I just tried (and failed) to write a test for the recent magic
inference of phrase searches. That's a feature that makes me *really*
uncomfortable to not have an automated test. But I believe the
proposed modularization of the test suite should reduce some quoting
nightmares, so will hopefully make this easier.