Recycle problem runner workers between problems.

[rohangpta]

Why

Long SCBench runs execute many problems through a small process pool. Each problem builds a full agent/evaluation/
runtime stack: agent clients, Docker clients, subprocesses, temp dirs, logging state, progress threads, and third-
party library caches.

The code has explicit cleanup paths, but over a long horizon any missed cleanup, process-local cache growth, leaked
descriptor, or stale third-party state can accumulate inside a reused worker. When that happens, failures show up
late in the run and are hard to attribute to the problem that caused the leak.

Recycling the worker after each problem gives each problem an OS-level cleanup boundary. It does not replace normal
cleanup; it limits the blast radius when cleanup is imperfect.

What Changed

• Set max_tasks_per_child=1 on the problem runner ProcessPoolExecutor. (note: this adds process spawn overhead between problems. That should be small relative to agent/problem execution time)
• Added a focused unit test that verifies the runner configures worker recycling.

[gabeorlanski]

I like this fix and thank you for adding it. But why does it need all of the CLI utils / other parts?

[gabeorlanski]

Closing this in favor of the smaller one.

[rohangpta]

my apologies, #12 here is the right separate PR for the patch I had in the cli_utils.py file (also a small optimization that I hit live while running your benchmark on my custom harness)