ci: run integration tests against Weaviate 1.39.0-rc.0

[jfrancoa]

Summary

Moves the integration-test matrix from $WEAVIATE_136 (1.36.12) to a new $WEAVIATE_139 (1.39.0-rc.0) and adds the version to the test-package matrix.

This is split out of #1991 so that the version bump and its fallout can be reviewed on their own, rather than mixed into the alter-schema drop-vector-index feature. #1991 is stacked on top of this branch.

Why now

The drop-vector-index endpoint only exists from 1.39, so its tests cannot run in CI until the matrix moves. Independently, the matrix has been sitting on 1.36.x while $WEAVIATE_137 was already defined, so a range of version-gated tests never actually executed — 56 more tests run at 1.39 than at 1.36.12 (1072 passed vs 1016).

Pinned tag

There is no plain 1.39.0-rc.0 tag on Docker Hub yet, only commit-suffixed builds, so the newest one is pinned following the existing WEAVIATE_137 convention. Worth swapping for the clean tag once the RC is published.

Test change

One existing test needed adapting to an intentional server change. From 1.39 asyncEnabled is no longer stored per class but derived server-side as factor > 1 && !globallyDisabled (adapters/handlers/rest/restcompat/wrappers.go). Three assertion sites use a replication factor of 1 and therefore now expect False.

Rather than sprinkling version checks, the scattered is_at_least(1, 26, 0) branches are replaced by one _expected_async_enabled(version, factor) helper that encodes both the old and new server behaviour. The two sites using factor=2 are unaffected — factor > 1 still yields True there — and are left alone.

No coverage is removed: the pre-1.39 expectations remain asserted on older servers.

Verification

The full integration suite was run against 1.36.12 and against 1.39.0-rc.0 on an otherwise identical local stack built from ci/docker-compose.yml. Both produce identical failure sets (109 entries, all from compose instances not available locally), so the bump introduces no regressions in the ~1000 tests that environment can exercise.

Two caveats for reviewers:

• Tests requiring the other compose instances (RBAC, cluster, proxy, okta, backup, namespaces) failed identically at both versions and were effectively not compared. CI is the first real check on those.
• The comparison was run on ARM using the .arm64 image variants; CI uses the .amd64 tags pinned here.

🤖 Generated with Claude Code

[orca-security-eu]

Orca Security Scan Summary

Status	Check	Issues by priority
Passed	Infrastructure as Code	0   0   0   0	View in Orca
Passed	SAST	0   0   0   0	View in Orca
Passed	Secrets	0   0   0   0	View in Orca
Passed	Vulnerabilities	0   0   0   0	View in Orca