Run the step forever under the policy: a completed step resets the backoff and reruns at once (the step owns its own pacing -- poll waits and cycle delays live inside it); a synchronous fault classifies through the policy (Transient logs and backs off, Permanent rethrows); an asynchronous exception is never caught (tryAny), so cancellation tears the loop down like any other thread. The Void return makes "this loop never returns" a fact of the type.

One loop's supervision policy: the label its log lines carry, how a synchronous fault is classified, and the backoff its transient faults pace at. Loops with wiring faults that no retry can fix (an unconfigured handle reached at runtime) classify those Permanent; everything else defaults Transient.

What the supervisor does with a synchronous fault the step let escape. Asynchronous exceptions are never classified: cancellation propagates untouched, so the shutdown race can always tear a supervised loop down.

A bounded exponential backoff: doubling from the base towards the cap as consecutive failures mount, so a persistently-failing dependency is retried at most once per cap interval. A base equal to the cap is a fixed-interval retry.

The delay before the next retry, given how many failures have run consecutively: base * 2^failures, saturated at the cap. The exponent is clamped so the doubling cannot overflow before the ceiling applies.