test: bound and clean up managed-process test lifetimes

[Ehsan-hash]

An assertion failure or Ctrl-C before the usual cancellation step can leave the test's detached supervisor and command running. This adds unconditional test cleanup and limits the busy-loop child to eight seconds, so a missed cancellation cannot consume CPU indefinitely.

The change is confined to test cleanup and its regression checks. Supervisor lifecycle behavior is a separate issue.

Validation on macOS: eight focused checks cover the original cancellation test, nested pass, assertion failure, observation timeout, Ctrl-C, interrupted start, outer-runner timeout and a child arriving during cleanup. The four affected process-test files pass 46 checks, with four Linux Bubblewrap checks deselected because that sandbox is unavailable on this host. The final supervisor/busy-child process scan is empty. Production process code and repository instructions are unchanged.

Developed with OpenAI Codex assistance and independent code review.

[Ehsan-hash]

The Linux CI run finished with 1,521 passed, 67 skipped and one failure: test_research_paths_do_not_import_digest_algorithms rejects the hashlib import in research/planning/workspace_analysis.py.

That failure is also present on the PR base a76120bf42be17cb153be4111ee33afd78f0bf7d. I replayed the exact check against the tracked Python sources at both the base and PR head 34ea71b; both reject the same file. The check and that production file are byte-identical between those revisions. This was a source-bound replay of the failing check, not a full baseline CI run.

This PR changes only tests/test_agent_processes.py. I have kept the unrelated production/check policy outside this test-cleanup change.