So my working model has changed. I will need to do some more research and report back.

When I count the tests that should run manually, I get 38 tests. However, the script seems to only run between 28 and 36 tests. What's particularly strange here is that I can call done_testing() and it works if I do it before I shut down an external etcd daemon used in the testing. I had previously thought the error was truncated output because done_testing didn't seem to work but this is more of an output order issue.

Running the test manually confirms a variable number of tests actually run before I come to an end-of-test condition. I now think the problem is coros being scheduled but not run yet But they seem to always run in the desired order, so I am confused here.

Not what you asked about, but possibly related to your scenario, if you are starting a background daemon using an object and expecting the object destructor to shut down the daemon when the object goes out of scope, you need to make sure to undef the object before the end of the script. At the end of the script Perl enters "global destruction" and just starts running destructors in any old order, which then tend to fail because they wanted to look at data that already got garbage collected.

I'll let you confirm or deny this possibility before I speculate further.