|Welcome to the Monastery|
I hate having to say, "Oops".
Everything that I said about BEGIN is true. But my promise that Perl 5.8 will have a version of Carp which fixes the begin-time Exporter issue is incorrect. Here are some details on the situation.
Suppose that you have 3 packages, A, B, and C. Suppose that B and C both inherit from A but C does not inherit from B. Today when C makes a call into B and then B makes a call into A which generates an error, Carp will not label the error as coming from A, B, or C, but instead will report it from whoever called C. In 5.8 the error will in this case be reported from C because Carp has not been told that B trusts C.
This is good because the fact that B inherits from A is an internal implementation detail which C should not have to know about. What is not good at the moment is that the trust relationship is synonymous with "inherits". But the code for Carp has been written in a way where this can be fixed by just replacing the trusts_directly() function in Carp::Heavy with something smarter. This is all intended.
What I had not thought through, though, is the situation where C calls a method in B, which actually is inherited from A. Now when Carp looks at the information, what it sees from caller is that C called a function in A. But C trusts A, so that call is trusted. In fact if the call generates an error, then C should be fingered because B does not trust C, and where the error is reported from should not depend on that implementation detail. But the critical piece of information required is not reported by caller, on a method call we don't know what call was intended, we only know what function was called. If we had a way to get that information, then set $called to the appropriate thing in the short_error_loc() function in Carp::Heavy, and the underlying problem would go away in Perl 5.8. But unless someone feels motivated to champion this on p5p, that won't happen and in this situation the error will continue to be incorrectly reported.
This last situation is exactly what we get with Exporter and playing games with @ISA during compile time. If you do not play the games, then while your code is being compiled you do not inherit from Exporter during your use invocations, so you can be correctly fingered as causing errors. If you do set @ISA bright and early, your use of other packages will go to Exporter's import() function, and then when Carp tries to figure out the error it will decide not to finger you because what it sees is you calling a function in a package you inherit from.
An incidental note. Capturing information about method calls in caller would allow that information to be properly propagated in Perl's SUPER mechanism. This is a necessary step towards having Perl's SUPER mechanism play nicely with multiple inheritance. Of course making that change in Perl 5 raises backwards compatibility issues, so likely a new keyword would need to be created. That would require even more championining. Since I am personally of the opinion that multiple inheritance leads to overly fragile and complex design issues, I won't be seen caring much either way, but some others might...