[#13516] improvement(ci): Split core tests into explicit database lanes

[nevzheng]

What changes were proposed in this pull request?

This pull request splits core testing into explicit unit, H2, MySQL, and PostgreSQL lanes so each lane has independent ownership, results, timing evidence, and coverage input.

It:

• adds dedicated Gradle tasks with single-backend selection and sequential database execution;
• preserves legacy :core:test behavior for local compatibility;
• records normalized test-identity manifests and reconciles the three database contracts;
• verifies that legacy execution is the exact multiset union of the split lanes;
• publishes separate JUnit, HTML, timing, identity, and raw JaCoCo artifacts; and
• extends [#13459] improvement(ci): Split build and Backend IT into sharded sub-jobs #13495 with six build shards and generates one combined core coverage report from raw execution data.

Because #13495 is still open, this pull request temporarily contains its exact head commits. After #13495 merges, the branch will be rebased so this pull request contains only the #13516 delta.

Why are the changes needed?

The existing :core:test task combines unit tests with three database backends. That makes ownership, timing, failures, and coverage inputs difficult to inspect independently, and it prevents safe follow-up concurrency work.

This change establishes explicit boundaries and fail-closed equivalence checks before further database-test optimization.

Fix: #13516

Related to: #13515, #13495, #13471, and #12584.

Does this PR introduce any user-facing change?

No. There are no production API, storage-format, or user-facing configuration changes.

The new gravitino.core.test.backend system property and Gradle tasks are test-only interfaces. Legacy :core:test remains available.

How was this patch tested?

Completed locally:

• python3 -B -m unittest discover -s dev/ci/tests -p 'test_core_test_identity.py' — 9 tests passed.
• ./gradlew :core:spotlessCheck :core:compileTestJava -PskipWeb=true -PskipDockerTests=true
• ./gradlew :core:coreUnitTest -PskipWeb=true -PskipDockerTests=true — 1,961 tests passed.
• ./gradlew :core:coreH2Test -PskipWeb=true -PskipDockerTests=true — 519 tests passed.
• Unit/H2 normalized identity intersection — zero.
• ./gradlew rat -PskipWeb=true
• bash -n dev/ci/test-shards.sh
• actionlint .github/workflows/build.yml
• Simulated merge of current upstream/main and the exact [#13459] improvement(ci): Split build and Backend IT into sharded sub-jobs #13495 head, followed by compilation and range-diff verification.

Appendix: local proof data

MySQL evidence collected so far:

• Full coreMySQLTest lane: 519 tests executed; 518 passed and one failed during unchanged fixture teardown after its functional assertions had passed.
• The teardown failure was a transient MySQL CommunicationsException / No route to host while dropping tables. Source and runtime inspection confirmed that the lane is serial and the affected test and teardown ordering are unchanged by this pull request.
• Focused rerun of the affected method: 2 tests passed in 34.845 seconds (BUILD SUCCESSFUL).

PostgreSQL evidence:

• Full corePostgreSQLTest lane: 519 tests executed; 517 passed, 2 skipped, and zero failed.
• JUnit duration: 1,961.648 seconds. Gradle task duration: 1,977.395 seconds (32 minutes 57.395 seconds).
• Normalized identity digest: 22844eb5cc3bf4e6a0d85f494bb0f4a4510e2be6a3451213ac2b5f32f382f035, matching the H2 lane.

TBD — four-lane identity reconciliation, legacy partition comparison, and combined JaCoCo evidence will be added when available.

Timing evidence

The comparable baseline is the successful
#13495 build (17, core) job.

CI measurement	Before: combined core tests	After: split core lanes
Core job wall time	49m35s	Pending
Gradle step	47m47s	Pending
:core:test task	45m13.251s	Replaced by four lanes

The #13517 workflows stopped during startup because of the ASF action allowlist, so a comparable after-CI measurement does not exist yet.

Local split-lane evidence:

Lane	Result	JUnit duration	Task duration
Unit	1,959 passed, 2 skipped	2m33.520s	2m38.064s
H2	519 passed	3m40.981s	6m13.684s
MySQL	518 passed, 1 teardown failure	1h09m08.839s	unavailable
PostgreSQL	517 passed, 2 skipped	32m41.648s	32m57.395s

The focused rerun of the MySQL teardown case passed both invocations.

Using the four local JUnit durations, the modeled serial total is 1h48m04.988s and the ideal concurrent critical path is 1h09m08.839s. This represents 38m56.149s, or 36.0%, of idealized local elapsed-time reduction—a 1.56× speedup. This is parallelization headroom, not an observed CI before/after result.