`borsh::io::Error::new` diverges from `std::io::Error`

[kayabaNerve]

borsh-rs/borsh/src/nostd_io.rs

Line 229 in a34f324

pub fn new<T: Into<String>>(kind: ErrorKind, error: T) -> Error {

https://docs.rs/borsh/latest/borsh/io/struct.Error.html#method.new

Into<String> vs Into<Box<dyn Error + Send + Sync>>.

I don't see a reason why it should be so specialized. This is technically a breaking change for the no-std API, as a type may impl Into<String> but not Into<Box<dyn Error>>, but I'd argue it as a bug fix.

I open this issue not for the hell of it, but because I did run into this while mapping borsh's no-std io to some work of my own.

[dj8yfo]

@kayabaNerve i guess i did notice this discrepancy too when moving the module back and forth.
it's the same in https://doc.rust-lang.org/1.67.0/std/io/struct.Error.html (and looks like everywhere below too, i tool a look at 1.21).
agree that it should be put to a bugfix category

i'll submit (or will accept) a pr with this change Into<String> => Into<Box<dyn Error + Send + Sync>>

[dj8yfo]

I don't see a reason why it should be so specialized.

yes, ok, the reason is that https://doc.rust-lang.org/std/error/trait.Error.html was never transferred to borsh::nostd_io ,
probably for the sake of simplicity.

If one starts to transfer that, then he/she would become unsure of when to stop.

@kayabaNerve you'd have to share the results of your experiments, in which way specifically borsh is blocking you.
Maybe we can work out a way without (over?) complicating borsh.
As the doc string suggests, the implementation was minimal enough to support

/// let custom_error = Error::new(ErrorKind::Other, "oh no!");

thas is all, which is the only thing borsh needed to stay polymorphic over no_/std::io in order to be able to compile and run

https://doc.rust-lang.org/std/string/struct.String.html#impl-From%3CString%3E-for-Box%3Cdyn+Error%3E

[dj8yfo]

this thingy looks relatively actively used, especially recently, https://crates.io/crates/no_std_io
borsh::nostd_io can be tried to be duct-taped to this.

but then it doesn't look too officially and actively maintained

[kayabaNerve]

There is now core::error::Error if the MSRV bump is acceptable. There doesn't have to be a reproduction of it.

Otherwise, I'd advocate a trait Error: Debug {} with a default impl as a better shim if the MSRV bump isn't acceptable. All instances of the actual Error trait, as necessary to satisfy it on std, will satisfy that. The exact depth it should further model I don't immediately comment on.