Documenting the guarantee that fingerprinter won't emit duplicate tokens for the stame field.

[fgregg]

Right now this is true because we are careful to make sure that every predicate returns unique keys.

It would be safer, and sometimes more efficient to move the code to the fingerprinter itself.

something like

block_keys = {block_key + pred_id
              for pred_id, predicate in predicates
              for block_key in predicate(instance, target=target)}
for block_key in block_keys:
    yield block_key, record_id

[fgregg]

one reason why we did it this way is to reduce large cartesian products for compound predicates. making this happen later would clean that up allow that to happen.