Deep reinforcement learning (DRL) enables adaptive spectrum and power allocation in highly dynamic vehicular ad hoc networks, but full-size policies are often impractical for resource-constrained onboard units. Conventional policy distillation and model compression mainly preserve average return and may distort decisions in rare states associated with outage, deadline violations, or packet loss. This paper proposes a risk-preserving compression framework in which a constrained teacher transfers its action distribution, upper-tail cost quantiles, and safe-action labels to a lightweight student. The student is subsequently compressed through critical- state-aware structured pruning and mixed-precision quantization. Evaluation in a Python-based VANET simulator covers highway and urban topologies, unseen vehicle densities, mobility shifts, imperfect channel-state information, observation delays, burst traffic, and missing neighbor data. Compared with vanilla distillation and conventional pruning or INT8 quantization, the compressed policy retains 96.6% of the teacher utility, reduces deadline violations by 33.3% under severe CSI uncertainty, and limits the compression-induced ∆ CVaR0.95 to 0.034. It also reduces model size from 3.61 to 0.72 MB and P99 inference latency from 4.24 to 1.88 ms. These results demonstrate that onboard efficiency can be improved without discarding risk-critical communication behavior.