test: adapt integration tests to Weaviate 1.37.5

[jfrancoa]

Motivation

Weaviate 1.37.5 introduces two intentional server-side behavior changes that break existing integration tests. This PR bumps the pinned WEAVIATE_137 version in CI from 1.37.1-4e61e26.amd64 to the released 1.37.5 and adapts the affected tests. Both failures stem from deliberate server changes (verified against the Weaviate source), not client bugs — the tests were asserting behavior the server no longer provides.

Approach

Two unrelated server changes drive the test updates:

1. maxWorkers removed from async replication config (server commit bb3d157004 "async replication scheduler", shipped in v1.37.3+). The server no longer accepts or returns this field, so max_workers now reads back as None. Rather than asserting on a field the server has dropped, the async replication tests now round-trip through hashtree_height (and propagation_concurrency), which are available on all async-capable versions. The client still exposes max_workers for backward compatibility with older servers — it's just silently ignored on 1.37.3+, so no client code changed.

2. Property name suffix _searchable is now reserved for internal indices (entities/schema/validation.go). Properties named field_index_searchable / nested_searchable are rejected by validation, so they're renamed to field_searchable_index / nested_searchable_index. The index_searchable=False behavior under test is unchanged — only the property names move.

The one test with a real semantic shift is test_replication_config_unset_single_async_field: after the maxWorkers removal, a single-field partial update now replaces the whole async config (unspecified fields revert to server defaults). The test is re-gated to >= 1.37.3 and rewritten with hashtree_height + propagation_concurrency to still exercise that "unspecified field reverts to default" invariant.

Key areas for review

• test_collection_config.py:test_replication_config_unset_single_async_field — the only test with changed semantics (config replacement on partial update); verify the version gate and the revert-to-default assertion still capture the intended behavior
• .github/workflows/main.yaml — version pin moves from a pre-release build hash to released 1.37.5
• The remaining async tests are mechanical max_workers → hashtree_height substitutions; safe to skim

Risks and mitigations

Low risk — this is test maintenance, no production client code changes.

• Older-server coverage: dropping max_workers assertions means we no longer verify max_workers round-trips on servers < 1.37.3 where it's still honored. Mitigated by the fact that the field is still exposed by the client and the substituted assertions (hashtree_height) cover the same async-config round-trip path across all supported versions.
• Version gate correctness: test_replication_config_unset_single_async_field is gated to >= 1.37.3 rather than 1.37.5 because the config-replacement behavior lands with the maxWorkers removal; running it on 1.37.3/1.37.4 is intended.

Testing

Full integration suite run against a live 1.37.5 cluster: 1162 passed, 88 skipped, 0 failures.

🤖 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

[jfrancoa]

Covered in https://github.com/weaviate/weaviate-python-client/pull/2042/changes