The way that ArcGIS computes watersheds probably doesn't match with your intuition: only areas upland of a specific pourpoint will drain to that spot, so if you locate the pour point at the center of the lake, 'lower' elevation spots drain into other basins. As Dan mentioned, instead you want to locate it at the top of the outflow stream, where the water from the lake heads to rivers below.
Another option to produce this effect is to continue locating the pour point in the center of the lake, but build an artificial 'dam' at the bottom of the watershed by 'burning': create a new line feature, convert it to a raster with a large value (say, 300). Now the DEM should be forced to empty at your desired pour, with the only artifacts being introduced by the manually produced dam itself.
If you're looking for something more sophisticated, you may want to look into the ArcHydro tools (beta version for ArcGIS 10), which provide a more detailed set of tools for working with hydrologic features -- see this tutorial for details.
I wouldn't clamp the values, as suggested by the pseudocode, because the curvature artifacts do not necessarily correspond to high-curvature areas. Better cures for the problem include:
Consider smoothing the DEM before computing the curvature. This would slightly reduce most curvatures but might give reasonable curvature values at all locations. Smoothing the DEM can be done with local medians or means.
Find a robust way to estimate curvature. (You won't find this in any Arc* software, but with the availability of general-purpose tools like R, this is a feasible recommendation.) In general, curvature is estimated by fitting a quadratic (or some specialized low-order polynomial) in each neighborhood. The usual fitting method is least-squares, which can be sensitive to slight local variations. By expanding the neighborhood and using a fitting method that downweights outlying residuals (such as IWLS), you might be able to get representative and accurate curvature estimates everywhere.
Replace the extreme curvature values with NoData. This is still a draconian measure, but would be appropriate if the extreme values are really artifacts of the interpolation (and don't, therefore, reflect truly high curvatures).
Choose an interpolation method that produces smoother surfaces!
The 'map algebra' syntax for replacing values with NoData uses the SetNull command, as in
SetNull(Abs([Curvature grid]) > 10*Abs(m), [Curvature grid])
By the way, m should not be the median curvature: the median often will be close to zero and won't reflect typical large curvatures (positive or negative) A large multiple of the median absolute curvature might be a reasonable choice. Better yet, examine the histogram of curvatures to find where to cut off the values.
If the SetNull approach pokes too many NoData holes in the grid, think about whether the curvature grid should be smoothly varying. If so, and if those holes tend to be isolated (or in narrow stripes--which would be a strong indicator that a poor interpolator was used), you can fill them by computing a focal mean and pasting the mean values over the holes. The pasting is done with the conditional operator, as in
Con(IsNull([Modified curvature grid]), [Focal mean grid], [Modified curvature grid])
Note the use of IsNull to detect the NoData cells.
Best Answer
You can go to Data Management -> Raster -> Raster Processing -> Clip, and Spatial Analyst is not needed. There just check the checkbox "Use input features for clipping geometry".
From help: If the checkbox "Use input features for clipping geometry" is checked, then the output raster is clipped based on the perimeter of the polygon shape. If the checkbox is not checked (default), then the raster is clipped based on the extent of the polygon.