Compare the number of folders in the gh-pages branch
https://github.com/NVIDIA/cuda-python/tree/gh-pages/docs/pr-preview
with the number of open PRs that we have:
https://github.com/NVIDIA/cuda-python/pulls
I think the numbers mismatch.

Some logic was oversimplified in the doc_preview action, for example:

These steps are only run when the target branch is main. So one class of cleanup failures came from the previous path finder work, where PRs targeted a feature branch instead of the main branch. However, this bug alone does not seem to account for the discrepancy fully; there seem to exist other issues that I could not tell immediately.