Query time has the obvious drawback that this makes the query more expensive.
However, the clustering results at query time are supposedly better, because at that time, more information has been seen and user feedback can be incorporated.
Note that technically, this is probably more frequent pattern mining than cluster analysis.
Maybe you should just try this variant of frequent pattern mining on your whole data set. You might not even need to store which documents were tagged which way - the solr engine should already be optimized to retrieve them again when needed.