brew test-bot fails 3rd-party formulae with requires on GitHub runners

[gromgit]

All my macOS FUSE repo builds now fail on GitHub runners with errors like:

Warning: Unreadable formula in /Users/runner/work/homebrew-fuse/homebrew-fuse/ratarmount-mac--1.2.2.arm64_tahoe.bottle.tar.gz:
ratarmount-mac: cannot load such file -- /opt/homebrew/Cellar/ratarmount-mac/1.2.2/require/macfuse

All the FUSE formulae require two files now, because a recently-introduced audit restricts each required file to contain only one class. Even if I could merge all the necessary code into a single class, the above error suggests that I will now have to copy-paste the relevant code into each of 36 (at last count) different formulae, and even the smallest code tweak would require manual updates of all formulae.

I assume that this new restriction is meant to ensure that bottled formulae are logically complete, and while I can get behind the intent, the current implementation results in a painful third-party experience and loss of DRY with increased scope for hard-to-track bugs, when identical code chunks in different formulae start drifting apart from each other over time due to simple human error.

I think this can be resolved in one of the following ways:-

1. Relax the above bottled formula check for third-party taps. After all, third-party formulae need their corresponding taps to be present on user machines anyway, so the need for completeness in the bottled formula file should be reduced. That said, if any part of Homebrew absolutely needs to load that file, an error would of course be thrown.
2. Follow requires in the bottled formulae and add the referenced files to the bottle. This would certainly be preferred, but I suspect there will need to be some path restrictions in place (e.g. require only files in same directory as the formula).
3. Document this "no requires in formulae" restriction, then folks like me will have to figure out how to implement Requirements in a different way.

[Bo98]

You seem to be conflating a few different problems into one which is making this difficult to follow, perhaps from frustration here. As far as I can gather, the issue of require needing two files now is irrelevant given it's failing on the first file anyway. I'm going to focus purely on the warning you received in your CI logs.

The formula file in the bottle has always been used when installing directly from a bottle tarball. What has changed is which path the formula file it is pretending to be. Previously it pretended to be in the tap directory but now it more accurately is where it is actually installed to: Cellar/<version>/.brew/<formula>.rb. There's a few reasons for this:

• It now correctly matches what is actually being read - the bottle file has always been used here and not the tap file.
• It now aligns with postinstall, which did not pretend to be the tap file (postinstall would have failed before under your setup if you had any)
• The formula file in the bottle matches the version being installed whereas the tap file may not always be the same.
• Homebrew relaxes deprecation rules for bottle-packaged formula and faking to be a tap file when it wasn't breaks that.
• Determining the tap file to fake never actually worked correctly if you used --only-json-tab (the preferred packaging for GitHub Packages) - if a homebrew-core formula was named the same it would have faked being that homebrew-core formula file instead.

However, there is an element of backwards compatibility that was retained here. In the event the bottle file fails, it will fall back to use the tap file. So CI isn't failing with that message - it is simply just warning. However something in that process seems to be triggering a download and thus leading to the Couldn't find manifest matching bottle checksum error, perhaps because the bottle file doesn't contain a bottle block but the tap file does. I don't have an answer for this right now - I'll need to dig into this deeper and try to reproduce the bug as that error should not be happening.

Separately, there is an interesting point on whether we want to handle shared code better. We have a similar setup in our portable-ruby formula, though that's not a "real" formula so doesn't hit the same issues. The point being though that the general structure of using require is understandable. Homebrew/core formulae never use split files for different reasons as split files do not work with API installs.

[Bo98]

I don't have an answer for this right now - I'll need to dig into this deeper and try to reproduce the bug as that error should not be happening.

Homebrew/brew#21743 contains a fix for this issue.

[gromgit]

Thanks @Bo98, that seems to have solved it.