retry on connection errors

[xzrderek]

Note

Medium Risk
Touches rollout execution lifecycle and shared-session cleanup behavior across parallel runs, which can affect concurrency and resource management. Retry behavior also changes for remote HTTP failures and could mask some server-side issues if misclassified.

Overview
Improves retry and resource handling for rollout execution. rollout_processor_with_retry no longer calls rollout_processor cleanup per run; cleanup is now performed once in evaluation_test.py (including the priority-scheduler path) to avoid closing shared sessions while other parallel runs are still in flight.

Expands and refines retry triggers. Default retryable exceptions now include aiohttp connection/disconnect errors, and RemoteRolloutProcessor treats HTTP 5xx from /init as a ConnectionError (retryable) while keeping 4xx as non-retryable failures.

Tests are updated to stop asserting per-run cleanup from rollout_processor_with_retry.

^{Written by Cursor Bugbot for commit a11c7dd. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}