Recovering complete control-flow graphs (CFGs) from stripped binaries is a fundamental challenge for downstream tasks such as binary lifting, recompilation, and security auditing. However, current indirect control flow (ICF) recovery methods are fragmented, focusing primarily on indirect calls and jump tables while leaving variants like indirect tail calls and returns largely unaddressed. Furthermore, the performance of existing machine learning models is often obscured by noisy ground-truth labels and structural data leakage. We propose ICFlowNet, a unified multi-task learning framework designed to resolve all major ICF variants within a single model. ICFlowNet leverages an augmented heterogeneous graph attention network (HGAT) to capture shared structural dependencies across different dispatch mechanisms. Evaluated under a rigorous, leakage-aware protocol on a large-scale corpus of 15,901 binaries, ICFlowNet-MTL achieves state-of-the-art performance, reaching a 98.73% F1 score for indirect calls and 99.43% for jump tables. Our work demonstrates that combining architectural innovation with strict evaluation hygiene is essential for reliable, real-world ICF recovery.