Suppose you have a network with a hundred devices and an addressing scheme that allows 256. You add an extension such that you can use 65536 addresses, and add another 150 devices. Now you have six addresses left under the old scheme (and lots under the new).
But do you have a way to discover whether any of the first hundred devices break soon? Or even the 150 new devices? You could add a device on a new-only address before you run completely out of the old-compatible addresses, but that won't test completely. There might still be devices that will connect, but log the connections improperly, or have a buffer overrun and crash randomly, or, or, or. There might be software on a device that supports only old addresses, even though the device itself supports the extension scheme.
If the network is a big one, where different parties own and operate the devices, then you may well be completely unable to test whether the extension scheme is universally supported. Before the day when you run completely out of unextended addresses, that is.
Contrast this with a dualstack approach, where you can make a list at any time: THESE devices are on v4 only and will be a problem when v4 runs out, THOSE OTHERS are on v4+v6.
I am overcome by a desire to mention an anecdote... about the financial-services company with three old servers that didn't support the new buzzword, but there was some magic compatibility and everything seemed to work. Until there was a lawsuit, a subpoena and it turned out that data sent among the three old servers used some legacy protocol without mandatory logging as demanded by law.
Could you explain what this means?