After you transform the point by the projection matrix in the vertex shader, clamp Z to 0 if it's negative. This will push everything behind the camera onto the near plane. You need to be careful though, since this means nothing behind the near plane but within the side planes will ever be clipped. You should also have a culling system in place to take care of objects completely outside the light view frustum.

For my orthographic shadow maps, I transform the caster bounds into light space, then simply get the extents of the bounds. The XY extents of the bounds represent the view "window" of the orthographic projection, and the Z extents represent what the Znear and Zfar should be. If the calculated Znear is less than the light space camera's near, then I move the camera back by the difference, effectively placing the light space camera's near at Znear. Of course, for an orthographic transform, this doesn't change the size of the view, results in a tight bounds in terms of maximizing the shadow map area and maximizing the depth precision, and is extremely simple code.

The only problem I've faced so far (apart from the usual depth and slope bias tweaking issues) is that GPU skinned object's bounds are not known when rendering into the shadow map, so I just make an approximation which seems to work well enough. In fact, before shadowing skinned objects, I never needed a scissor rect or anything to prevent the clipping of the casters because they always fitted well. Now I require a scissor rect/viewport/clip region that clips one pixel around the border :-(

perhaps of interest
http://aoeu.snth.net/?p=7