This is an interesting idea. But I have two concerns.

1. It is hard to implement refcount-based optimization on Python implementations which do not use reference counting (i.e. PyPy). If the effect of this optimization will be significant, people will become writing a code that depends on it, and this will cause problems on other implementations.

2. Currently list1 + list2 + list3 returns a list which allocates the exact amount of memory needed to contain its content. But with the proposed changes the result list could preallocate more memory. If the result is a long living object, this can cause to wasting of memory.