I've also noticed that sometimes (in CI, less than 5% of the runs) the installation or run will fail deep in chromium. It's actually spawning a chrome process and tunneling remote commands to the child process to do work, so, I guess the flakiness is (almost) excusable?
Restarting the test always resolves the issue, fwiw.
I too have experienced flakiness in CI where Chrome doesn't always seem to start or quit within a reasonable amount of time. I get around this in production with timeouts that give any lingering Chrome processes a series of increasingly aggressive requests to quit, but my hosted CI environment doesn't like it when I send kill signals to random PIDs.
I'm hoping that this Firefox client will be somewhat more reliable in that regard.