At the time of the split, did the relation already have from, via, and to members, or just a to member? It looks like the code for splitting only tidies up relations that already have the full complement of members:

This documentation points out that it’s valid for there to be no from member, which would be unlike any other relation type that uses from/via/to roles. I disagree with the note on that page about “via” nodes being nonsensical – they would make sense for the same reason as “via” nodes on any other kind of relation, such as restriction relations. This issue has been raised on the talk page as well.

Yes, the problem occurs with complete relations. I've just created a complete one to show the issue, relation 14305516, and then split the to member. The attached print shows the problem.
I also disagree with that note. I think that 'intersection' nodes are important/necessary. 'via' ways are not part of the original proposal, but used "informally". By the way, it'd be nice if iD implemented some validation of these relations.

OK, so the issue is that the role is intersection instead of via, which causes iD to not recognize the relation as being the sort of relation that has turn restriction–like roles. It’s quite unfortunate that the proposal co-opted from and to roles but with different rules for the from role and a novel via-like intersection role. We could work around that design problem by equating via and intersection in the hasFromViaTo() method above, but it would be prudent to limit this special case to destination_sign relations. I also wonder if we really should place such importance on the relation type’s original design, since the proposal was rejected.

OK, limiting equating "intersection" and "via" to destinmation_sign relations seems fine. I also agree with not placing much importance to that proposal (rejected but widely used, ex. 31500+ in Germany, 5500+ in USA, 4700+ in France, 1400+ in UK, 850+ in Brazil). Anyway, it'd be interesting to have a minimum validation such as requiring exactly one "to" member.
I personally dislike optional "intersection" and "from" but that's the de facto usage.

Hello, what's the status of this issue? I mean, avoid inclusion of multiples 'from' and 'to' to the destination_sign relation when a way ('from', 'to' or 'via') is splitted, so as not to break/change the relation.