This addresses issue #2.
Results of timing comparisons (using a slightly modified implementation that's more amenable to simultaneous benchmarking in a single Perl interpreter) can be found in arc@ba54b38; see the notes at the bottom of the timing.pl file in that commit. Short version: we get about 12% performance improvement from caching AUTOLOAD methods.
Cache AUTOLOAD-generated methods for future use
I'll be merging this soon, tomorrow most likely.
Thank you, @arc! This is great!
FYI: release 1.950 uploaded to CPAN