Interesting. It occurs to me that SQLite, even, is overkill for this purpose. A relational database isn't necessary and the unavoidable duplication of data is undesirable. All that is needed is an R-Tree index and, eventually, the capacity to use spatial operators and predicates in views. The Rtree package that Howard and I have been working on could serve well. It need store only the keys of CouchDB documents, so there's no duplication of spatial data. SpatiaLite (which appeals to me in many ways) uses the same geometry library as Shapely does (GEOS), so a Python query using Rtree and Shapely has essentially the same implementation as an OpenGIS SQL query of SQLite/SpatiaLite. I suggested in comments that GeoCouch might want to take advantage of the GeoJSON group's work, on geometry formatting at least.
Via Simon Willison.