mansart-persistence/mansart-jpa-bench is a standalone Maven project, intentionally outside the normal reactor. Its classpath-only provider dependencies are test-scoped and isolated by provider profiles; none enter a production Mansart module. The runner activates exactly one provider profile per run and builds it in its own directory, target/<provider> (POM property bench.build.directory, default target), so cleaning one provider never wipes another’s build.

The JMH fixture uses the same H2 in-memory database, manually-created schema, 128 seeded rows, entity class, @Version policy and resource-local transaction boundaries for Mansart, Hibernate ORM and EclipseLink. Shared and provider-specific second-level caches are disabled; Jakarta Validation mode is explicitly NONE for every provider. Each measured operation creates a new entity manager, preventing first-level-cache hits from standing in for SQL. The workloads are primary-key find, JPQL query/materialization (32 rows), versioned update and insert/update/delete in one transaction. The timed update writes a label derived from the next @Version value and returns the persisted version, so every invocation is a dirty, versioned UPDATE rather than a no-op commit. Before timing, setup verifies exact database row and query counts, exercises insert/update/delete, and runs the timed update method twice on the same row, checking through plain JDBC that both calls changed the label and incremented the version. JMH runs with -foe true, so any setup, preflight or benchmark failure stops the run with a non-zero exit (JMH’s default would only report the error and exit 0). JMH reports throughput, iteration confidence intervals and allocation via its GC profiler.

Build/install the provider and processor from the Mansart reactor, then run each provider separately with an explicit Temurin 25:

Every run stores its raw evidence, never cleaned by the runner and never overwritten, in target/bench-reports/<run-id>/<provider>/: jmh.log, jmh.json, maven.log, java-version.txt, test-dependencies.txt and test-classpath.txt. BENCH_RUN_ID groups the three providers of one comparison; it defaults to a UTC timestamp.

Versions and the full measurement protocol are fixed in the standalone benchmark POM and runner. These runs establish comparative observations on Java 25; they are not vendor certification claims. Raw scores, hardware, JVM, commands, and comparison notes are recorded in mansart-persistence/BENCH.md. No performance winner is claimed unless the measured evidence supports one.

Run mansart-persistence/mansart-jpa-processor-module-it/run-aot-smoke.sh with Temurin 25 and Maven to build and launch a small named application module using APT-generated managed access. The smoke asserts CRUD and JPQL results, checks the generated service provider is part of the application module, and fails if the entity package is opened or exported to Mansart. It creates a JDK Leyden AOT cache, reruns the same modular application with that cache, and checks the startup log reports cache use. Before recording, the script deletes only its own named cache (target/mansart-jpa-smoke.aot) and logs, so a stale cache can never satisfy the checks; each JVM writes to target/aot-record.log or target/aot-cache-use.log, and a non-zero JVM exit prints that log and fails the script with the same status. Use an explicit Temurin 25 JAVA_HOME. This demonstrates a Leyden AOT cache and is not a GraalVM native-image proof.