(1) Can't that be accounted for in some manner, especially when you have so much data? Typically when I design tests at work, it is not unusual to add many rules and exceptions to make sure the data is as accurate as possible.

(2) This is where it becomes tricky. You have to understand that these decisions are made at a very local level by small teams with specific goals(for example, "bump revenue from sidebar links by x%"). Also, it is unrealistic to suggest that a company the size of google can only work on super big items when at Google's volume, small improvements can add up and result in hundreds of millions in additional revenue.

(3) Not a fan of the new gmail but using their blue shade experiment as supporting evidence for why their design sucks isn't very powerful.

Wouldn't #1 be an argument for testing to pick the best color (or at least a big part of any decision) - it doesn't matter what the designer picks, it matters what people see!