<p>Vehicular networks increasingly require reliable and adaptive task execution frameworks capable of handling rapid mobility, varying traffic scenarios, variable network conditions, and diverse task semantics. This work proposes a risk-sensitive and semantic-aware computation offloading framework for mobility-coupled vehicular networks. To handle delay and energy consumption uncertainty, the offloading process is modeled using Conditional Value-at-Risk (CVaR), which captures the expected cost in adverse conditions. A binary offloading model is developed that accounts for vehicle dynamics, semantic task priorities, channel variability, and Roadside Unit (RSU) availability. We introduce a stochastic vehicle-to-RSU association model, formulated through spatial-temporal connectivity probabilities, and analyze its properties under mobility-induced uncertainty. The overall system cost is redefined using a CVaR-based risk metric to capture worst-case performance. A joint optimization problem is formulated to minimize the CVaR of a composite cost function combining service delay and energy usage, subject to binary offloading constraints. This problem is solved using a Lyapunov-augmented Double Deep Q-Learning (L-DDQL), which enables distributed learning of optimal offloading decisions under dynamic network conditions. Simulation results validate that the proposed method reduces latency by 8.6% compared with DDPG, 16.5% compared with DQN, and 33.8% compared with the random method. Moreover, it also minimizes the energy consumption by 5.6% compared with DDPG, 9.6% compared with DQN, and 32.0% compared with the random method.</p>