Nondeterministic behavior when running zinc multiple times in different folders

[lihaoyi]

steps

• Check out 2nd attempt at Reproducible out folder contents com-lihaoyi/mill#4642 at commit 6398ba3
• Run the reproducibility integration test via

./mill -w 'integration.feature[reproducibility].local.server' mill.integration.ReproducibilityTests.diff`

• Extract and decompress the zinc analysis files via

cp out/integration/feature/reproducibility/local/server/test.dest/sandbox/run-1/out/mill-build/compile.dest/zinc zinc-1.txt.gz && cp out/integration/feature/reproducibility/local/server/test.dest/sandbox/run-2/out/mill-build/compile.dest/zinc zinc-2.txt.gz &&  gunzip zinc-2.txt.gz && gunzip zinc-1.txt.gz

problem

This results in two different files

Attached

zinc-1.txt
zinc-2.txt

It seems that this libraryClassName key in the zinc config picks a random class from each jar to store in the analysis file, and which class gets picked differs every time

expectation

I would expect the zinc files to contain the same contents every time. I already have custom ReadWriteMappers that successfully normalize all the file paths, but this class-based non-determinism doesn't seem related to any files

notes

It seems like libraryClassName should contain a list of every class available in each library, but somehow it's only containing a single class per library, presumably picked arbitrarily. Not sure if we're configuring zinc incorrectly or something

[lihaoyi]

@szeiger did you bump into this when you were working on reproducibility at Databricks?