Move update job from GitHub

[mpadge]

To some hosted GPU instances. https://salad.com seems good, as it uses residual compute from a heap of machines, including private ones. That then offers much smaller and cheaper GPUs than most, which is fine for this task.

TODO

• Build full Docker image that downloads all current data
• Rewrite Dockerfile here to have a single entry point that runs the update function and uploads data to this repo when done

[mpadge]

Hmmm,. can't build a full image with downloaded data because the GitHub runners error with message:

System.IO.IOException: No space left on device : '/home/runner/actions-runner/cached/_diag/Worker_20251106-100248-utc.log'

I'll have to remove the data download line, and just leave it with full install of repo plus ollama models.

[mpadge]

Hmm, even removing that and just leaving install of pkgmatch and ollama gets things a bit further, but in the final "exporting to image" stage still ends up with:

ERROR: failed to build: failed to solve: ResourceExhausted: mount callback failed on /tmp/containerd-mount2361214157: failed to write compressed diff: failed to create diff tar stream: failed to copy ... no space left on device

Have to also remove the ollama install bit, and also have that as an entrypoint command

[mpadge]

That latest commit restructures the whole container to build directly on r2u (830MB), rather than pkgcheck (11GB). A local test build with ollama install gave a container of 4.4GB, so still plenty of space within the GitHub upper limit of 10GB. Should be able to then revert to:

• install both embedding models
• download and cache pkgmatch data