There is another possibility, that direct (verbal) user feedback is discounted as being "anecdotal" and not included when making "data driven" design decisions. But as you mentioned, this is all speculation without an insider's POV.

My intuition, however, is that their data-driven technique over-optimizes the various minutiae while allowing broader flaws to persist. I'd chalk that up to a lack of design vision to guide the testing.

I think there was a clear line drawn at least with the G+ integration, since they started consolidating the appearance of all services under the same UI and same color codes. There is a now a clear "design vision" in Google (even if it clearly sucks bad) and it's clearly not as Engineers-driven/Fonctionality-driven as before. Having Icons without text next to it is not just a style issue: it's a well known flaw in terms of usability. Everyone knows icons+text are superior to icons only. Especially when your icons are dark grey on light grey background. No differentiation whatsoever. I can't believe someone was actually paid to do this.