The comment spam is likely a byproduct of RL, it lets the model dump locally relevant reasoning while writing code.

You can try asking it to not do that, but I would bet it would slightly degrade code quality.