<p>Space robotic operations increasingly require long-horizon execution of timed, multi-stage tasks rather than isolated pre-scripted maneuvers. However, closed-loop deployment of generated references remains challenging because temporal structure is often implicit, local artifacts may reduce executability, and retimed execution can induce cumulative phase drift. This paper proposes Task-to-Motion-to-Control (T2MC), a unified framework for timed task-stream reference synthesis and retiming-robust tracking. A timed task stream specifies primitive identity, commanded duration, and optional tempo modulation, enabling reproducible task construction and direct evaluation of duration and boundary fidelity. The streamed generative backend combines a structured-prior variational autoencoder with conditional latent diffusion for stable segment-wise synthesis with bounded per-segment cost. A lightweight Contact-Preserving Correction projection is applied consistently during training and inference to improve reference executability while avoiding train–test mismatch. A plug-in online phase-alignment module further corrects reference indexing during execution, improving robustness to non-uniform retiming without retraining the tracker. Experiments on a physics-based robotic tracking benchmark demonstrate that T2MC improves timed-task fidelity, reference executability, closed-loop robustness, and execution fidelity under retimed execution. These results show that the proposed framework effectively supports long-horizon, space-operation-motivated robotic execution by jointly addressing explicit timing, executable reference synthesis, and robust online synchronization.</p>