If you have time I would appreciate a full review. If not, I would still appreciate your weighing in on two things:

• How can I test my changes in FieldBase.h to FieldBase<std::string>::get[Const]Reference in Python? I think the changes are necessary if this code is reachable at all, but all of my attempts to access a string field using reference notation failed.
• Is my change to error handling in BaseRecord.cc commit 0c65e9f acceptable? I think the old test was inadequate because it ignored the case that the left-hand side was fixed and the right-hand side variable.

I'll do a full review, but I'm afraid the answer to your first question is that it is not possible to test the reference changes from Python; that would require (at least) a mutable string type in Python.

Thanks for the helpful comments. I adopted all suggested changes except one, as discussed on github.

I have discovered a misfeature with variable-length strings in C++: one cannot specify a new variable-length field unless one also specifies a value for the doReplace flag. E.g.

# this works

schema.addField<std::string>("foo", "docs for foo", 0, false)

# this fails: the compiler says it is ambiguous:

schema.addField<std::string>("foo", "docs for foo", 0)

I'm not sure if there is anything reasonable that can be done about this. It is worrying because the compiler warning does not provide much help in figuring out the workaround.