First of all, thanks for working on this! :) Overall it looks good.

My main concern with this PR is changing semantics. From what I can tell you're introducing a bunch of changes in behavior, albeit it corner error cases. What is the risk to compatibility? My gut tells me there's at least a slight risk.

Secondly, you've touched a lot of critical code. Please make sure to run the benchmark suite to ensure the PR doesn't slow down Python. :)

Also, there a number of places where I wanted to suggest a better spelling. However, such changes would be slightly riskier and would mostly clutter up the PR, obscuring the core changes. So I've left out those comments and focused mostly on checking correctness.

Finally, the most likely thing I might have missed in this review is refcounts. You've added quite a few places that exit early when there's an error. I'm not sure that I checked to make sure everything was properly decref'ed in those new error cases.