BUG: TermRefs with uninstantiated type variables get incorrect signatures

[smarter]

This PR should not be merged, it demonstrates a bug.

After frontend, pickleinf.scala looks like:

package <empty>.type {
  class Bar[N]() extends java.lang.Object() { 
    type Bar$$N
    private[this] type N = Bar$$N
    def bar(name: Bar.this.Bar$$N, dummy: scala.Int): Bar.this.Bar$$N = name
    def bar$default$2: scala.Int = 42
  }
  final lazy module val Test: Test$ = new Test$()
  final module class Test$() extends java.lang.Object() { this: Test.type => 
    def test(): scala.Unit = {
      {
        {
          val $1$: Bar[scala.Int] = new Bar[scala.Int]()
          $1$.bar(10, $1$.bar$default$2)
        }
        ()
      }
    }
  }
}

The bug only manifests itself after unpickling, but the problem is already present when we pickle the type:

sel: Select(Ident($1$),bar)
sel.tpe: TermRef(TermRef(NoPrefix,$1$),bar)
sel.tpe.widen: MethodType(List(name, dummy), List(TypeRef(TermRef(NoPrefix,$1$),Bar$$N), TypeRef(TermRef(ThisType(TypeRef(NoPrefix,<root>)),scala),Int)), TypeRef(TermRef(NoPrefix,$1$),Bar$$N))
sel.tpe.signature: Signature(List(java.lang.Object, scala.Int),java.lang.Object)

Notice that in the signature we get java.lang.Object instead of scala.Int for the first parameter and the return type, this happens because the signature is computed and cached before N is instantiated to Int.
CC @odersky

[smarter]

I guess the problem is that we try to take the signature of bar at a point where the type of its first parameter is an uninstantiated type variable.

[smarter]

I managed to reproduce the problem in master, I'll update this PR and the description.

[smarter]

I think this particular issue could be fixed if EtaExpansion#lift would call fullyDefinedType, otherwise we can get a ValDef with an uninstantiated type parameter which should not be allowed.

[smarter]

Proposed fix at #823.