In the latest set of changes I've finally dealt with the source of all address-related problems. It's this piece of code from ZydisEncoderDecodedInstructionToEncoderRequest:

if (dec_op->encoding == ZYDIS_OPERAND_ENCODING_DISP16_32_64)
{
	ZydisCalcAbsoluteAddress(instruction, dec_op, 0,
		(ZyanU64 *)&enc_op->mem.displacement);
}
else
{
	enc_op->mem.displacement = dec_op->mem.disp.has_displacement ?
		dec_op->mem.disp.value : 0;
}

It opened the gates of hell because it introduces special treatment for absolute addresses coming from moffs variants of mov instructions. So mov [address], eax could result in different encoder requests depending on which variant of mov it was ("normal" vs moffs). Dealing with this on the other end of the story (the encoder itself) resulted in convoluted processing of displacement-only and moffs operands. At some later point 16-bit workarounds were added to the mix.

Now things are done The Right Way ™️. All displacements are the same and they are handled just like in other parts of Zydis (decoder and utils like ZydisCalcAbsoluteAddress) i.e. they are 64-bit signed and interpreted in the presence of external size information. For encoder this information is now the address size hint.

This took way more time and effort than I've originally anticipated for something that initially seemed like a minor oversight but I think the changes are worth it. New behaviors are truly uniform and shield the user from all the ugly quirks of mov. Handling of none hint should be more aligned with the expectations of a typical user as well (when encoding from scratch).