Over on the chemed-l list, the annual recrudescence of the
sig-digs discussion is in progress.
My take on all this:
-- Counting digits tells you nothing. It gives neither an
upper bound nor a lower bound on the uncertainty.
-- Round off for convenience, provided it doesn't cut into
your accuracy.
-- Keep plenty of guard digits on intermediate results.
-- Do not round to express uncertainty.
-- Instead, express uncertainty separately. Anybody who really
cares about the data does it this way.