CI failing on master: NavigationTest#test_can_restart_bundle gets 500 instead of 2xx #41

Summary

master's CI is currently failing on every push/workflow_dispatch run. Discovered while investigating unrelated CI failures reported on the JRuby 10 upgrade PR (#40).

Failing test

NavigationTest#test_can_restart_bundle (test/integration/navigation_test.rb):

Failure:
NavigationTest#test_can_restart_bundle [test/integration/navigation_test.rb:15]:
Expected response to be a <2XX: success>, but was a <500: Internal Server Error>

The test POSTs a RESTART_PLUGIN node command (for org.apache.felix.shell) to a real Kill Bill server (killbill/killbill:0.24.0, started via docker/docker-compose.ci.*.yml) and gets a 500 back instead of a 2xx.

Evidence this is not related to any recent PR

Triggered workflow_dispatch directly on unmodified master (completely independent of any open PR) and it failed identically:

• Run: https://github.com/killbill/killbill-kpm-ui/actions/runs/34312886138
• Same failure signature, reproduces on both mysql2 and postgresql matrix legs, and with the old jruby-9.4.15.0 (i.e. pre-dates any JRuby 10 changes) as well as plain ruby 3.2.2.
• Re-running the same failed job (gh run rerun --failed) reproduces the identical failure every time — deterministic, not flaky.

Ruled out likely external drift as the trigger:

• killbill/killbill:0.24.0 Docker image: unchanged since 2023-01-03 (per Docker Hub API last_pushed).
• killbill/postgresql:0.24 / killbill/mariadb:0.24: last rebuilt 2026-08-19, which predates the last known-good CI run on this repo (2026-09-05), so not the trigger for a more recent regression.
• killbill-client gem: last released 2026-06-17, no new version to explain a recent regression.

What's still unknown

The actual Kill Bill server-side Java exception behind the 500 wasn't found in the CI job's captured Docker logs (the debug dump didn't include a matching stack trace in the searched window). Root-causing further would need direct access to Kill Bill server debug/trace logging around the node-command/plugin-restart subsystem, which is out of scope for this UI repo.

Impact

Blocks CI on master and on any open PR (e.g. #40), since all 4 matrix legs show red — 3 are actually just fail-fast cancellations of siblings once the first leg fails, not independent failures.