I finally read through the excellent article in Racecar Engineering. It is certainly worth the purchase price of the magazine this week (as a bonus there's an excellent article on picking spring rates). Some conclusions...
1) Picking a Subaru as your only inspiration for your cage design may be to narrow of a focus since it was in a Subaru that Mark Lovell and Roger Freeman died in the 2003 US Championship Rally when their Impreza collided sideways at speed with a tree. It's not clear what if anything you are doing differently then was done in that car. Here's a picture of Lovell that shows some of the cage in the immediate area around the driver.
2) Here's an example of a Peugeot 206 from 2003. While the picture is from a Tamaya model, it gives you some idea of how the cage was laid out.
Here's another image of a complete tub with cage.
Unfortunately in 2005 Michael Park died in a Peugeot in 2005 Wales Rally Great Britain in a similar accident to Lovell's, so it's not clear if this style cage performed any better. Here's a picture of Park in his car.
3) Crush zones and load spreading is key. The door area in a rally car is especially bad since there is only about 30-35 cm between the outer door skin and the nearest occupant. How the designer manages that 30-35 cm's is key to the occupants survival. To give you an idea of the significance, check out the attached chart which was included with the article based on work done by some Cranfield Engineering students.
Basically, at 100G deacceleration -- You're dead.
At 80G's -- you'll have massive head injuries.
At 60G's -- you'll have a concussion.
(Note that there is some disagreement regarding the actual levels, but I think it's save to walk away here with an idea of the magnitude of the forces that we're talking about. Note that in other situations (such as Indy Cars), and if properly restrained drivers have survived over 100Gs.)
Disclaimer: I'm not an engineer, and I didn't sleep in a Holiday Inn last night. I'm just a safety concious 911 driver (and sometimes racer).