[Shubham-Padkonde]

Some Unicode configuration paths encode to Base64 containing /. open_with_lockfile() then treats that slash as a directory separator in the temporary lock path and raises FileNotFoundError before opening the configuration file.

Use URL-safe Base64 so the generated lock name stays a single filename. Add tests for Unicode paths, competing lock acquisition, reopening after release, and cleanup when the context raises, plus a changelog entry.

Validation: 17 utility/update-check tests pass, and changed-file Flake8 passes. The two new Unicode cases fail before the fix. A broader run encountered two Big5 charset-detection failures, both reproduced with unchanged upstream code; the full suite was interrupted during slow network tests, so I cannot claim a complete full-suite pass.

[Shubham-Padkonde]

I compared the CI failures with unchanged upstream 5b604c3, keeping this PR at dab3a07.

On Windows / Python 3.12.6, the utility, update-warning and CLI-help files pass all 19 tests. The two Big5 detection tests fail identically on this branch and upstream (the Big5 subset has 10 passes and 2 failures on each).

The other inspected CI failure is the invalid --pretty choice message: the test expects quoted choices, while CPython 3.12.14's argparse uses str rather than repr for the choice list. Loading that unchanged standard-library module in a Python 3.12.6 compatibility harness reproduces the same 3 passes / 1 failure on both branches. This is not a full native Python 3.12.14 run or a full CI-matrix pass.

No code or dependency pins were changed for this diagnosis. Would you prefer the parser/charset compatibility work handled separately so this PR stays focused on Unicode lock filenames?

Follow-up from a full native Windows comparison with CPython 3.12.10, the version reported by the Windows 3.12 CI job, and matching pip 26.2.1. The PR at dab3a07 had 1005 passed, 3 failed, 19 skipped, 4 xfailed. Upstream 5b604c3 had 1002 passed, 3 failed, 19 skipped, 4 xfailed. Both failed the same 3 cases, including the invalid --pretty choice and both Big5 checks. All three lock-file regressions passed on the PR. Running those same three tests against upstream produced two failures and one pass.