The logic for how to communicate with the compiler server was spread out amongst a number of types in the code and it was causing a few problems:

1. Made the server protocol hard to understand because you had to dig through 3-4 different types to understand the interactions.
2. Made testing real scenarios difficult because the logic was in the wrong places (why should the csc.exe know anything about how connections are managed?)
3. Lead to too many knobs being put into the APIs that created scenarios that needed to be tested that never actually occurred in production code.

This change was mostly spurred by (2) above. In a recent bug fix I ended up spending far too much time trying to write a very simple test. It motivated me to finally fix this up.

The bulk of the change is in BuildServerConnection.cs. That is now the central point for all logic around how the server is started, build requests are processed and how the server is shut down. All logic for that set of interaction should now be centralized here.