Unless you are dealing with a surveyed peak, I would assume all the
elevation data is derived from USGS NEDs (national elevation dataset).
For the most part, they are only 1/3 arc second. I'm speculating that
some of these programs are reporting data from the NEDs as if they are
section of flat areas on the 1/3 arc second grid, while others are
interpreting the elevation using a weighting scheme.
http://seamless.usgs.gov/

Have you considered running SPLAT! ? It uses 1/3 arc second data. It can
predict line of sight. Generally I find I do better than the SPLAT!
prediction, so if SPLAT! says no, the answer is maybe, but if SPLAT!
says yes, you have line of sight. The program has hooks for predicting
signal strength, models knife edge diffraction, etc.
http://www.qsl.net/kd2bd/splat.html

Similar software is
http://www.cplus.org/rmw/english1.html

For pure line of sight, there is GRASS. However GRASS has a very steep
learning curve.
http://grass.osgeo.org/

What takes maybe 5 minutes on SPAT can take 2 days on GRASS. I don't
know if the results are that much better.

For the bay area, I suspect SPLAT! is fine. I think it's shortcomings
are in areas where there is really rugged terrain that probably isn't
modeled well.