Performance regression tests

The release benchmark uses deterministic generated ontologies at three representative sizes. The fixture definition is versioned in performance/fixtures-v1.json. Every generated graph combines a branching taxonomy with existential restrictions, disjointness, inverse/transitive/symmetric object properties, data properties, and named individuals; the medium and large graphs also include equivalent definitions and general class inclusions. Each fixture measures the same six existing paths:

• isolated project/QC snapshot capture;
• HermiT class-hierarchy and class-assertion reasoning;
• asserted SPARQL cache snapshot construction and Turtle serialization;
• Jena SHACL validation;
• asserted semantic diff; and
• verified serialization with an isolated round-trip reload.

performance/materialization-v1.json separately pins the 0.8 materialization ceiling: 50,000 produced axioms, at most 512 MiB sampled incremental peak heap, a 500-axiom live batch and 100 ms model-thread stall ceiling, five-second measured cancellation effectiveness, and the common 2.0x/250 ms regression rule. MaterializationScaleTest exercises the complete 50,000-axiom category and records completeness, elapsed time, sampled peak heap, and cancellation evidence. MaterializationLiveScaleTest invokes the production Swing EDT gateway and exact Protege adapter commit core, including authorization rechecks, collision analysis, revision fingerprints, listeners, history, and the single-Undo assertion. The fixture uses a test OWLModelManager adapter; the separate live integration harness covers packaged Protege/OSGi wiring.

performance/baseline-v1.json records the reference measurements, reference machine, warm-up/sample counts, and the allowed regression factor. Fast operations also receive a small noise floor. These values are regression guards, not hardware-independent latency promises or end-user service-level objectives. Changing a fixture or reference requires a reviewed baseline update with a retained measurement artifact.

The opt-in benchmark is excluded from normal mvn verify. Run the same enforced gate locally with JDK 17:

mvn -B -pl plugin -am \
  -Dtest=PerformanceBaselineTest,MaterializationScaleTest,MaterializationLiveScaleTest \
  -Dsurefire.failIfNoSpecifiedTests=false \
  -Dprotege.performance=true \
  -Dprotege.performance.enforce=true \
  test

The machine-readable results are written to plugin/target/performance-results.json, core/target/materialization-performance-results.json, and plugin/target/materialization-live-performance-results.json. GitHub Actions runs all three gates weekly and for every release, retaining all result files so a failure can be compared with prior measurements.