I bisected just to be sure what happened. It was #51152 that broke #49929, and then #51762 fixed it, which is confusing as I thought that was just a crash fix.

So, maybe I just wait for #51771 to be merged before doing this.

I've modified this PR to instead bump package.json to the latest nightly + an override to make our dev tooling accept TS 5.0, followed by:

$ rm -rf lib && cp -r ./node_modules/typescript/lib . && git restore ./lib/.gitattributes

Which does basically the same thing that I used to do with git checkout v4.8.4 -- lib but for a nightly build.

This has the benefit of keeping everything in sync and letting tsbuildinfo not break every time LKG changes (since it encodes a bad version from main), at the cost of not being auto-updated for patches since that would be very noisy.

I'm curious what people's opinion is of that (it's in the same spirit as my git checkout-ing, so, I kinda like it, but I also feel like we could totally forego LKG altogether).

Thinking about this again, the main benefit is being able to use the new inference features in #49929 without breaking ts-eslint, since without bumping node_modules, it will use 4.9 and not have the new inference features.

I've added a task which automates installing a new version of the nightly into package.json/node_modules and copying it to LKG:

$ hereby bump-lkg-to-nightly
Using ~/work/TypeScript/Herebyfile.mjs to run bump-lkg-to-nightly
Starting bump-lkg-to-nightly
> npm install --save-dev --save-exact typescript@next

up to date, audited 353 packages in 626ms

found 0 vulnerabilities
Finished bump-lkg-to-nightly in 922ms
Completed bump-lkg-to-nightly in 922ms

We can of course update this out of band when needed, but to bring them back in sync again, this task can be used.

Eventually per #52199 we'll drop lib entirely, in which case this won't be needed anymore.

I'll redo this PR tomorrow once #51771 is in.

This PR should be ready for review; all needed PRs are merged in main and released in nightly.