ci: retry managed infrastructure failures once

[lilly-luo]

Hypothesis: managed-config instability can be contained by distinguishing infrastructure timeouts/runner loss from assertion failures and retrying only the former once on fresh runners.

Changes:

• bound the managed subprocess to 24 minutes so evidence can be uploaded before the 30-minute job deadline
• persist the subprocess exit code with the integration artifacts
• classify exit 124 or a non-success job conclusion as infrastructure
• schedule one fresh-runner retry for both managed lanes
• do not retry ordinary nonzero assertion failures

This is one of several CI structure experiments prompted by #755.