A.scala

// class Q

object A { val x = 3 }

B.scala

object B { val x = A.x }

1. compile
2. Uncomment line in A.scala.
3. compile. Both sources are recompiled because Q introduces a change to the hash for <init> (in the compilation unit A.scala).

The underlying reason is that name hashing is tracking hashes for names at source file (compilation unit) level. The hash for <init> includes hashes for both A.<init> and Q.<init>. Therefore name hashing works well only when a member that is modified doesn't introduce any nested members. Therefore, it works for changes to val and def but not to class and object.