ROS 2 kinematic_guard catches mismatches before hard E-stop

Most mobile robot stacks rely on timeouts: if /cmd_vel stops arriving, the base driver halts the robot. But timeouts only detect silence, not physically inconsistent motion. A robot can receive fresh commands while slipping on a wet floor, suffering odometry drift, or reacting to bursty Wi-Fi replays. ros2_kinematic_guard bridges that gap by comparing command velocity against measured odometry over a short sliding window. When the residual exceeds a threshold, it emits a structured JSON diagnostic with fields like causalAlignment (ALIGNED/BROKEN) and dominantCause (WHEEL_SLIP, LOCALIZATION_JUMP, etc.). In guard mode, it can clamp or brake the outgoing command before the safety-rated E-stop is forced to act.

The package runs as an inline topic filter: Nav2 or teleop publishes to /cmd_vel, the guard publishes to /safe_cmd_vel for the base driver. No modifications to planners, behavior trees, or proprietary drivers are needed. Three modes allow staged deployment: observe (log only), passthrough (wiring test), and guard (active intervention). The README includes a 5-minute demo using a lightweight mock robot that injects wheel slip—no Gazebo or Isaac Sim required. For persistent fault debugging, setting slip_duration_sec:=9999.0 holds the condition stationary. This pre-E-stop diagnostic layer aims to reduce unforced hard stops, mechanical stress, and post-incident debugging time in production AMR/AGV fleets.