[refactor] Replace XSuite with a JUnit Platform engine, remove JUnit 4, move test support to the test-jar

[duncdrum]

Stacked on #6790 (JUnit 4 to 5 migration); review that first.

Summary

Part of #6037. After #6790 the only thing keeping JUnit 4 alive was XSuite, the runner for the XQSuite and XML tests (a JUnit 4 ParentRunner). This PR replaces it with a JUnit Platform engine, converts all 37 suites, then removes JUnit 4 and the Vintage engine from the build and bans them. It also moves the test support code (org.exist.test.*) out of exist-core's production jar into its test-jar, so test libraries stop leaking to consumers and into the server distribution.

What changed

1. Test libraries provided in exist-core, so the distribution stops shipping hamcrest, jupiter-api, platform-commons and opentest4j .
2. TestEvents sink. The six XQuery callbacks report to a framework-neutral interface instead of a JUnit 4 RunNotifier.
3. XQSuiteTestEngine and @XQSuite. One test per test function. A suite that cannot be discovered fails instead of vanishing; a test that was discovered but never reported fails instead of being silently dropped.
4. All 37 suites converted (24 in exist-core, 14 extension modules); XSuite deleted.
5. Parallel files and hang detection (below).
6. JUnit 4 removed from the build, with an enforcer ban (a new JUnit 4 test would otherwise compile but never run). exist-ant gets a small helper instead of Ant's JUnit 4 BuildFileRule.
7. Test support moved into the exist-core test-jar. The server fixtures, assertion helpers and test runners (26 files, moved with history) leave the production jar; the four test libraries become plain test scope. 15 modules gain the test-jar dependency (the two JMH modules at compile scope, since their benchmarks are main code), and exist-xqts gets it at runtime for the external XQTS runner, which also needs junit-jupiter-api and junit declared.
8. Review and CI follow-ups: Codacy PMD findings resolved, assertTrue(x instanceof y) converted to assertInstanceOf with the static imports tidied, and the identical junit-platform.properties removed from lucene and restxq (their copy and the one from the exist-core test-jar made surefire warn that only the first of two files is used).
9. JMH benchmarks compile again with -Pperf-tests. exist-core-jmh declared junit-jupiter-api with test scope and exist-indexes-jmh not at all, but the benchmarks (main code) use ExistEmbeddedServer, which now implements the Jupiter extension callbacks. CI does not build that profile, so it went unnoticed. Both modules now declare the API at compile scope, with the dependency analyzer ignores and the reason; checked with mvn compile dependency:analyze-only@analyze -pl exist-core-jmh,exist-indexes-jmh -am -Pperf-tests.

Same tests as the old runner

By console totals: exist-core 3,023 tests (39 skipped); the 14 extension modules 1,471 (61 skipped); 0 failures; identical module by module. Comparing the two found mismatches JUnit 4 had hidden, because it only counts tests that start: same-named test functions collapsed into one, hyphenated module prefixes, raw <task> text in XML test names, and modules that cannot be compiled without a database. All fixed.

Parallelism: what we found

Update: #6796 (stacked on this PR) addresses these findings: it gives test files their own collections and makes CoreTests and XQuery3Tests parallel, and records what still blocks the other suites.

@XQSuite(parallel = true) runs a suite's files concurrently (tests within a file stay sequential), capped by exist.xqsuite.parallelism. It is opt-in, because parallel is only safe for suites whose files do not share database state, and almost none are:

Suite	Under parallel = true
NumericOpTests (12 files, no database access)	Stable: 40 runs, 72/72 every time, up to 10 files at once. Now enabled.
CoreTests	147 to 150 failures
RangeTests	63 to 67 failures
LuceneTests	2 to 5 failures and errors per run
XQuery3Tests	files drop and recreate the same collection

• There is almost nothing to gain. The safe suites finish in about a second. And 150 of XQuery3Tests' 157 seconds are one file (transform/catalog.xql); the other 100 files take about 7 seconds together, so per-file parallelism cannot speed it up. Why that one file is slow is a separate question.
• NumericOpTests is enabled to keep the feature exercised on a real suite in CI, not for speed.

Hang detection

Applies to every suite. A file that reports nothing for exist.xqsuite.hang.threshold.minutes (default 5) is failed with its name and the tests that were running; the others carry on (the old runner cancelled everything). The old runner's test never actually hung anything; the new ones use fixtures that really do.

Limit: giving up on a hung file is reliable, stopping its query is not. eXist only stops a query where it checks for being killed, and some loops never do. Such a thread keeps the database from shutting down, so the engine bounds the shutdown, fails the suite and skips the ones after it, instead of hanging as before.

Behaviour changes

• XSuite and @XSuiteFiles are gone; use @XQSuite from the exist-core test-jar.
• org.exist.test.* (ExistEmbeddedServer, ExistWebServer, ...) is no longer in the exist-core jar; external users need the exist-core test-jar (<type>test-jar</type>). The only external user found, the XQTS runner, is covered through exist-xqts.
• Surefire now groups results under the suite class (for example XQSuiteTests) instead of xqts.org.exist-db.... Test names are unchanged.
• Settings are JUnit configuration parameters: exist.xqsuite.parallelism, .hang.threshold.minutes, .hang.watcher.interval.seconds, .hang.grace.seconds.

Not in this PR

• xmlunit-legacy: tests in exist-core, lucene and ngram still use XMLUnit 1 assertions, which extend JUnit 3's Assert, so those three modules keep junit:junit at test scope (the one documented exception to the ban). Migrating about 100 assertions changes comparison semantics and deserves its own review.

Test plan

• Full reactor verify passes (compile, license, dependency analyzer, JUnit 4 ban)
• All 37 suites: same counts as the old runner, 0 failures, re-verified from a clean build after the move (macOS)
• exist-core integration tests and the XQTS runner's start-up checked on the moved code (isolated build)
• 22 engine tests: scheduling, hang handling, bounded shutdown, settings
• Test and documentation workflow green on Linux, macOS and Windows
• XQTS workflow green
• Rebased onto develop at 9c639a0ce6 on 2026-10-05 (no conflicts; VectorSearchTests keeps the query-vector-k.xqm that Fix ft:query-vector k being ignored / throwing past 341 docs #6741 added); reactor test-compile, and the Lucene module (951 tests) and the engine, runner, CoreTests and XQuery3Tests runs (2106 tests), all with 0 failures, run at the top of the stack ([test] Isolate XQSuite test files, run CoreTests and XQuery3Tests in parallel #6796)
• mvn compile dependency:analyze-only@analyze -pl exist-core-jmh,exist-indexes-jmh -am -Pperf-tests passes

close #6037

🤖 Generated with Claude Code

https://claude.ai/code/session_01KqkMrHZx1uvZFuwvMrc3gj

[github-actions]

📊 XQTS result comparison

Comparison of this run against develop.

Metric	develop	this run	Change
➖ Passed	28,867 (90.73%)	28,867 (90.73%)	0 (0.00 pp)
➖ Failures	1,547	1,547	0
➖ Errors	134	134	0
➖ Skipped	1,267	1,267	0
🧪 Total tests	31,815	31,815	0

Relative to develop: 0 newly passing, 0 newly failing, 0 new errors, 0 newly skipped — counting only tests recorded in both runs whose outcome changed.

_{Runtime: 439.7s (-60.66s vs develop).}

[line-o]

Thank you very much!