The real kicker is this: serde-rs/serde#2120 (comment)

The serde data model lacks the concept of a homogeneously-typed fixed-size array/tuple. This means there is no universally agreed upon compact representation of a fixed-size byte array. The best you can do is using serialize_bytes on the underlying serializer, but this will add a length prefix even if the length is known in advance.

The only way to avoid a length prefix is to use serialize_tuple, however for some message formats like MessagePack this will include type information with each element, leading to a [u8; N] serialization which is much larger than N bytes.

Ya exactly. I notice you said in the linked thread you were switching to serialize_bytes. In which crate is that happening? I don't think it's necessarily a good option here, but I'm curious to see.

It was in the serdect crate here: RustCrypto/formats#1112

It's designed specifically for constant-time bytes handling. Using serde_bytes was the only way to ensure encoding in MessagePack isn't data-dependent. The additional and often unnecessary length prefix is unfortunate.

It looks like there is one project that is using this, here and here.

I also realized, I don't think the p256 impl is DER-encoded. The serde impls appear to just be a thin wrapper around to_bytes, which is good. So this all drops to the GenericArray serde impls. So I guess the move here is to document everything and offer some sample code for doing the conversion. Shouldn't be hard.