Hi, sorry for the delay in answering your question.
While we didn’t investigated the minimum system that runs the full node, let me tell you that we use m4.large with 2vCPUs and 8GB memory for our testnet and load test environments.
The basic node I have been able to run it with 1vcpu, as low as 1GB of ram with lean30 for mining and mining set to off (it will still generate a block per generation).
The suggested one is at least 4GB ram for the mean algorithm or it will run out of memory.
So I could pretty much tell you, you may be able to run the node on commodity specs, but it will take you several configuration changes and it may fill up the harddrive with the blockchain when it gets higher transaction volumes
to the contrary, with stopped mining, the node builds up a candidate, which is cheap operation - or at least, there is small overhead over validation - but it doesn’t start PoW
Did you observe different behaviour? We may have a bug then.
Up to versión 24 I saw the node generating candidates and still taking 100 of 2 cpus. so I assigned only one cpu and left mining to autostart. I thought that was default behaviour