Efficient orchestration of latency-critical Service Function Chains (SFCs) in mobile edge networks requires careful coordination across heterogeneous and dynamically changing resources such as CPU and memory, especially under bursty and time-sensitive workloads. Traditional single-resource optimization models overlook the nonlinear dependencies between Virtual Network Function (VNF) deployment delays and cross-resource contention, limiting their effectiveness in collaborative edge environments. This paper proposes a two-phase adaptive orchestration framework that integrates Proximal Policy Optimization (PPO)-based VNF placement with a Karush-Kuhn-Tucker (KKT)-based resource allocation strategy. The first phase employs an action-space-pruned deep reinforcement learning (DRL) agent to dynamically determine VNF placement, eliminating invalid exploration through topology-aware constraints and dependency-aware action filtering. The second phase introduces a latency-aware KKT-based resource allocator that minimizes execution delay through adaptive multi-resource allocation while explicitly accounting for VNF reconfiguration overheads and server capacity constraints. Extensive experiments demonstrate that our approach reduces average SFC execution latency by 13.38% compared to benchmark methods. Furthermore, the proposed adaptive pruning strategy accelerates DRL convergence by 30.43%. Source code available at https://anonymous.4open.science/r/men-sfc .