Spam filters work for lots of reasons but not because spammers can't figure out if they're caught or not.
If most people's willingness to pay for non-abusive communications media is $0, then... I expect they will continue to get abuse.
Shadow banning is evil and intransparent. It may work, but at what cost?
In either case, straight up ban or Shadow ban, someone is silenced as a punishment. In either case, the user likely isn't learning their lesson. In either case, the tools are ripe for abuse.
What is the exceptional cost of shadow bans besides them being opaque?
HN allows you to turn on shadowbanned posts.
In the several years I've been here, I've seen maybe 3 or 4 shadow banned posts that were of value. In those cases, other users replied to the shadow banned user and told them to message the mods about getting the shadow ban lifted.
Heavy moderation of the tone of message works. In a similar vein, the neutral politics sub on reddit is successful because they ban people who are rude.
I think that if we moderated tone more often, that a lot of the toxic communities wouldn't even form.
Shadowbans aren't akin to SMTP tarpits insofar as the target of force is less likely to be a bot. Behind the silence is a human being who is trying to engage others.
SMTP tarpits are like shadow bans because the perpetrator isn't informed about the ban/tarpit, and its meant to slow them down by wasting their time in their bubble that they are getting "work" done.
SMTP tarpits are more likely to involve a perpetrator who is operating a bot, and less likely a perpetrator who is personally engaged in the conversation and community.
Yet Hacker News isn't using these.
Because shadow banning, for example, works. Having someone's voice lost while they don't know about it (keep living in their bubble) works.
The lesson here: don't assume that your solution is the only solution.
Usenet, e-mail, and forums have taught us already that the solution to this problem is advanced, silent filtering (which have been innovated in the anti-spam arms race [1]) essentially akin to shadow banning. The enforcement of law also tends to help.
[1] RBLs, Bayesian filtering, and the already mentioned shadow banning