Accurate localization of P-, QRS- and T-wave boundaries is essential for quantifying clinically relevant electrocardiogram (ECG) intervals. In conventional ECG analysis, waveform boundaries are commonly detected using rule-based algorithms, which often show limited accuracy under waveform variability and distortion. Developing more accurate supervised AI systems, however, requires extensive expert boundary annotations. We propose a weakly supervised framework in which multiple rule-based algorithms generate complementary candidate boundaries from heterogeneous data, knowledge and predictive subsets. Candidate uncertainty is incorporated into an iterative distributional clustering procedure to infer stable, high-confidence consensus boundaries for model training. On the Lobachevsky University Database, the framework correctly detected 96.42\% of reference boundaries, with a mean localization error of 11.10 ms across P-, QRS- and T-wave onsets and offsets. Without dense expert annotations, the proposed method achieved performance comparable to conventional supervised networks, enabling accurate automated delineation of large-scale unlabeled ECG data and providing a reliable basis for quantifying clinically relevant waveform durations and cardiac intervals.