Person re-identification (ReID) models trained on small, curated datasets often suffer from severe domain shift in real-world deployments. While large-scale datasets could mitigate this issue, their annotation cost is prohibitive. Raw surveillance videos, though abundant, remain underutilized because existing unsupervised pipelines are tailored to modest benchmarks, relying on computationally expensive clustering and noisy label supervision. We propose a weakly supervised, domain-generalizable ReID framework that learns directly from unlabeled surveillance videos and generalizes zero-shot to unseen domains. Our method leverages off-the-shelf pedestrian detections to form tracklets and employs a contrastive objective at two levels: (i) single images for appearance invariance, and (ii) consecutive frame pairs for temporal coherence. Camera labels–available without identity annotation–are further exploited via a lightweight principal-component alignment module to suppress view-specific variations. Experiments on an unlabeled dataset three orders of magnitude larger than public benchmarks show that our approach consistently surpasses prior unsupervised methods, while remaining linear-time and memory-efficient.