Every case. Every signal.

Research context, not a live safety demonstration. These saved held-out fall and everyday-activity cases were previously used for diagnostic analysis, not fresh blind validation. Falls are shown first. Recorded candidate, veto, rescue and final-alert states are offline results. Alert intervals refer to measured source-frame times, not real-time latency or anticipation. Missing cache coverage is unavailable, never assumed OFF. A candidate is not a final alert. Curves are display-only shape-preserving interpolation. Cached F4 joints are already One-Euro-filtered image estimates with short-dropout holds, not raw 3D measurements. Flagged gaps are masked in plots; source values, peaks and missing samples remain unchanged. Sparse semantic observations remain discrete points. Compression changes playback timing, not the source-time signal axis.