ARCHITECTURE.md

ArchUnitJava architecture

ArchUnitJava analyses its own compiled main classes during the test phase. The self-test reads target/classes as class-file data; it does not load those classes or execute application code.

The enforced dependency direction is:

1. graph is the identity and dependency kernel and has no dependency on higher packages.
2. importer and model are one extraction cluster. They may collaborate and use graph, but do not depend on rules, reporting, or integration entry points.
3. projection depends on graph, not extraction, rules, reporting, or integrations.
4. rules is policy code. It may use extraction, projections, selectors, slices/layers, PlantUML input values, execution, and result values, but not reporting or integration entry points.
5. report formats graph/result/layer/slice values and does not reach back into extraction, projections, metrics, or rules.
6. cli and junit are integration seams. Core extraction, projection, policy, and reporting packages do not depend on them. The Maven/Gradle bridge package may depend internally only on cli; it receives compiled outputs and never starts another build.

The dogfood suite deliberately ignores dependencies whose targets are outside the imported ArchUnitJava main classes (for example JDK and JUnit APIs). It does not exempt any internal package or type. Every boundary is an ordinary public ArchitectureRule, adapted to an independent JUnit dynamic test through ArchitectureTestCases, so a violation uses the same structured result and failure rendering as a consumer's test.