Hide posts until legitimacy of the poster has been verified. Allow them to post and respond, but don't show it to anyone yet, except to the moderators. If they're posting something sensible, unhide them. If it's spam, shadow-ban them. Don't let the user know. Let them guess why nobody is responding to the spam.
For that reason, it may also be a good idea to post an announcement that nobody should respond to this spam. Tat way the spammer won't know if he's being ignored manually or auto-hidden. Let him waste time and frustration on that.
Only use this against people who are this malicious. For regular hot-headed people who accidentally break forum rules but do want to meaningfully contribute to the community, always remain open and honest. Give people the opportunity to learn. Only people who are determined not to learn and to remain purely destructive, do you use the shadow ban.
Of course once they catch on, they'll probably start making new accounts with some legitimate posts, and once people start responding, they go back into spamming mode. This is tricky. Ideally, it'd be nice if you had a system that could automatically detect that sort of spam. If someone suddenly starts posting ascii art, bigimages, all caps, or anything like that, or goes on a rapid posting spree, automatically put them back on probation with hidden posts requiring approval, so you can check this change in posting behaviour.
For IP bans, they are now using illegitimately acquired or fraudulent IP space (the guy is not intelligent enough for this, he's almost certainly just buying proxies with in-game gold or some BS - but there is a criminal element here), including what might involve significant hijacking of AT&T, CenturyLink, Level3, and Windstream network resources.
(if you work at one of those places and are clueful, I would be very interested in asking about this)
With a troll this persistent (and willing to spend money on it), your best bet is definitely shadow bans and moderation queues.
If the troll is using residential proxies, you might try abuse@ the handful of services that offer such things. There aren't that many. I don't know if they actually take abuse seriously, but it can't hurt.
Luminati has an abuse form: https://luminati.io/report-abuse
Then move on to more advanced stuff, e.g. browser fingerprinting and behavioral pattern matching. If it's automated, there'll be a pattern. But the basic stuff can go a very long way.
As a bonus, you can put all new accounts into a single group, so that they would see each other posts, but without making these posts visible to the approved accounts.
Doesn't work for the VPNs that get you on residential IPs though.
There’s more stuff on the client you can do to prevent attackers from just hitting your APIs and forcing them to run a full client with having client solve sort a problem that it needs to provide to the API. You can detect headless browsers as well in JavaScript with a few open source bot detector JavaScript libraries.
I would advice sending a registration token as a clear picture to the user email (really short one, like four letters). Then the spammer need to do pattern recognition and if he can't program he will not bother.
The amount of legally blind users should be low enough that there can be a ad hoc queue of some sort. I guess any spammer can spam down that queue too though?
Maybe there could be a simple sound captcha for blind people with links to some audio files with letter in them to listen to in order. If the script kiddie figures out how to match byte streams add white noise to make it an DSP problem.
Also, with them using Selenium, I doubt that they can't put together some rudimentary pattern generation...
I adds complexity for the users and it is hard to know for user what is happening since it is a "secret" that someones messages doesn't show up. When the moderation process breaks down or there is a bug the user wont know how to appeal etc. It's the Google way of user interaction.
Just add a "This message will not be showed to others before your account has been approved" or whatever and it is not a shadow ban anymore.
The spam program might be shipped with Selenium. If the spammer can fix a trivial OCR text then OP knows that the spammer is a programmer not a "script kiddie".
Imagine a picture with four rows of letters and a rectangle around four letters. The rectangle moves around between registrations. Surely that would require a handwritten OCR to solve if the spam app ships with some boilerplate OCR? And be trivial to solve for humans with bad eyesight.
But they're registering hundreds of thousands of accounts. And, I assume, many of those accounts are creating posts. Moderating that many posts doesn't seem like a solution.
This can also be easily spotted and combatted by using a half-shadow-ban for accounts that are seen repeatedly browsing threads that just had a shadow banned post made; the user will not see random shadow banned posts, making detection even more difficult as you have to now cross reference multiple accounts.
You can make it arbitrarily hard to circumvent shadow banning.
* Don't delete ban accounts, don't notify them in any way, but tag their IPs and cookies to auto shadow-ban any sock puppets, so that these don't even make into an approval queue.
* Use heuristics to automate the approval process, e.g. if they looked around prior to registering, or if they took time to fill in the form, etc.
* Add a content filter for messages, including heuristics for an ASCII art as a first post, for example, and shadow-ban based on that.
* Hook it up to StopForumSpam to auto shadow-ban known spammers by email address / IP.
* Optionally, check for people coming from Tor and VPN IP, and act on that.
Basically, make it so that if they spam once, they will need both to change the IP and to clear the cookies to NOT be auto shadow-banned. You'd be surprised how effective this trivial tactic is.
All in all, the point is not to block trolls and tell them about it, but to block them quietly - to discourage and to frustrate.