Reading through the wiki and forums, I'm having a really hard time wrapping my head around how all this VLAN/VPLS/MPLS stuff works together.

If I want to transport VLAN traffic over VPLS, is my l2mtu 1516 or 1530? mrz posted in this thread that the ethernet header is not part of the l2mtu. Yet, the wiki article referenced, shows a diagram that does include the ethernet header, yielding a l2mtu of 1530.

Is mrz mistaken? Is the ethernet header included? Is there an error in the diagram, and the ethernet header is not included? Why does the diagram show 2 ethernet headers, including one in the total but not the other?

If you are using VLAN, VPLS needs to encapsulate the appropriate ethernet header as well.
So in that case, you have an outer ethernet header (for Transport on wire) and an inner ethernet header (where your VLAN is contained).
Therefore, your L2MTU is 1530.

Bummer. I guess VPLS is out of the question... I'm not about to start provisioning a VPLS tunnel per customer.

Is there any way to transport a VLAN over a routed network without fragmentation? Bridging is fun, but I'd rather get away from it if I can. I hate seeing every last bit of broadcast traffic showing up in every little nook and cranny of my network.

You don't need to provision them manually.
Try using BGP autodiscovery for VPLS (there's a wiki entry for that). It works like a charm.
It's a lot less work than fondling around with VPLS tunnels by hand.

Also, I wouldn't break too much of a sweat about having to fragment.
As RouterOS is using VPLS Controlword, it transparently fragments packets going through a VPLS tunnel,
just like an EoIP tunnel would.

From tests that I did, two Routerboards 600A running at 533 MHz are able to push in excess of 600 Megabits per second though a VPLS tunnel between them,
even if they have to fragment and reassemble (I set L2 and MPLS MTU to 1500, while transmitting 1500-byte sized packets). An EoIP Tunnel on the same machine only managed about 280 Megabits. So VPLS is really worth it, even if you have to fragment

Is mrz mistaken? Is the ethernet header included? Is there an error in the diagram, and the ethernet header is not included? Why does the diagram show 2 ethernet headers, including one in the total but not the other?

If you look at the diagram more carefully, you will see that it is not included when calculating L2MTU until VPLS encapsulation occur.
Ethernet header is there only to show you that it exists and that whole ethernet frame (including ethernet header) is encapsulated into VPLS packet.

OK, help me understand fragmentation... will it be end-to-end, or just on the one ethernet link between R4 and R5?
If it is end-to-end fragmentation, how will this impact performance on the wireless links? Will there be a reduction in available bandwidth?