Some valuable perspective was provided by Răzvan Sandu in bugzilla on this issue, I'm shamelessly cut and pasting it here.

Hello,

Description of problem:

Even using the rich-language feature, it is still rather difficult to figure out
how to use firewalld on a RedHat/Fedora/CentOS system that is used as a router (a "transparent" system).

That's because:

a. administrators will need different sets of rules/restrictions for access to the router itself and to the various services that run beyond the router (using or non using NAT).

b. It is not very clear how/where the predefined firewalld zones implement their policies (ACCEPT or DROP) and when these policies apply to traffic bounded to the router system or to traffic that traverses the router.

For example, an administrator needs an easy method to restrict VNC access to the router itself (INPUT), but may want free VNC access to some server located behind the router (FORWARD). In the second case, forwarding may (or may not) imply NAT, depending if he goes on the Internet via the external interface or simply goes in another LAN segment beyond the router.

c. It is not very clear how/where the predefined firewalld zones implement their trafic rules ( exceptions to ACCEPT or DROP default policies) and when these rules apply to traffic bounded to the router system or to traffic that traverses the router.

Additional info:

Even it is not dynamic, the Shorewall application (http://shorewall.net/) acts as a higher-level language over iptables, offering the same concepts of "zones" for interfaces. Much of its conceptual architecture is directly applicable ("portable") to firewalld, if accepted by developers.

Somewhat different from conceptual point of view, the "zones" are "levels of trust surrounding the router", including thr FW zone for the router itself.
(unlike firewalld, the shorewall zones have no "sources" or "services" embedded in them).

IPv4 and IPv6 zones are completely separated (they actually represent different levels of trust).

Administrators may directly define policies, i.e. allow default actions to be done when an packet travels from a zone to another (ACCEPT, REJECT). The most sane policy between any two zones is REJECT (with further exceptions defined as rules, see below).

Rules are exceptions to policies , explicitly defined (based on various criteria such as source IP, destination IP, ports, etc.)

Rules may be expressed via predefined (or customised) "macros" (which are the direct equivalent of firewalld's "services").

IPv4 and IPv6 policy and rules are completely separated (IMHO that's good, since the use of global IPv6 addresses pose completely different security problems than NATted & externally firewalled IPv4).

Best regards,
Răzvan