Fall detection systems are critical for elderly care; however, cross-dataset generalization and practical edge deployment are challenges in existing approaches. This paper presents an efficient wearable fall detection system based on CNN-LSTM that achieves robust performance across multiple benchmark datasets, with real-time inference on resource-constrained microcontroller units (MCU) while maintaining low energy consumption. The proposed architecture combines convolutional layers with long short-term memory cells to capture spatial-temporal patterns in tri-axial accelerometer signals and their RMS magnitude. Using LOSO validation on the KFall dataset, the model achieves a subject-averaged accuracy of 99.3% and an F1-score of 98.97%. For cross-dataset validation, zero-shot transfer to SisFall achieved 98.3% accuracy and 97.9% F1-score without retraining, representing approximately 1.0% performance degradation under controlled laboratory conditions. The trained model is successfully deployed on an ESP32-S3 MCU, achieving an inference latency of 113.6 ms per 4.0-s window with an average power consumption of 5.78 mA, enabling up to 7 days of continuous operation or approximately 3 weeks under typical daily usage cycles. The proposed system offers a highly practical, energy-efficient baseline that paves the way for future real-world elderly monitoring applications by successfully integrating efficient MCU execution with robust cross-dataset transfer on simulated falls.