Electronic – Downscaling Sensor output

A comparator running at 5V and a voltage divider will work fine, you wont be attenuating the signal in any non-linear fashion so the signal going into the comparators will work for you.

You stated the input voltage will be a range (6 - 50V) and as you said you would use a 10:1 divider, then at 50V your max input would be 5V and at 6V your max value would be .6V. I'm sure any decent comparator will be able to distinguish differences in the 10's of millivolts (conservatively). I looked over the Microchip application note you linked, The circuit shown on page 14 looks great, it provides a virtual center point so you don't need a positive and negative voltage supplies.

Concerning the power supply it seems as though your best bet would be to supply the higher voltage (6 - 50V) then use a buck converter to bring the voltage down to 5V, something like the TI simple switcher designs like http://www.ti.com/product/lm46000 would work (I don't work for TI, I just like their simple switchers). They even have an example schematic.

Yes, an op-amp unity-gain buffer is a reasonable approach in the ADC does not have a high-impedance input.

The minimum value of the sum of the two resistors is determined by how much current you can draw from the source without unduly affecting the accuracy. The power dissipation might also come into play if the resistors are low value.

The maximum value of the two resistors paralleled is determined partly by how much error you can tolerate due to op-amp bias current and/or leakage. Also practical considerations (resistors of lower value tend to be more stable, at least down to 1M or 100K). For example, metal foil resistors are not available much above 100K.

For relatively low accuracy applications, a few M ohms is fine, and the bias current error will depend on the op-amp you select.

If you follow their app note on the op-amp part (looks like there is an error in the plain divider calculation) they suggest an OPA344 - bias current should be less than 1nA at reasonable temperatures, which implies (for < 0.5LSB error on a 12-bit converter):

\$R1||R2 < \frac{5.0}{2^{-9} \cdot 2^{13}} \approx \$ 30K

Your 105K || 3.3M ~= 100K would have a bit more error at high temperatures, but still should be acceptable for most purposes.