To quote the Cg tutorial:

"Notice that the code for the shadow map query is virtually identical to the
code for the projective texture lookup in Example 9-6. To a Cg programmer,
shadow maps are like other projective textures, except that instead of
returning a texture color,tex2Dproj returns a value representing shadow
coverage. The underlying hardware implementation automatically recognizes when
a texture is a depth texture (instead of a color texture) and performs the
shadow map comparison instead of an ordinary texture fetch."

I suspect that the Cg compiler always compiles tex2Dproj to a TEX instruction
with the SHADOW2D target and relies on OpenGL implementations always
transparently using normal sampling on non-depth non-stencil textures.