Strange, that stuff shouldn't have changed. It's probably my copy of linux screwing with things again. Probably updated the math library or something and now everything is off. Try "cl_showmiss 1" and see if it spams a bunch of stuff. Also try the build 20 server update executables on the client. Which server were you testing this on?
Edit: Well, I just tried on the EV1 server with both a build 20 and 19 windows client, and neither stuck. I ran around on carpathian and tried some of the same stuff in your demo. What I'm running on the EV1 server is the debug build, though, so it's probable the math calculations are different in release mode. Not sure why these would change all of a sudden, but I go back to my previous statement. Linux backward compatibility is terrible.
Edit 2: Put up a test server with the release build and I'm getting all kinds of prediction misses on angled walls. Stupid linux.