Cross-subject generalization remains a critical barrier for EEG-based imagined speech decoding, limiting clinical viability of brain-computer interfaces for paralyzed individuals. We evaluate rapid calibration through over 28,000 experiments across three architectures, two channel configurations, seven calibration methods, and nine budgets on a 15-subject five-class dataset. Using leave-one-subject-out evaluation with topographic feature maps, we show that zero-shot transfer fails at chance level, while rapid calibration recovers performance substantially: the best configuration reaches 53.71% with 25 minutes of calibration. Most notably, we find a strong positive correlation between source-domain training accuracy and calibration effectiveness, suggesting that adaptation efficiency, rather than cross-subject generalization, may be the more useful design objective for calibration-targeted BCIs in this regime.