fix(bench): keep benchmark results free of symlinks - #2318

Summary

Fix both failures in the first live calibration run, which compares hosted and local builds of the same source.

The local build failed because fetching from a depth-one checkout imported the commit without its parents or a shallow boundary. git log then failed, leaving GIT_SHORT_HASH and GIT_DATE undefined. Result export failed separately: the profiler created latest_* symlinks inside result/, and the executor refused them.

Changes

• Fetch cached build snapshots with --depth=1, repairing already-incomplete imports too.
• Run each profiler in disposable working and temporary directories outside result/. Keep metrics and process logs in the results.

The installed executor needs no reinstall: it already downloads the benchmark scripts for each new job.

Testing

All 50 affected tests passed on Linux:

uv run --no-project python -m unittest \
  scripts.tests.test_e2e_bench scripts.tests.test_e2e_ci \
  scripts.tests.test_e2e_entry -v

The two regression tests fail before the fix and pass after it. No Rust build or live calibration rerun was performed.

Security Considerations

The executor still rejects symlinks in exported results. No proof or verifier code changes.

Breaking Changes

None.

Accepted debt

• One test requires TMPDIR to equal the working directory; the contract only requires both to stay outside results.
• Old incomplete imports may remain in cache reflogs. Their effect on later Git garbage collection or repacking is unverified.
• The existing build.rs still hides git log errors.