tests: fix mtime flakiness in git gc tests

[yuja]

#3537

@thoughtpolice can you try this patch? I cannot reproduce the problem reliably at my end.

Checklist

If applicable:

• I have updated CHANGELOG.md
• I have updated the documentation (README.md, docs/, demos/)
• I have updated the config schema (cli/src/config-schema.json)
• I have added tests to cover my changes

[thoughtpolice]

I just blindly ran for x in $(seq 1 20); do cargo nextest run --workspace; done since in my experience the failure rate is higher than ~5%, and unfortunately after 10-15 runs it failed again, with a slightly different LHS/RHS:

--- STDOUT:              jj-lib::runner test_git_backend::test_gc ---

running 1 test
git version 2.44.0
test test_git_backend::test_gc ... FAILED

failures:

failures:
    test_git_backend::test_gc

test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 315 filtered out; finished in 0.32s


--- STDERR:              jj-lib::runner test_git_backend::test_gc ---
thread 'test_git_backend::test_gc' panicked at lib/tests/test_git_backend.rs:154:5:
assertion `left == right` failed
  left: {CommitId("395234a1c7b881d22499564b5d2196203ac6d782"), CommitId("ae36ec61acbca7209121052dda43f9ca9db90817")}
 right: {CommitId("ae36ec61acbca7209121052dda43f9ca9db90817"), CommitId("7231fd316deb9d5c8f479e625dba6a9004dddc30"), CommitId("395234a1c7b881d22499564b5d2196203ac6d782")}

Maybe 1 second isn't enough? I haven't spent time looking at the test since I figured you'd have a better idea anyway :)

[yuja]

left: {CommitId("395234a1c7b881d22499564b5d2196203ac6d782"), CommitId("ae36ec61acbca7209121052dda43f9ca9db90817")}
right: {CommitId("ae36ec61acbca7209121052dda43f9ca9db90817"), CommitId("7231fd316deb9d5c8f479e625dba6a9004dddc30"), CommitId("395234a1c7b881d22499564b5d2196203ac6d782")}

Ok, this appears to be an mtime problem in gix. GitBackend::gc() first updates refs internally by using gix API, then spawns git gc process. If git gc quickly rewrites the packed refs, gix can't detect the change because mtime is the same. Perhaps, we can get around the problem by reloading the refs store or gix repo.

[yuja]

Updated tests to not rely on in-memory cache. I've tested the change on old ext3 image (of 1sec timestamp accuracy) where the problem can be ~100% reproduced.

[martinvonz]

Thanks for fixing.

[thoughtpolice]

My test of 20 runs went through fine, thanks a lot!