One bug fix and a simple GEOS geometry cache to improve performance when wrapping coordinate data stored outside of Shapely. http://pypi.python.org/pypi/Shapely/1.0.4.

In comments on a post about Javascript geometries, Paul Ramsey and Martin Davis wondered about a pure Python version of JTS. It might be an interesting project, and maybe you could coax reasonable performance out of it with psyco or Cython, but it wouldn't perform as well as something based on C (for C Python) or Java (for Jython) or C# (for IronPython). Consider Ian Bicking's HTML parsing benchmarks: lxml, based on libxml2, is far faster than anything else. The need to calculate spatial operations in C rather than Python is greater yet.

If Google would provide some basic spatial indexing (for App Engine), you could try to go the rest of the distance with the help of an external web geoprocessing service. You might not need a PyTS after all.