Reliability Assessment of Wi-Fi CSI-Based Activity Recognition Under Environmental Shift and Structured Measurement Loss

Wi-Fi channel state information (CSI) supports contactless activity recognition, but performance in one acquisition setting does not establish reliability after environmental change and measurement loss. We evaluate 1,260 genuine recordings from 30 participants in three indoor environments using participant-disjoint source partitions and leave-one-environment-out tests. An initial 60-fit model comparison and a separate 90-fit component-removal study retain mixed and adverse results for compact missingness-informed and residual networks. A further exploratory execution freezes 75 source-fitted models across five prespecified configurations before their test evaluation. Fixed log-mean-power inputs feed observed-endpoint temporal moments and generalized covariance spectra. The declared feature-only MLP arm E achieves mean held accuracy of 75.17 ± 1.00% on complete input, 73.82 ± 0.94% with random 50% loss, and 60.60 ± 1.37% with contiguous 50% loss. These deviations describe five training seeds after within-seed mask averaging. Its contiguous-loss macro F1 is 59.16%; two seed accuracies remain below 60%. The primary matched-endpoint versus observed-cell normalization contrast in the hybrid model is only +0.54 ± 0.65 macro-F1 percentage points. E's result therefore does not establish that support matching alone causes the improvement. Source-only calibration, recognition, proper probability scores and repeated-view dependence are documented alongside verified checkpoints and saved predictions. An additional fixed-power-input MiniRocket reference reaches 53.84% complete and 19.41% contiguous-loss accuracy. Its clean-source training and mean filling differ from E's loss augmentation and observed-mask features; this comparison does not isolate architecture. Different cohorts occupy each environment, and missingness is simulated after packet-index normalization. Historical target outcomes influenced development; external replication and matched controls are needed before general performance or deployment claims.