OOM situation investigation revealed that com.cloudbees.jenkins.support.impl.SlaveCommandStatistics#statistics is always growing. For instances with dynamically provisioned nodes this is effectively a memory leak. Detection:

println "Listeneres " + ExtensionList.lookupSingleton(com.cloudbees.jenkins.support.impl.SlaveCommandStatistics.class).@statistics.size()
println "Computers " + Jenkins.instance.computers.size()

Values from some of the instances with varying size/uptime:

Listeneres 1990
Computers 5
---
Listeneres 62
Computers 33
---
Listeneres 16521
Computers 16

The instance torn down had 30K instances.