i think this will 
- increase the probability that a developer will agree to dogfood LUCI
- increase our confidence in adding much more people to dogfood (e.g. everyone)

Beta will end in a few weeks when we start converting builders so I think we should only pursue this if the work is very small.

> How hard it is to implement? 

Hard. 2 days for a first crappy version, the decision to fall back won't be remembered across restarts. And CQ code will get worse, I don't like this.

But even if this was easy, I don't think I like approach:

0. How do we know 3 failures aren't because CL itself is bad and we aren't just wasting capacity?

1. 3 failed builds is still easily 1-2 hours of delay.

2. 3 failed builds on the same builder is above CQ retry policy, so either we increase retries (which leads to more capacity wasted in Chromium), or we make CQ code even more complex.

3. Suppose this is implemented, what signal do we get from dogfood anyway?
If builds are green, devs won't be looking at them anywya. If they are red, there will
be fall back to buildbot build.


WontFix for now.

sg