It's all done sequentially (as in no parallel initialization tasks) so two runs of the same test should produce the same consistent outcome.

Still, not all elements in the sequence use a deterministic order rule, for example the order in which entities are discovered might depend on platform, JDK version, even filesystem being used as it simply will iterate on the elements maybe in alphabetical order, maybe inverse, or maybe one platform doesn't account for case, etc...

So if it's order related that might explain while it's very hard for Hardy to reproduce, still that doesn't explain why you have inconsistent runs. Do you have some parallelism during initialization of your stack? Any other process which might interfere with it?

Did you nothing anything unusual in the logs?