I think I may have found the answer.  It is related to Reverse Short Channel Effect (RCSE)

Here's what some guy wrote:

Arya Raychaudhuri
3rd June 2004 - 05:32 PM
You can look at it like this. For example, if you plot the surface
potential along the channel region of the MOSFET near but below the
threshold voltage (by a device simulator, perhaps), you will observe
high surface potential near the source/drain
edges (almost 0.55 + phif), and then it gradually falls below the 2*phif
value near the middle of the channel. As you know, the inversion of the
substrate (channel formation) takes place when the surface potential
at the center of the channel reaches 2*phif. But, beyond 2*phif, it
requires higher and higher gate voltage increments to increase the
band bending (surface potential). For short channel MOSFETs, the
gradually falling surface potental gradients adjacent to the central
region help increase the band bending at the channel center. So, to
an extent, this helps lowering the threshold voltage. At the same time,
it has been seen that the substrate doping concentration at the
center of the channel region increases slightly when the channel length
is decreased. So, this is an effect that should increase the threshold
voltage. So, the threshold voltage can go thru a maxima as the channel
length is decreased from long channel.
---

And here is a more concise explanation that makes more sense to me:
RSCE in nMOSFETs is mainly caused by the boron
dopants pile-up phenomenon at the edge of the source and
drain regions.

So, if the channel is really short, there is little space between source and drain; it is mainly made up of these high-concentration regions.  High concentration requires a higher Vth.  Makes sense to me!