Steve,
Like Terry, I too question the real-world value of a super-fast toolchange time. I can't imagine that it would save more than a few minutes per day at the cost of significantly increased complexity, and opportunity for disaster, due to higher speeds, and generally higher precision required. I think you can quickly get to the point where doubling speed will take 10X the time and 2X the cost, to make it as reliable as a slower, simpler machine.
Re: the speed of a Geneva, this, too, I think is a bit of a red herring. Perhaps a high-speed Geneva will always be a bit slower than a max-performance direct-driven mechanism, but I think the overall difference in toolchange time would still be small. I'm running mine VERY conservatively, and it runs under a second per tool pocket. So, even with a 24-tool carousel, max seek time would be 12 seconds. I'm sure I could easily double that. Beyond that point, you have to start worrying about the tools flying out as the thing moves. The biggest job I ever did only used about 12 tools, and that one took hours to run, so an additional 3 minutes in seek time would be lost in the noise.
I'm not sure if our posts 'crossed in the mail' or if you misunderstood. My setup will not be 'super fast', just fast. And even that is a subjective term. I think the only number I mentioned was 5 seconds, but really I won't know what it will tolerate until it is running and I can push it to failure. What I don't want is 20 second tool changes. Seconds add up when you are multiplying them by orders of magnitude. If you doubt that, just run Vectrix generated peck drilling code for 30 deep holes . . . .

Certainly doing a variable-tool-size chain-driven machine is quite do-able, but, again, at considerable cost in terms of complexity. You'd have to not only program each tool length, but now the "pocket" width as well. And, setting up a job will mean a fair amount of disassembly/re-assembly to move the tools to their required positions. Or, have pre-configured slots of different sizes, but then you have to deal with mapping tool numbers so each tool lands in an appropriately-sized pocket. My biggest tool is a 4" face mill, so I just sized all the pockets accordingly.
The jury is still out on this and I have not really thought it thru. Thanks for you observations. Every comment helps.
The idea probably falls apart without a smart servo drive that can find the tool spots on its own. I had in mind just a certain number of 'small' spots and 'wide' spots. You put your 4" face mill in an existing wide spot and your 1/4" drill in a 'small' spot. The only disassembly would be to alter the ratio of small to large to suit your typical jobs. The position of the tools can be defined as a distance from 'home' and could be calculated easily thus; 3 wide spots x 4" and 2 small slots x 3"; that tool is at 18". All you would need is a table of tool positions stored in the drive or the controller.
It would just be a way to shoehorn a couple more tools on a belt which may be all you need for a specific job. If I build in the capability, that doesn't mean it has to be used. Even if the slot width was variable, they could all be set the same, so no harm no foul . . . . IF it turns out to be easy to implement. If it does result in a lot of complexity, then I would agree with you that it may not be worth the effort, but I don't know that yet. Certainly the tool postions could be completely random and be probed, but that is most likely over the doable-but-not-practical line.
The chain belt option is a definate, but variable spacing is just a discussion point at the moment.