writing interactive music for ensemble ?

You will get bleed-over from the microphones no matter what you do, and this will make itself apparent more with fiddle than with most other objects.
The best advice I can give is this: working with Max/MSP in real-time with microphones requires more preparation on stage than a normal concert. It is similar to using video in real-time; alot of time has to be spent finding solutions for the acoustic problems. One of the primary problems will be the placement of the mics in relation to the loudspeakers, and the choice of microphones will play a big role. The use of sound-absorbing screens can help alot, as it does in recording classical music.
Whenever I start a new piece, I begin working directly with the setup I will use at the end, so that I don't start making decisions which I later have to change or abandon. Basically, the use of the setup becomes an important, formative influence on the music. The difficulty then becomes when I perform the same piece in another space; I have to make sure that I have plenty of time to try things out before the performance; sometimes an entire day (for one piece, there was a three-day buildup time.) I don't write this to be discouraging; I find this sort of work fascinating and I realize that the work I put into the preparation is an important part of creating something which cannot be had in any other way.

As for the music: I have used a great number of physical switches in some pieces; in this way the musicians have direct control of what signal goes where and the composition can be in a tight, exact format. If I use threshold switches, or fiddle, or bonk, then I have to use the resulting error quota as part of the piece. I can reduce this quota alot by constructing elaborate filters and gates before sending a signal to the object, but the errors remain.

Max/MSP seems like a "well without a bottom" (as a technician once told me) but with a bit of analytical consideration, you can begin to see that the real possibilities become variations of a few basic catagories. For example, what can you do with live instruments? You can sample and loop them, apply distorting filters and effects and analize thier signals to use as switches and controllers for those samples and effects. These three basic directions describe everything I have ever done with live audio. Of course, the results may well be infinite in their scope, but the basic operations you can perform are limited in catagory. For this reason, I have a few standard patches which I use over and over again, and each is constructed in a way in which I try to cover the whole gamut of what should be possible with the basic concept. An example: my sampling patch can sample in real-time, immediately loop what was sampled, cut the sample into parts and also save the sample for later use. I have another patch for threshold triggering, another for switch-triggering, another for sequencing events, etc. With time, I only need to mix and match these patches to get the results I want.

With live instruments, I try to make an audible difference between what is played directly and what is looped, echoed or recorded. Otherwise, it would be just as well to use pure playback and use the instrumentalist as an actor (hey; it works for the pop-music industry...)

Hope this helps.