Harmony, primarily embodied by chord progressions, forms the structural and stylistic foundation of diverse musical genres, ranging from Classical and Jazz to Pop, etc. Harmony-precedence music composition often begins with establishing a fundamental chord progression, upon which melody, accompaniment, texture, and instrumentation are developed. To mimic this chord precedence in human composition, chord constraints are commonly utilized in the music generation task within the field of Artificial Intelligence. However, there remains a notable gap in explicitly modeling the stylistic richness inherent in chord progression sequences. In this paper, we propose ChordCrafter, a novel framework for learning stylistic distributions of chord progressions, emphasizing interpretability and controllability. We conceptualize chord progressions on a high-dimensional musical feature manifold, characterized by musically-informed filters that capture harmonic properties such as consonance, tension, voice-leading smoothness, and root motion statistics. Monte Carlo sampling is employed to estimate the probabilistic model of the stylistic distributions and generate diverse chord progressions from the learned style-conditioned distribution. Experiments on both objective and subjective evaluations demonstrate the superiority of ChordCrafter, indicating the generated chord progressions are stylistically faithful, musically coherent, and diverse.