Yes, you're missing the fact that IEEE floating point numbers aren't exact representations.

Update: See for instance Equality checking for strings AND numbers for links and discussion.

Another Update: After actually running it and poking with Devel::Peek this is strange because that's showing a 2 for the NV part.

  DB<8> x Dump( $x )
SV = PVNV(0x8e98f0) at 0x800e64
  REFCNT = 2
  FLAGS = (PADBUSY,PADMY,NOK,POK,pIOK,pNOK,pPOK)
  IV = 2
  NV = 2
  PV = 0x435c40 "2"\0
  CUR = 1
  LEN = 36
  empty array
[download]

More poking: If you look at the number the representation's different under the hood, so it's definitely a float precision issue somewhere.

  DB<11> x unpack( "h*", pack( "F", 2 ) )  
0  0000000000000004
  DB<12> x unpack( "h*", pack( "F", $x ) ) 
0  2000000000000004
[download]