The problem is you're both right. Containers worked great to isolate and manage a large class of faults (though arguably e.g. BEAM would handle the same class of faults better).

Let's say fault chance due to program design was x and container orchestration fails also to correct an independent y, so we went from reliability (1 - x) to (1 - x * y). Total system reliability goes up even thought we didn't get better at writing programs, though so does complexity.

But actually, the market generally only wants (1 - x) reliability. That y was invented for the few situations situations where x alone was unacceptably large relative to other classes of software. Now everyone is using it, which means x drops proportionately across the board. If container orchestration made you 10x as reliable, most CFOs and customers will be even happier with 2x as reliable at the same cost, so your actual software will degrade to be 5x more unreliable.

So for people who want to use "software" and not "systems" (which includes developers, hobbyists, and many people in specific lines of technical work who combine prêt-à-porter with bespoke software) the world now sucks horribly.

Agreed. TL;DR: Application-level reliability did not improve. Orchestration/Ops System-level reliability made up for it.