eval: define a consistent framework-timeout policy for background processes across arms

Follow-up from #2971 review. Arms differ in how subject-owned background processes survive agent shutdown vs framework timeout (see the DSH PTY-tree teardown found in the 89-task run). Define one policy, apply it to every arm, and test it. (Raised by @M4n5ter in the approving review.)

Start with the review follow-up in #2971 and inspect the DSH PTY-tree teardown identified in the 89-task run. Compare how each arm handles subject-owned background processes during agent shutdown and framework timeout. Done means one documented policy is applied across every arm and covered by tests.

Written by the indexing model from the issue text.