Thanks for the input ikegami, much appreciated - I will be making some changes as a result...

Re: "I don't see the point of determining the capability in advance", that depends on whether "it's a bug to call call the same code_that_touches_db() whether transactions are available or not":

I'm working under the assumption that if a user has a DB that is not transaction-capable (not sure what DBs those might be these days...) that they may decide to proceed at their own risk, assuming the risks are clearly laid out in the documentation. If that's the case, then there's a benefit to checking one time in the constructor, so I don't have to do any extra work while attempting (possibly many) transactions later on.

Not entirely sure if that is the best approach... Maybe I could have the constructor die() if transactions are not possible, unless the user supplies a "no_transactions_are_ok" parameter...