Instead of trying to prevent LLM training on our code, we should be demanding that the models themselves be freed.

You can demand it but it's not an pragmatic demand as you claim. Open weight models aren't equivalent to free software, they are much closer proprietary gratis software. Usually you don't even get access to the training software and the training data and even if you did it would take millions of capital to reproduce them.

But the resulting models must be freed. Any model trained on this code must have its weights released under a compatible copyleft license.

You can put into your license whatever you want but for it to be enforceable it needs to grant licensee additional rights they don't already have without the license. The theory under which tech companies appear to be operating is that they don't in fact need your permission to include your code into their datasets.

block the crawlers, withdraw from centralized forges like GitHub

Moving away from github has become a good idea since Microsoft has purchased it years ago.

You kind of need to block crawlers because of you host large projects they will just max out your servers resources, CPU or bandwidth whatever is the bottleneck.

Github is blocking crawlers too, they have restricted rate limits a lot recently. If you are using nix/nixos which fetches a lot of repositories from github you often can't even finish a build without github credentials nowadays with how rate limited github has become.