[P1] reproduce: cleanup deletes unrelated files in the user-supplied work directory

[reacher-z]

Problem / impact

clawbench-reproduce --work-dir <existing-directory> recursively deletes the entire supplied directory after comparison unless --keep-cache is supplied. The directory need not have been created by this invocation, so unrelated user files are removed along with downloaded traces.

Source: reproduce.py. The code calls shutil.rmtree(args.work_dir, ignore_errors=True) without tracking ownership. mkdir(..., exist_ok=True) earlier accepts an existing directory.

Verified reproduction (offline, temporary directory only)

1. Create a temporary user-work/ with unrelated-user-file.txt and downloaded-traces/.
2. Mock download to return downloaded-traces/ and rescore to return a matching 129-task summary; invoke main with --model deepseek-v4-flash --work-dir user-work.
3. Actual: exit 0; both the unrelated file and user-work/ no longer exist.

No real data, downloads, or judge API calls were used for verification.

Acceptance criteria

• Allocate a unique invocation-owned child directory beneath the chosen work directory; remove only that owned child.
• Preserve pre-existing files and other runs' caches on PASS, FAIL, exceptions and cancellation.
• Keep --keep-cache behavior explicit, and report the exact owned cache path.
• Add a regression using a sentinel sibling file plus an existing work directory; verify it survives normal completion and failures.

Suggested priority P1: user data loss. Separate from #296 / #313 (rescore path defaults).

@Perry2004 please review the cleanup ownership contract and prioritize this fix.