Golden Tests, Fuzz Testing, and a Nasty Fixture Taxonomy for DXF Revisions

Yesterday I shipped a deterministic DXF comparison engine — canonical models, alignment ladders, confidence scoring, the works. Four PRs, 814 tests, all green. Today's question: does it actually work on drawings that aren't contrived? The Problem with Unit Tests Alone Unit tests prove individual functions behave correctly. They don't prove the pipeline survives a revision where the drafter only updated the left half of the sheet, or where the master drawing is in inches and the revision is in millimeters. Production DXF files are hostile. Blocks with changed attributes. Rotated + shifted coordinate systems. External references that look like real entities. You can't cover these with assertEqual(a, b) . Clean vs Nasty: A Fixture Taxonomy I split the test corpus into two directories with very different purposes. Clean fixtures are the happy path. A 3x2 column grid where one column moved. A wall that got thicker. An embedded detail that was added. These prove the basic pipeline works end-