NAT with VPNs is complicated. This happens because the Encryption Domain of a firewall is used for two things which are related, but subtly different.
First, when the firewall receives traffic, before it processes firewall rules or NAT rules, it checks the source to see if it's in an encryption domain and the destination to see if it's in an encryption domain. If so, and if the endpoints with those encryption domains share a VPN community, this is used to flag the traffic as "interesting" for the VPN between them.
Second, when the traffic is about to be encrypted (after access rules, routing and all NAT!), the VPN community is used to pick the exact IDs which will be sent in the IPSec negotiation. If you translate a single IP address, but the encryption domain contains a /28 network covering that IP, the /28 network will be proposed rather than the single IP.
This is expected behavior, and a side-effect of consulting the encryption domain for different purposes at different times relative to NAT and rule decisions. You need to include the untranslated addresses for the matching logic, and the translated addresses for the negotiation logic.