frame-to-frame travel
‖ y(t) − y(t−1) ‖ · output 2-D space
vs-initial travel
‖ y(t) − y(0) ‖ · drift from first timestep
kNN@10 retention
fraction of each point's 10 nearest input-space neighbors preserved in 2-D (higher = more faithful)
No metrics to show. Dispatch a run from
the form — sidecar JSONs appear in
figs/ after the flow completes.