;
- view_specific_cheats: separate camera/view variants allowed, not one rigid asset.
Feed the chosen policy into fit-repair-optimizer before any geometry, UV, or lighting work.

Hard gates

Do not restart modeling until canonical policy exists.
Do not treat top-view aura/context as body depth.
Do not average contradictory dimensions unless the user explicitly asks for an approximate stylized compromise.
Record part-count disagreements separately from bbox disagreements.

Script

scripts/view_constraint_report.py reads a JSON file of per-view body bboxes and emits shared-axis compatibility and a suggested canonical policy.