ci: smoke job fails intermittently on link-remove and repository-info tests unrelated to the change

Summary

The smoke (linux-x86_64) job in PR Validate is failing intermittently on tests unrelated to the change under test. Over the last week, on branches from several authors, each run fails a different one or two tests out of ~1040, and a rerun on the same commit passes.

Observed (PR Validate runs, last 7 days)

Run	Branch	Failing test(s)
35215639231	bound-branch-push-lookups (#186)	test_repository_info_local_on_fresh_repo
35215974397	fix-empty-branch-revision-info (#184)	test_link_commit_per_link_message, test_link_remove_keeps_uncommitted_local_edit
35216121252	miri-growvec-ub (#210)	test_link_remove_keeps_uncommitted_local_edit
35151934781	miri-growvec-ub (#210)	test_link_remove_keeps_ignored_file_under_mount, test_link_remove_keeps_untracked_file_under_mount
35140317945	miri-growvec-ub (#210)	test_link_remove_keeps_untracked_file_under_mount
35100409458	ap/merge-commit-authorship-metadata-main	test_link_remove_refuses_when_path_case_differs, test_link_remove_keeps_uncommitted_local_edit, test_link_remove_keeps_untracked_file_under_mount
35083243074	fix-build-rs-unwrap	test_link_remove_refuses_when_path_case_differs, test_link_remove_keeps_staged_local_edit

None of these branches touch link removal or repository info. #184 is a 12-line change to the empty-branch RevisionInfo handler; #186 changes branch-push verification; #210 is a lore-base GrowVec fix.

Pattern

• Nearly all failures are in test_link_remove_* (scripts/test/test_link.py); one in test_repository_info_local_on_fresh_repo (test_compact.py), where repository info --local returned empty output.
• The smoke job runs pytest -m smoke -n 4 with --basetemp=/dev/shm/lore-smoke (tmpfs, remounted to 12G). The link tests mount a second repository under the first; every failure so far is in a test that runs a second loreserver/repository alongside the first, which points at a per-worker isolation or timing interaction rather than a product regression.
• Reruns on the same commit pass, and different tests fail each time.

Impact

Fork PRs cannot be rerun by their authors, so each flake needs a maintainer rerun before a PR can go green. Happy to help dig into the link tests if useful; pointing at the isolation problem is as far as I've got.

CC @ragnarula, since #132 is yours, and #210 just failed on this