I'm finding this thread interesting. I recently acquired a 7d and have become quite frustrated with some focus behaviors that appear essentially random.
I'm mostly concerned with critical focus on the eye of a subject with a narrow DOF (large aperture). I have literally thousands of these shots in my portfolio shot with a 30d (and prior Canon DSLRs) and a number of lenses (primarily 50mm, 17-50mm, 70-200mm, but others as well at times) at very large apertures, and the behavior of the 30d has always been predictable.
The eye of a typical subject will always have high contrast areas, whether it be the pupil/iris, the eyelid/sclera, the eyelashes/eyelid, etc. On the 30d, it was certainly important to know where to place the focus point (generally the center double precision point) and to accept that some small percentage of shots may have the corner of the eye in critical focus, so great care should be given to subject positioning and focus methodology.
I will generally shoot in one-shot focus, but as conditions dictate, I may shoot in AI-servo. Similar results are expected and experienced with both modes, as long as they're used judiciously.
On the 7d, I've experienced (as stated above) what appears to be essentially random focus behavior. I've used several lenses and especially in low-EV situations (EV 6 or less), I will see back focus OR front focus (without any evident pattern) by no less than 1x - 2x DOF. This means that if my DOF is only 2", the shot will be back focused or front focused by 2" - 4" -- which means that the focus target (the eye) is completely out of focus. No part of the subject which is in focus is even remotely near the active AF point that was used for the shot. No focus and recompose is being used in my testing.
I have run dozens of test shots in both AI-Servo (to try to rule out any for or aft subject or camera movement, although certainly nothing has actually changed by simply picking up a different camera body) as well as One-Shot focus mode, and I'm seeing absolutely no consistency whatsoever. I've used both manual point selection as well as spot manual point selection, and have used the high precision center point as well as peripheral cross-point AF points.
I've even gone so far as to shoot test charts and static subjects (on a tripod, of course, with MLU and a remote release), tried to MA the lens, and used other methods to try to definitively attribute the behavior to a controllable variable- but yet I'm still seeing arbitrary focus come and go. I might get a number of shots sequentially with perfect focus, and then see random front and back focus -- while the camera is on a tripod shooting a static subject in One Shot AF mode, with spot manual AF point selection, and focusing on a subject with high contrast in the focusing area. Of course, in One Shot AF mode, the camera is happily blinking active AF point and beeping to indicate that it's acquired focus.
I'm certainly aware of the behavior, positioning, precision, and size of the AF points on Canon cameras (having owned and used more than half a dozen models since the D30, which is sitting on my desk as I type, was released in 2000) and at this point I'm confounded by the AF behavior of my 7d.
It is absolutely capable of amazingly sharp, detailed photos, has great autofocus, and is overall a stunning camera, but where the focus really matters (low DOF, wide aperture, low light), it's simply not consistent.
The more I mull over this, the more I see parallels with the 1dMk III, where the sub-mirror was actually returning to a slightly different resting position after each shot, causing a physical difference in light-path distance to the AF-sensor, and a corresponding "random" focus behavior. I know this was primarily experienced in AI-Servo mode, but I still recall people complaining about random One-Shot AF performance as well. I'm not saying that this, or anything like it is the cause of the problems that some very small number of 7d owners are experiencing, but merely noting that "random", arbitrary, or unpredictable behaviors do exist and sometimes have origins in hardware as well as software (firmware).