I'll take the out of date comments. They provide a red flag that either the original author was confused, or the behavior changed over time. At least then you can do code archeology to piece together descriptions of what the original intent was, and how it changed using the commit history and figure out where things went off the rails, and thus determine how to actually fix it rather than patch over it even worse.
The problem with unit test as documentation is that over time they end up reflecting the same misconceptions that the code has. Someone does a refactor, misunderstands how the original code works, and "fixes" the unit tests to pass. Now you have tests that lie just like comments can lie.
