Eventually, we needed more flexibility than Solr easily offered though. For example, we've added far more efficient sharding, document modifications (updates and deletions), flushing, and near real time search than either Lucene or Solr support out of the box (and they were much easier to add to Lucene than Solr, since Lucene makes fewer assumptions about your dataset/use).
I think Solr is a great tool if your needs happen to fit into their model - but if they diverge a lot, it sometimes makes more sense to build your own custom framework on top of Lucene.
We didn't use them outright since we have fairly different requirements/constraints (our data has some pleasant properties that makes it easier to shard and facet than the general case) and we wanted something a bit simpler.
Shoot me an email sometime though (email in profile)! I'd love to buy you lunch and pick your brain ;-) You guys clearly know what you're doing!