It's always good to get a glimpse into the approaches taken by non-OSS JCR implementations: In a recent technical article on the developerworks website Malarvizhi Kandasamy describes how IBM goes about JCR fulltext search. The actual engine is
Juru, which is a Java library developed by the IBM Haifa research lab
According to the article Juru is capable of some natural language processing like stemming or finding similar spellings.
IBM uses a JCR compliant repository in a number of their products, e.g. Lotus Web Content Management or WebSphere Portal.