> Returned error values are sometimes errors.New (e.g. io.EOF) and sometimes random fmt.Errorf strings that couldn't possibly be handled (e.g. all of tls.go basically), sometimes actual structs (e.g. PathError),
This is the advantage of having errors be interfaces rather than concrete values. Once you learn how to use it, it's a strength. The usage of structs that satisfy the interface (as in PathError) is clearly documented and completely predictable. If your concern is that a certain package could use a struct satisfying the interface and does not, well, one advantage of errors as interfaces is that you could propose this change without breaking the backwards compatibility guarantee.
It's really important to keep in mind that errors are adopted from error values in C, which means that their idioms and usage is inspired by that rather than by exception handling. Many new Go programmers (including me, when I started) were more familiar with exception handling in higher-level languages than error handling in idiomatic C, so it does take some getting used to.
> and sometimes panics.
Do you have any examples of the standard library using panics to signal handle-able errors (not counting cases in which the panic is recovered at the top-level in the library and never visible to the caller)?
> The math library, as mentioned above, is pretty darn iffy.
The only complaint I have had about package "math" is that all operations are defined on float64s, which can be annoying, but is far better than any alternative. That's an annoyance, but it's certainly not an inconsistency in the library.
Do you have any other examples of inconsistencies in the math package?