My first version saved a GPS fix and showed distance plus a compass bearing back to it. I assumed that would basically solve the problem. It didn’t. The first real tests were indoors and in parking structures, and the naive approach fell apart.

This is the failure in miniature. Under weak signal, v1 trusts a single drifted fix and a noisy compass, so the arrow wobbles and points at the wrong place. v2 follows the trail you actually walked and knows when not to trust itself.

Simplified simulation of the idea, not recorded field data. White dots are raw GPS fixes; the ring is where you actually are.

The short demo uses cuts and 2× walking segments (labelled on screen). No navigation behaviour was altered, and to back that up the unedited capture is published alongside it, with licence plates blurred for privacy.