[jabrailkhalil]

Failed comparisons can leave automatically created result files in the temporary directory. Cleanup now runs after both benchmark execution and result loading, while explicitly requested output files remain untouched. An already-removed temporary file is tolerated; other cleanup errors still surface.

The regression tests use real temporary files and child Python processes. They cover success, nonzero exit, launch failure, invalid JSON, unsupported schema, explicit output on success/failure, and a child removing its temporary output. Four regressions fail on the old implementation; all eight pass with the fix.

Validation:

• All 40 Python tooling tests pass on Windows / Python 3.12 and Linux / Python 3.12.14, using the repository's pinned NumPy/SciPy requirements.
• Actual compare.py CLI checks reproduce the leak before the fix and verify cleanup afterward, with original exit codes and explicit output preserved.
• Ruff 0.14.0 check/format, mypy 1.18.2 and git diff --check pass for the changed Python files.

This changes Python tooling only. The C++ library and compiled bindings were not rebuilt. AUTHORS/CONTRIBUTORS are updated as requested by CONTRIBUTING; the Google CLA already passes on my existing Google contributions.

Fixes #2349.

I used AI assistance for analysis, implementation, tests, validation, review and drafting this contribution.

[jabrailkhalil]

The Google CLA check has passed. The workflows currently show action_required; could a maintainer approve their run? The Python tooling suite passes all 40 tests on both Windows and Linux.

AI assistance: checking CI status and drafting this update.