You're correct that this used to work differently. The reason for the change as far as I remember was so you could use the detune feature to detune agy combination of pu1, pu2 and wav against each other. Before this change, you couldn't detune pu1 an wav against each other. You can of course create another instrument where you set finetune to 0 and play on either pu1 or pu2 to get less detuning.
The detune happens in the Gameboy sound chip\s internal values, which means it's neither measured in cents nor Hz. The frequency, given a certain GB pitch value, is givne by the formula:
f = 131072/(2048-x)
As you go higher in pitch, there are fewer frequency steps, and the error compared to the desired note increases.
Here's a handy table:
http://www.devrs.com/gb/files/sndtab.html
Let\s look at a couple of notes to demonstrate this.
A-3, supposed to be 110 Hz. The closest GB value is value 854. f = 131072/(2048-854) = 109.78 Hz
The next value above it is 855, giving f = 131072/(2048-855) = 109.87 Hz (So actually, it turns out that the table is inaccurate, but that doesn't matter. This is not the note table LSDj is using as far as I'm aware. And I'm just interested in approximately how big a one step difference is in that general range.)
This means that for A-3, the slowest beat frequency you can get between two detuned notes is 109.87-109.78 = 0.09 Hz.
A-5, supposed to be 440 Hz. The closest GB value is value 1750. f = 131072/(2048-1750) = 439.84 Hz
The next value above it is 1751, giving f = 131072/(2048-1751) = 441.32 Hz
This means that for A-3, the slowest beat frequency you can get between two detuned notes is 109.87-109.78 = 1.48 Hz.
A-7, supposed to be 1760 Hz. The closest GB value is value 1750. f = 131072/(2048-1974) = 1771.24 Hz
The next value above it is 1975, giving f = 131072/(2048-1975) = 1795.51 Hz
This means that for A-3, the slowest beat frequency you can get between two detuned notes is 24.27 Hz.