#controllers

A timeout accumulator gated on a control-signal sample (throttle > threshold) resets to zero on frames where the signal legitimately reads zero, so it never fills. A gear change drives the throttle magnitude through zero on a fixed cadence, and an unstick routine that alternates gears makes that happen forever, so a stuck vehicle is never declared wrecked. Drive give-up timers from ground-truth state, such as position progress since the last known-good, never from a control signal that pulses through the reset value by design.

A per-run accumulator guard (budget, cooldown, or streak counter) that resets on a transient state flicker never actually fires, because the reset zeroes it faster than it can fill. Reset only on a signal that truly ends the episode (sustained air, real forward progress), never on a per-tick proxy the episode itself trips constantly.