Synthetic human mobility trajectories support urban planning, transportation simulation, and epidemic analysis by representing when, where, and how people conduct activities. However, existing approaches often model individuals independently or over single days, limiting joint representation of overnight activity, demographic heterogeneity, household interactions, and multiday regularities. The required evidence is fragmented across census, survey, geographic, and anonymous GPS data that cannot be linked at the individual level. We present PopTraj, a hierarchical multi-source generation framework for producing household-coordinated weekly activity-location trajectories for synthetic urban populations. PopTraj organizes generation into three source-aligned levels. A sample-free synthesizer reconstructs a household population from aggregate census constraints. A survey-calibrated language model produces a coordinated-schedule teacher corpus, which is distilled into a Hybrid 3D conditional variational autoencoder for scalable generation. Finally, HEARTH completes each dictated week as home-anchored excursions, using a vector-quantized persona latent to connect GPS-learned behavioral archetypes with a survey-derived demographic prior. In a metropolitan-scale Tokyo deployment, PopTraj generates 395K synthetic person-weeks. HEARTH achieves a mean divergence of 0.0252 across six realism metrics, compared with 0.0412 for the strongest of nine baselines. The deployed schedule pipeline reduces household coordination error by more than 16-fold relative to six learned baselines, while the persona prior halves demographic divergence with little change in aggregate realism. Source-removal, architecture, and error-propagation analyses show that the persona prior supplies demographic separation and the household schedule supplies cross-level coordination. PopTraj therefore provides a scalable route from unlinked population statistics and anonymous mobility data to simulation-ready human mobility trajectories.