Conceptually, this all makes sense.

My only hesitation on approving and merging is that I can't make sense of the example provided for the "proof it works" case. The testbed tests have all been marked as failed... but with no test failure or coverage problems.

I'm presuming this is an artefact of rewriting history, and you manually added a "return 1" to the test case or something?

However, even then - only Windows and Linux have presented app data artefacts after the build - iOS, Linux and macOS have no app data. I can't tell if this is a side effect of the $HOME changes, or an artefact of how the test was made to fail.

As for whether these artefacts are useful - they're generally only useful if a test fails.

The app data is particularly useful for the canvas tests, which generate images of the appearance of the canvas when they fail. This was essential when developing those tests, because we were getting failures caused because the rendering of the canvas wasn't pixel-perfect the same with desktop machines (for any number of reasons). Exfiltrating the app data gives us the actual output under test conditions, which we can compare with local output and work out if there's an actual problem, or just a sub pixel variation caused by some local test condition.

The logs are there mostly for additional debugging when the failure is for Briefcase failures - those don't normally happen, but when they do, it's useful to know the exact machine state when the failure occurred. There's an argument to be made that we should always be running with --log, and then uploading all the artefacts if a failure occurs, in case there's something else going on.

Heh... ok, I've found the source of the two outputs.

The Windows one is a known problem. The app data is a set of images that can't be reliably cleaned up because of Python.net retaining open file references. The "forced fail" means we see the mess that was left.

The GTK app data is some transient storage from the webview test. Again, the forced fail means we see the mess; these files would be included if there was a canvas or image failure, but they'd normally of no significance.

So - looks like you're right that nothing has changed. The artificial nature of the test was just surfacing things we wouldn't normally see.