I wondered about that too. As you say, connecting it to a floating input would probably just leave the charge on Chold since all you have is the leakage current to discharge it.

I noticed in figure 19-4 that they show a 2TAD "discharge" time between measurements, and they say this:
19.2.3 DISCHARGE
The discharge phase is used to initialize the value of
the capacitor array. The array is discharged after every
sample. This feature helps to optimize the unity-gain
amplifier, as the circuit always needs to charge the
capacitor array, rather than charge/discharge based on
previous measure values.
The PIC18 reference manual has the same sort of text, but is also has the statement
The charge holding capacitor (CHOLD) is not discharged after each conversion
so it sounds like they're talking about something to do with the internal ADC structure and not Chold.

It WOULD be interesting to see the same measurements without any averaging and with alternating input voltages for the in between "idle" cycles.