Operator was struggling to attain "leader" (using old SDK implementation of leader election) but this was a secondary issue. The real issue was that the liveness probe was too tight, and the pod failed to start in time. The pod was then restarted, but the leader lock deletion (handled by garbage collection) was lagging behind so the restarted pod could not become leader. The fix was to adjust the liveness probe settings to the current scaffolded defaults.