Transmission reconstruction--the inference of who infects whom in disease outbreaks--offers critical insights into how pathogens spread and provides opportunities for targeted control measures. We developed JUNIPER (Joint Underlying Network Inference for Phylogenetic and Epidemiological Reconstructions), a highly-scalable pathogen outbreak reconstruction tool that incorporates intrahost variation, incomplete sampling, and algorithmic parallelization. Central to JUNIPER is a statistical model for within-host variant frequencies observed by next generation sequencing, which we validated on a dataset of over 160,000 deep-sequenced SARS-CoV-2 genomes. Combining this within-host variation model with population-level evolutionary and transmission models, we developed a method for inferring phylogenies and transmission trees simultaneously. We benchmarked JUNIPER on computer-generated and real outbreaks in which transmission links were known or epidemiologically confirmed. We demonstrated JUNIPER's real-world utility on two large-scale datasets: over 1,500 bovine H5N1 cases and over 13,000 human COVID-19 cases. Based on these analyses, we quantified the elevated H5N1 transmission rates in California and identified high-confidence transmission events, and demonstrated the efficacy of vaccination for reducing SARS-CoV-2 transmission. By overcoming computational and methodological limitations in existing outbreak reconstruction tools, JUNIPER provides a robust framework for studying pathogen spread at scale.