Hey there.
Which document exactly lacks that information?
The regular "dhcp" wiki page I know about does state it clearly.
See http://wiki.openwrt.org/doc/uci/dhcp on the table described as "Below is a listing of legal options for dhcp sections".
There are both, "start" and "limit".
The LUCI input for DHCP configuration labeled with "Limit" has a question mark icon stating: "Lowest leased address as offset from the network address."
If you know about any documentation handling DHCP but missing either "start" or "limit", could you please link it to allow somebody to adjust it?
Btw: Currently you can still use "10.0.3.1" as "limit" value. Although *that* is not part of a documentation this setting currently works. Being a lacy reader, I just added that value to my LUCI panel, didn't think about it because it worked. Then read the wiki documentation about DHCP and remembered "wow, I just did it slightly different".
I don't know if this method is supposed to continue working but as of current stable BB it works.
As an additional note: Your description is kind of wrong when you talk about "I might want to join multiple networks in this one big subnet".
There are no sub-subnets within 10.0.0.0/16 if your subnet mask is /16. There is just one subnet.
You cannot have one computer at 10.0.0.1/16, a second one at 10.0.0.2/24 and a third one at 10.0.1.2/24 and expect them to interact properly. Either you have one 10.0.0.0/16 subnet and have all nodes within that /16 or you have multiple smaller ones. Then your router needs to have multiple interfaces, one for each small subnet. But configuring them "sub subnet within /16 and one overloard accessing the whole /16" isn't possible.
In one 10.0.0.0/16 you have the 10.0.0.0 as network and 10.0.255.255 as broadcast. The 10.0.0.255 as well as the 10.0.1.0 are just a regular node address.
In 10.0.0.0/24 you have the 10.0.0.0 as network as well but the 10.0.0.255 is not a regular node but the broadcast address.
So mixing different subnet masks for overlapping IP ranges leads to pretty unexpected errors.
Regards,
Stephan.