This is/was also my take.
I’m skeptical that a probe-based network can be granular enough to reliably pinpoint a city, especially when some paths are much better connected than others (fewer hops, uncongested fiber, no throttling).
However, ipinfo still appears to rely on active probing to triangulate geolocation data, which suggests they believe these routing asymmetries can be modeled or averaged out in practice.
This video [0] is relevant, though it actually supports your point - it shows Claude Code struggling with non-trivial tasks and needing significant hand-holding.
I suspect videos meeting your criteria are rare because most AI coding demos either cherry-pick simple problems or skip the messy reality of maintaining real codebases.
However, ipinfo still appears to rely on active probing to triangulate geolocation data, which suggests they believe these routing asymmetries can be modeled or averaged out in practice.
https://ipinfo.io/blog/ipinfos-probe-network