In recent years, diffusion models have demonstrated high quality and diversity even in time-series generation, attracting attention as a data augmentation for Human Activity Recognition (HAR). However, conventional diffusion-based augmentation methods struggle to generalize across waveforms captured from different people. In this work, we establish a novel diffusion-based data augmentation framework for HAR, which can (i) simultaneously condition on both activity and individual-specific characteristics, and (ii) extrapolatively generate waveforms from behaviorally unlabeled data of a known individual. Specifically, we propose a diffusion model that employs dual-axis conditioning using both of the latents extracted from a pre-trained activity discriminator and person discriminator. Notably, by incorporating both Classifier-Free Guidance and Cross Conditioning, we find that our diffusion model can generate behavioral waveforms conditioned on action labels transferred across individuals, thereby enabling cross-source data augmentation for waveforms. Comprehensive evaluation on the UniMiB SHAR dataset shows that our method significantly outperforms the baseline, achieving a +5.0% improvement in accuracy in unsupervised person adaptation using the generated waveforms with our framework. We hope our framework contributes high-accuracy and generalizable HAR personalization while significantly reducing the cost of obtaining activity labels.