Interesting idea :) a bus factor >1 for tectonic-on-arXiv would be nice 😄
Doing this on GitHub's infra might also be nice for the "lots of parallel compute" aspect. The setup sounds quite tricky but getting it split across multiple workflow shards should be doable..

I'm not sure how the limits for GitHub Actions work exactly, but compute limits might end up being annoying here. It seems like GitHub comps 2000 compute-minutes per month. For reference: My somewhat aging system takes ~8 CPU-hours per run. If I understand correctly, the compute-minute limit is actually for the 4-core standard runners, so ~16 runs per month.

That limit is specifically for private repos on the free plan (see this section) - as a public repository, tectonic can use more (though we do want to be a little cautious about not going overboard, so we'll see).

I haven't gotten to a point where I'd try splitting it, but that's an interesting idea. Could be pretty simple as just a matrix job - then wait for all of them to complete for the final summary. At that point I maybe would want to split the dataset, maybe into bundles of ~1000. I'd have to give up the collective dataset to do that if we wanted to stay free, I'm using Cloudflare which has a free tier with limits that should be fine for us, but it caps at 10 GB storage before I start paying for it (access limits exist but are far beyond what we'd use, 10 million file GETs per month, and if it goes over that I'll just be out a couple cents, which, whatever). Alternatively I can try to slim it down - find the largest files and remove some of them, since I suspect some of the size comes from diagrams and images that, while it is nice to test, may be fine to skip at least some of.

Checking rates - I'll just store more data if need be. The cost per month for like, 50 gigs is totally reasonable.

1 hour 22 minutes for one runner to do it. Split into thousands, that should be reasonable speed. Alternatively, figuring out how to speed up document compilation even a bit should result in good wins :P

Either way, just need to get it to upload the report as an artifact, turn on pulling from master, and it should be almost ready. I do want to look into requiring it to merge but only running on request somehow, given the compute requirements.