just to name a few: argument passing, process scheduling, register allocation, big vs little Endian - I realize this is a weak list but personally I expected more stuff to have some reasons behind the decisions.

Those are not decided by arbitrary conventions. They're usually chosen by some sort of criteria like performance, simplicity for teaching, portability, etc. There's definitely flexibility in choosing the specifics, with a lot of trade-offs to consider, but they do have specific design criteria and are not just conventions. They are conventions in the same sense that TCP/IP is a convention -- since if everyone did something completely different, there would be no interoperability.

Calling conventions (argument passing protocols) are designed to be efficient for a particular CPU architecture. Scheduling algorithms try to make the best use of all resources while still meeting other performance criteria (e.g. firing events at roughly the right time). Register allocators are usually trying to produce the fastest code possible. And there are arguments both for and against big and little endianness, but I don't remember what they are. :-)