提出停顿决策步态，让四足机器人在无地图情况下更精准爬楼梯。

Onboard quadruped inspection systems often share limited compute between perception and navigation, reducing the rate at which event-triggered controllers evaluate proprioceptive signals. We study this latency in stair-summit detection and propose a climb--settle ``stop-to-decide'' cadence for structured, mapping-free inspection. On a Unitree Go2, the integrated stair loop ran at $\approx$15 Hz. On a three-level stepped platform whose 50 cm top was shorter than the robot, continuous-climb overshoot increased with per-period advance $v/f$, whereas the climb--settle cadence held observed overshoot near zero (22/45 vs 1/45 pooled over $\approx$30/20/15 Hz; Fisher $p\approx2.4\times10^{-7}$). A logistic dose--response model gives a model-based critical rate of $\approx$19 Hz at 0.30 m/s; a pre-specified 40 Hz held-out check was consistent with the protocol-clean fit. We integrated the detector with line following and a three-segment 90$^\circ$ corridor maneuver in a fully onboard, learning-free stack using an IMU, foot-force sensing, three 1-D ranges, and one line camera. The corridor maneuver completed 20/20 trials without contact, compared with 14/20 completions and 12 wall contacts for in-place yaw; the full course completed 18/20 trials. Results are limited to one calibrated course, robot, and operator but identify loop rate as a deployment parameter for proprioceptive event detection.