You've certainly done your research...
From all of my experience with ansible what you're looking to accomplish, isn't supported. As you mentioned, ansible states that it does not require passwordless sudo, and you are correct, it does not. But I have yet to see any method of using multiple sudo passwords within ansible, without of course running multiple configs.
So, I can't offer the exact solution you are looking for, but you did ask...
"So... how are people using Ansible in situations like these? Setting
NOPASSWD in /etc/sudoers, reusing password across hosts or enabling
root SSH login all seem rather drastic reductions in security."
I can give you one view on that. My use case is 1k nodes in multiple data centers supporting a global SaaS firm in which I have to design/implement some insanely tight security controls due to the nature of our business. Security is always balancing act, more usability less security, this process is no different if you are running 10 servers or 1,000 or 100,000.
You are absolutely correct not to use root logins either via password or ssh keys. In fact, root login should be disabled entirely if the servers have a network cable plugged into them.
Lets talk about password reuse, in a large enterprise, is it reasonable to ask sysadmins to have different passwords on each node? for a couple nodes, perhaps, but my admins/engineers would mutiny if they had to have different passwords on 1000 nodes. Implementing that would be near impossible as well, each user would have to store there own passwords somewhere, hopefully a keypass, not a spreadsheet. And every time you put a password in a location where it can be pulled out in plain text, you have greatly decreased your security. I would much rather them know, by heart, one or two really strong passwords than have to consult a keypass file every time they needed to log into or invoke sudo on a machine.
So password resuse and standardization is something that is completely acceptable and standard even in a secure environment. Otherwise ldap, keystone, and other directory services wouldn't need to exist.
When we move to automated users, ssh keys work great to get you in, but you still need to get through sudo. Your choices are a standardized password for the automated user (which is acceptable in many cases) or to enable NOPASSWD as you've pointed out. Most automated users only execute a few commands, so it's quite possible and certainly desirable to enable NOPASSWD, but only for pre-approved commands. I'd suggest using your configuration management (ansible in this case) to manage your sudoers file so that you can easily update the password-less commands list.
Now, there are some steps you can take once you start scaling to further isolate risk. While we have 1000 or so nodes, not all of them are 'production' servers, some are test environments, etc. Not all admins can access production servers, those than can though use their same SSO user/pass|key as they would elsewhere. But automated users are a bit more secure, for instance an automated tool that non-production admins can access has a user & credentials that cannot be used in production. If you want to launch ansible on all nodes, you'd have to do it in two batches, once for non-production and once for production.
We also use puppet though, since it's an enforcing configuration management tool, so most changes to all environments would get pushed out through it.
Obviously, if that feature request you cited gets reopened/completed, what you're looking to do would be entirely supported. Even then though, security is a process of risk assessment and compromise. If you only have a few nodes that you can remember the passwords for without resorting to a post-it note, separate passwords would be slightly more secure. But for most of us, it's not a feasible option.
Best Answer
Since ansible is agentless it means it would need to open new connection (or re-use existent if configured properly) to host in order to restore the state.
If on the current step changes to firewall would not allow the next connection to succeed there's no way to recover obviously.
Thus it means you have to put testing-recovery task on the host itself which would try to apply new rules and then wait to see if next step of playbook has been succeeded (typically you need at least SSH accessible, so testing step of playbook could be just killing that script before it restores previous ruleset). It also can do some self-originating testing of course.
Summary: plant a script that would not only apply the rules but also would withdraw them back in case those rules would cause connectivity regression.
Straight-forward approach using
iptables-apply: