[HereThereBeDragons]

Non-ephemeral Windows GPU runners are shared across super-repos whose checkouts pin different submodule commits. git reset --hard (run by actions/checkout) does not clear nested submodule working trees, so a newer nested submodule left on disk by a prior job can remain under an older superproject whose .gitmodules lacks its URL.

actions/checkout then runs git submodule foreach --recursive during its auth setup/teardown, descends into that stale tree, and aborts.

To fix this:
Add a step to the shared pre-job cleanup script that empties the top-level submodule working trees left by the prior job. Submodule paths are read with git config -f .gitmodules and the dirs removed with Remove-Item. git submodule deinit is unusable here: a prior job can leave the workspace partially corrupted (submodule URL in .git/config but .git/modules/<name> missing), making deinit abort with could not lock config file .... Physical removal is independent of .git/modules health. Runs before actions/checkout.

Cache-safe: keeps .git/modules objects and touches no untracked or ignored files, so ccache and downloaded deps are preserved.

Fixes #5742

[HereThereBeDragons]

.github/workflows/test_component.yml and .github/workflows/test_artifacts.yml already have this cleanup_processes step.

@amd-justchen please let me know if htis is a good location for this or you would prefer something different? (new file/rename to general windows cleanup etc)

[cgoea]

LGTM

[HereThereBeDragons]

Test run on rocm-libraries using therock pin from this branch here (+ changes to just run the windows tests):
https://github.com/ROCm/rocm-libraries/actions/runs/27283688359
rocm-libraries branch: https://github.com/ROCm/rocm-libraries/tree/users/lpromber/bump_therock_git_deinit_windows

[HereThereBeDragons]

run is happy and fixes the problem

[amd-justchen]

Thank you for the fix!