> Are there any plans to add STM to OCaml?  And I'm talking highly 
integrated 
> into the language, not just some bolt-on library.

I implemented 'some bolt-on library' as you call it for
myself (for understanding the details of STM) but did
not release it. The main lesseons I learned, however, are

a) A library will be nearly as good as something 'highly
integrated into the language'. The main difference is
that you cannot annotate fields to be transactional,
and define transactional record types; you need to
work with explicit transactional reference boxes and arrays.
This appears tolerable, and it is a lot simpler, both in
terms of implementation and in terms of using it.

b) The version of STM as designed for Haskell makes essential 
use of laziness for composing transactions in a very clean
way. You can construct more and more complex transactions
by not yet enclosing them into 'atomic'. This is probably
the most important benefit of STMs at all. Unfortunately,
this property is essentially impossible in OCaml: No matter
how highly integrated STMs are in the language, it will 
always be easy to write a program that uses ordinary (non-STM)
state to keep a record of something it shouldn't---messing
up semantics of course.

Sorry, I know this is bad news, but that is what I found.
I invite everybody to verify this for him- or herself,
or prove me wrong.

Sebastian.