After reading list and seeing statement that lazy/Lazy.force is faster then
closure/apply, I've decided to run microbenchmark with both.
Results are rather confusing:
 Byte code - Lazy is three times slower than closure
 Native code (x86) - it's either coredumps or leaks memory

Code attached as <<lazy.ml>>. With it's current form native code
leaks memory, by uncommenting first two lines it would coredump.

Also both compilers tends to generate rather strange error.
=-=-=-=-=-=
The files lazy.cmi and lazy.cmi make inconsistent assumptions
over interface Lazy
=-=-=-=-=-=
This error fixable by removing lazy.cmi from a disk.

Have I stepped on some known bug/feature?

Vladimir