Automated delivery stalls across both clusters

The component that reconciles deployments ran out of memory and restarted roughly every thirty seconds while managing around two hundred applications.

While it was down nothing was delivered: eleven applications drifted out of sync and in-flight operations hung having applied nothing, which is not obvious from the outside because the applications themselves stay healthy.

Its memory ceiling was doubled. Normal use is well under the old limit; the ceiling was only reached during bulk activity. The change is now in the configuration repository rather than applied by hand, so a redeploy will not undo it.