e2e: rerun failed tests once to absorb transient agent flakiness

Real-agent E2E ran single-shot (-count=1, no retries), so one non-deterministic agent response turned the whole agent job red. Add gotestsum --rerun-fails=1 so only the failed tests retry once; a test that passes on retry reports green. --rerun-fails-max-failures=10 (default) leaves broadly-broken runs untouched so real regressions still surface. Canary/roger-roger deterministic tasks are left strict.