Here's what we know:

* It's an OOM issue.
* Concurrency wasn't actually turned off (oops), but the instances still recovered, and the only mitigation done was to stop processing client.art builds.
* With a bit of logging, we know that the request that ultimately logs a critical OOM error might not be the actual failing request.
* This build looks particularly suspicious:
https://uberchromegw.corp.google.com/i/client.art/builders/bullhead-armv8-gcstress-debug/builds/238
https://screenshot.googleplex.com/BGmyjzuEv1F

This is a finish build, but the status is running.

TODO: Try to repro this locally by generating large builds similar to the one above.

Talked it over with nodir@.  It's still somewhat speculative, but I think we can reasonably believe this is the canonical root cause, and it has been written in the PM.