Ok, now that you've made your official recommendation, I would like to make a suggestion that is so bizzare that even if there were no obvious technical problems, I doubt it would be worth the risk.
Still I would like to know if anyone can say out of hand what would fail with the following fanciful solution:
First you would get one one of those Vivotek daisy-chain cameras that Ethan remembered. They are basically a camera with two rj-45's and a built-in two port switch.
The idea would be to get one of the rj-45 ports to the existing switch and the other one to the (new?) redundant switch.
Since you said cabling another run back to a redundant switch would be a massive amount of work, you could run both on one cable by using 1,2,3,6 for one port and 4,5,7,8 for the other and then using custom crimps at both sides to split back to two.
Next you would setup the Vivotek to send the stream as multicast. Because the built in two port switch is unmanaged it would end up sending out the multicast on both, because it doesn't know any IGMP info and so therefore must broadcast it.
Then you would setup on the managed switches STP, spanning tree protocol, to prevent a possible multicast storm at the point when the primary and the redundant networks converge.
Which would make the redundant switch drop the multicast. However if the primary network were to fail, STP would enable the forwarding of the multicast by the redundant switch, since it would not cause a loop.
Yeah, weird. But what part would definitely not work?