A Leak-Safe Within-User Benchmark for Compact Surface-EMG Grasp Decoding, with a Causal Sequence-Reasoning Decoder

Surface electromyography (sEMG) is the dominant control modality for myoelectric hand prostheses, yet a persistent gap remains between high offline classification accuracy and real-time, on-device control: models are often large, evaluated under leaky window-level splits, and reported with a single inflated metric that hides instantaneous behaviour. Our contribution is primarily methodological: a strictly leak-safe within-user extrapolation protocol (calibrating on a user's earlier repetitions and testing on a held-out later repetition, split by contiguous recording segment so no overlapping window crosses the split), paired with honest dual reporting of per-window (instantaneous) and per-execution balanced accuracy alongside the rest false-activation rate. We instantiate it with a compact (28,199-parameter, 27.5 kB int8) decoder for a functionally-motivated set of six hand grips plus rest: a depthwise-separable CNN, leak-safe per-user calibration, and a causal sequence-reasoning decoder in which a population-learned, zero-parameter transition grammar reasons over per-window evidence. We evaluate on NinaPro DB2. With five calibration repetitions the system reaches 94.0% per-execution balanced accuracy at 2.2% false-activation with the deployed operating-point gate (92.7% from the gate-free decoder stream alone), and 75.2% per-window accuracy, with a microcontroller-compatible 28k-parameter model. We report a rigorous ablation: a classical time-domain baseline is competitive on the saturating per-execution metric, while the learned representation wins on the discriminating per-window metric, and the causal sequence decoder gives a small but statistically significant per-window gain over a matched-window majority vote (Wilcoxon p = 0.001) at zero added parameters. Under cross-subject 40-fold leave-one-subject-out evaluation the same pipeline reaches 50.6% per-execution balanced accuracy with no labelled calibration (unsupervised per-channel normalisation only), rising to 89.1% with 20% labelled calibration, where the sequence layer contributes a larger gain than within-user. We position the work honestly against recent ultra-low-power and foundation-model EMG decoders and report compute cost rather than claiming on-device measurements. Code and configurations are available at https://github.com/seanb9/emg-leaksafe-benchmark.