Would it be possible to run the ci on 32bit arm (which does not have the 64 bit time_t transition from debian 13) to see if this indeed fixes the issue?

One of the nightly builds is on a 32 bit arm platform — https://github.com/libgit2/libgit2/actions/runs/20870863585/job/59972059810

It hasn't seemed to be flaky. But maybe it's being compiled slightly differently than yours?

We build and test within containers on Linux so it may be possible to repro our environment locally and / or update the build environment to match your case.

For context of how I discovered this:

I caught this assert in https://github.com/libgit2/libgit2/blob/main/tests/libgit2/repo/init.c#L445 (on a 32bit hurd system, which has a different order in the stat struct than linux)

repo::init::extended_1 [tests/libgit2/repo/init.c:445]
cl_assert(S_ISDIR(st.st_mode));

Looking at the stat struct in a debugger it was completely garbled caused by enormous values in st_ino and st_size and then I found the FILE_OFFSET_BITS=64 only for the tests.

This test also has been marked as "flaky" on debian https://sources.debian.org/src/libgit2/1.9.2%2Bds-2/debian/patches/disable-flaky-stat-tests.patch probably for the same reason.

The problem is that git_fs_path_lstat and the test disagree on the size and layout of the stat struct.
For this I went with removing from the tests as compiling libgit2 with FILE_OFFSET_BITS=64 is probably an ABI change.

I am reading #4631

and am really confused what happened here.

It seems i encounter the same issue but in reverse (casting stat32 struct to stat64)

The other pr mentions that libgit2 used to be always compiled with _FILE_OFFSET_BITS=64 so putting it to the tests aswell made sense.

I think the timelime is something like this:

• fac7eac : Compile tests with FILE_OFFSET_BITS=64
• ef4ab29 : moved it from libgit2 from src/CMakeLists.txt to src/libgit2/CMakeLists.txt
• 91ba089: Removed it from libgit2 entirely.
• e7fce1b : introduces GITTEST_FLAKY_STAT
  -v1.5 is the earliest version with "flaky" stat tests.