What's the right suite for #3?

I'd say bvt-cq is still the right suite. The test is fast enough to work there, and having it in the BVT will make sure that future failures get caught quickly. Having it in some rare set of tests that only gets run whenever we qualify an FSI or something means that you're sticking random bring-up teams with the debug for this whenever it breaks, with that breakage possibly being months old already at that point. The nice thing about continuous testing is that you notice quickly if something doesn't work anymore, have a small set of changes to find the culprit, and the responsible committers haven't forgotten the context of what they did there yet so they can respond quickly.

*ping*

Andrey, any movement here? Let's tie it to a release so that we don't forget it until the next ODM needs to spend several weeks being confused about this again...

Is this a true blocker for proceeding with a M71 DEV or being used as a priority marker?  Please untag as a blocker if the latter. If it is a blocker, we need a solution soon.  Thanks

Ok, let's start with this:
1) Update the test to check only on 4.4+ (see https://crbug.com/634341#c26).
2) Move the test to bvt-perbuild (see below).
3) Revert https://crrev.com/c/366135
4) Fix "gentle shutdown" messages on 4.4+ if needed to make the test pass.

I still see little value in spending CQ time on running it as a part of bvt-cq, though. Something changes only when we add a new device model, so we need to verify 'gentle shutdown' still happens there. And failures in this test shouldn't just turn CQ red. If it's a newly added board in EVT/DVT phase, it can live for a couple of weeks until it is fixed. And if it's an old board, it's hard for me to come up with an actual regression scenario for this check (false positives are possible: if the TPM device is not created at all at boot, there will be no 'gentle shutdown' line - and though that's a real issue, it's not an indication that we stopped making gentle shutdowns for that model). So, maybe run as a part of canaries testing, but not CQ. And that's bvt-perbuild iiuc. Any objections?

#6 I agree, take it out of the CQ, it really belongs to the canaries.  As you say it's highly unlikely that this will be a regression.