Some fun things callers need to be aware of:

• You may not call wirehair_read twice with the same packet ID, this causes decode failure.
• Memory lifetimes generally arent documented - during encode the memory you pass into wirehair_encode must be available until the last time you call wirehair_write (and maybe later? probably not), whereas the same is not true for writing.

In other news, for my application I hardcoded the chunk size, assumed 16/32-byte alignment (depending on where its used) for XOR, replaced the different-compile-unit XOR code with simple avx-/sse-based intrinsics and saw something like a 3x speedup :).