[magnusp]

Fixes #117

Three-layer contract. Hrana 3 is the source of truth for transport, values, metadata, batching, transactions, and error reporting; java.sql is the source of truth for the JDBC surface. xerial/sqlite-jdbc is a behavioral reference only. This PR runs the reference and does not assert parity with it.

Base 1bed15e4c8a5be7fa47fac173eada95ad3b0a7e5 · Head 9c9b66daa3749e637925e7fdb482eaa97c2e24e8

What changed

• .github/workflows/ci.yml and .github/workflows/pr.yml: new Xerial Control (reference probe) job that runs mvn --batch-mode -pl libsql-conformance -am -Pxerial-control verify. The existing Build & Test (Java 25|26) job names are unchanged.
• libsql-conformance/pom.xml: a generate-test-sources-phase guard fails when conformance.control.driver is set outside the xerial-control profile. The profile sets an internal marker, so the profile run is unaffected. The guard shares the phase of the existing vendored-upstream check, so mvn validate still resolves no test dependencies; binding it to validate made the CI Validate reactor step fail before the other modules were built.

The control run is a reference probe: a failure means a vendored test is not true for sqlite-jdbc itself (or the driver diverges without a recorded reason), not that libSQL must match xerial.

Red / green

• Red (guard absent at trunk): ./mvnw -pl libsql-conformance -Dconformance.control.driver=xerial generate-test-sources → exit 0. No workflow passes -Pxerial-control.
• Green (guard present at head): same command → exit 1 with conformance.control.driver is only set by the xerial-control profile. Run with -Pxerial-control instead....
• Green (probe): ./mvnw -pl libsql-conformance -am -Pxerial-control verify → 100 tests, 0 failures, 1 skipped, BUILD SUCCESS.

Lanes

1. Regression: behavior absent at trunk, present at head. Pass.
2. Control profile locally: 100 tests, 1 skipped. Pass.
3. Guard failure: BUILD FAILURE with the message above. Pass.
4. YAML parse of all 4 workflow files. Pass.
5. ./mvnw -pl libsql-jdbc test: BUILD SUCCESS. Pass.
6. Workflow diff vs trunk: only the new job. Pass.
7. Jobs: trunk set preserved; one job added per workflow. Pass.
8. ./mvnw -q validate: BUILD SUCCESS. Pass.
9. Control job log in PR checks: Xerial Control (reference probe) ran all 9 selected org.sqlite classes (100 tests, 0 failures, 1 skipped), BUILD SUCCESS.
10. Required checks: Build & Test (Java 25), Build & Test (Java 26) still the required contexts; both pass, and all PR checks are green.

Perf: ./mvnw -pl libsql-jdbc test median 2.08 s at trunk vs 2.06 s at head (3 interleaved runs each), no Java change.

Divergences

• pr.yml is outside the plan's Files list. Lane 9 requires the control job log in this PR's checks, and ci.yml only runs on pushes to main. pr.yml is a near-identical copy of ci.yml, so it gets the same job.
• mvn --batch-mode instead of mise exec -- ./mvnw. CI uses actions/setup-java and never installs mise; both existing workflows call mvn --batch-mode, so the new job matches them.

Evidence logs: /tmp/xerial-parity-i117/.