backtrace() should prefer the source's backtrace if it has one

[dtolnay]

#[derive(Error, Debug)]
pub struct Error {
    #[source]
    source: anyhow::Error,
    backtrace: Backtrace,
}

Currently the generated backtrace() method would just return Some(&self.backtrace). Instead it should be returning Some(self.source.backtrace().unwrap_or(&self.backtrace)). The lower-level error will tend to have a deeper more relevant backtrace.

Note: I am not confident about this. Which behavior is the right one?

Separately, we should return source backtraces if self doesn't contain its own backtrace, which we don't currently do. self.source.backtrace() in the following case.

#[derive(Error, Debug)]
pub struct Error {
    #[source]
    source: anyhow::Error,
}

[Jezza]

Yeah, source's backtrace should be more relevant.

Maybe a good idea would be an explicit #[backtrace].
Only needed if it's ambiguous, or even better, default to the source's backtrace, but emit a warning or something.
That way nothing breaks, the users can still select the one they want, and we get better default backtraces.

[TyPR124]

What if the user launches a thread that might return an error, and they include that error as a source on the main thread? Would that produce two completely separate (but potentially equally important) backtraces?

If that is correct, I feel like hiding the outer backtrace may be a bad idea, especially as the inner backtrace would still be accessible through source.

[mathstuf]

An explicit #[backtrace] could elevate its precedence over the source backtrace.