Butler reading information (in particular the observation time and length) out of an input dataset's file representation in order to provide rendezvous with calibration data in another repository (that does have a registry). Today, that read is handled by genInputRegistry.py so that the butler doesn't need to look into the dataset itself. If there's no registry, such a read will be necessary.

The ultimate test is to run processCcd.py (a CmdLineTask that runs Instrument Signature Removal) on a repository that contains raw frames that require more than one set of calibrations and have it pick up the correct ones.  That would have to be condensed down into a unit test.