Hard questions to answer, as I don't believe anyone went for such advanced/specific use yet.
For the question about collisions, as I said previously the built-in physic behavior is based on box2D.
You can check this demo that shows quite a few elements on screen with physic behavior. It's perfectly fluid on my FF 9.0.1 on win7 (NVIDIA GeForce GTX 560 Ti).
So I guess JS via C2 sounds powerful enough on computation level (I'm not sure if the webGL rendering helps in any way with the computing of physic though).
And for the AI part, same thing I guess. I think I'm not knowleageable enough on this.
The furthier I went as for "AI"/computing is the pathfinder behavior and it has always appeared to work pretty good on my computer (I started it on XP with a dual core 2.5Gh and finished it on win7 with a quad core 3.5Ghz).
I guess if you really push it you will hit limitations, but hard to precisely tell you "you can do up to X collision at the second" for now 'cause it sounds it depends so much on the gear/OS/browsers/etc...
As you said, at worse you can always do "simpler" thing with C2 it is not totaly wasted money.
Keep us informed about your experimentations/results <img src="smileys/smiley1.gif" border="0" align="middle">
Also moved to the SDK forum, it is less likely to drown answerless.