First, thanks to @weston.smith for assisting with JIRA, I have created an account there as I’ve had other needs for it.
For posterity – I did find a solution to this. It turns out that this was in part caused by bad code that was flooding couchbase with n1ql queries. Now, the code that was connecting to couchbase shouldn’t have been doing this - but at the same time I would maintain that it’s not proper behavior for the system to overrun an explicitly set memory limit. If the trouble is flooding I would prefer to see something like a query timeout, or an outright rejection if the server is too busy to respond. Growing in memory until the whole thing crashes isn’t such a great behavior.
The message “unknown n1ql query error, this is usually tied to a memory condition” that I was seeing in the node SDK was a bit of a red herring. Docker stats indeed does indicate that it’s a memory problem, but that’s not really the root cause.