Typically D is a function of lambda, such as a half wave antenna for example. You have D squared on top so the overall value does indeed drop with reduced lambda. Is this enough of an answer?

Nothing significant happens at the boundary. Everything varies smoothly. It's arbitrary. The Universe has been around a lot longer than the definition of the boundary.

I've seen other definitions.
...
2) Where the wavefront becomes reasonably flat (rather than spherical).
3) Where a certain approximation in GTD/UTD becomes reasonable.

The 2D^2/Lambda is really not a boundary.
Typically it's used as a required distance for anechoic chamber antenna testing. i.e. length from your unit under test to the transmit antenna at the other end of the antenna chamber.
This 2D^2/Lambda distance represents about 23 degrees phase error in a spherical wavefront across your antenna from a point location at the center of your transmit antenna. D is the largest dimension corner to corner of your test antenna (or reference gain standard). "D" size is sometimes important on your transmit antenna in the chamber, but not often since there is some assumption that your chamber has been tested and works properly (good phase and amplitude taper).

Often times a customer will ask if you've tested at 2D^2/Lambda. If you tested at say 1/4 this distance, your results are suspect. Shallow antenna pattern nulls is typical evidence of being too close. Low measured gain is another.
D^2/Lambda does pretty well to for testing and I've read it's ok for "engineering tests".
0.4D^2/Lambda is the distance for peak field from an antenna.

kch