Vietnamese automatic speech recognition (ASR) for edge devices remains challenging because practical systems must simultaneously satisfy recognition accuracy, low latency, low memory overhead, and transcript readability. These requirements are particularly demanding for Vietnamese, where tonal and prosodic cues are important for sentence structuring, while many lightweight ASR systems still produce raw lowercase text without punctuation. In this work, we present an edge-ready Vietnamese ASR system built on VietASR Iteration-3 and fine-tuned on nearly 1,000 hours of professionally curated Dollya Vietnamese speech data with punctuation and capitalization annotations. Our work makes three main contributions. First, we develop a deployment-oriented fine-tuning pipeline for a compact 68M-parameter VietASR model, targeting practical offline inference on resource-constrained devices. Second, we introduce an end-to-end structured decoding formulation in which punctuation marks and capitalization indicators are incorporated directly into the output vocabulary and predicted jointly with lexical units, eliminating the need for a separate post-processing module. Third, we validate the system on both recognition and deployment aspects through evaluation on standard Vietnamese benchmarks and on a Qualcomm QCS8550 edge platform. Experimental results show that the fine-tuned model achieves 6.41% WER on the VIVOS test set, while also generating well-structured transcripts with strong punctuation performance (micro-F1 of 0.94). On the target device, the deployed system processes a 30-second utterance with about 380.0 ms encoder latency on CPU and 321.0 ms on NPU, while decoder and joiner costs remain negligible. Comparative analysis further indicates more stable Vietnamese output and better formatted transcripts than Whisper-large-v3-turbo and the pretrained VietASR baseline. These results suggest that compact Vietnamese ASR models can jointly address accuracy, formatting, and deployment constraints, making them suitable for real-time, privacy-preserving, on-device speech applications.