Forgot to mention one oddity I noticed when debugging this. To reproduce
faster, I reduced the powerdevil delay to suspend the screens to 1 minute. This
works, and while the bug may not happen if I resume just after the monitor
suspends, waiting an extra 20 seconds (so a total of 80 seconds) seems to be
enough.

But I noticed that after my monitor suspends, it is kind of suspended a second
time 30 to 40 seconds later (the monitor reactivates to show "HDMI POWER
SAVING", then is disabled again). But I don't have to wait this second suspend
to get the bug, and it also happens on 3.2.35, so I'm not sure this is a clue.