Note: the AVM2 specification says that dxns/dxnslate opcodes should be rejected at verification time if the SET_DXNS method flag isn't set. (relevant avmplus code: https://github.com/adobe/avmplus/blob/master/core/Verifier.cpp#L1169)

I don't think the stack is correct here; what FP does kinda resembles the stack, in that when you call a function and set dxns, when you exit from the function the callee has its previous dxns back. But this alone could be implemented by just setting dxns as a field on the Activation - and maybe this is enough for an MVP?

But in reality, in FP this is more complex unfortunately, in avmplus (see MethodFrame::findDxns) it's a property of MethodFrames/MethodEnvs (kinda what our BoundMethods are)/scopes. It recurses MethodFrames... presumably because sometimes the current frame might inherit dxns from the parent? But in practice, with normal function calls, a function doesn't inherit and just uses its own default namespace so not sure what that's about.

The easily visible difference is that the default namespace of the MethodEnv depends on the environment (scope chain?) at the point in time when it was created, see the example:

function f(){
    var a = <xml xmlns:a="a" xmlns:b="b"><data x="1" a:x="2" b:x="3">asdf</data></xml>;

    trace(a.data.@x); // 1

    var inline_f1 = function(){ trace(a.data.@x); };

    default xml namespace = "a";
    trace(a.data.@x); // 12

    var inline_f2 = function(){ trace(a.data.@x); };

    default xml namespace = "b";
    trace(a.data.@x); // 13

    inline_f1(); // 1
    inline_f2(); // 12

    create_custom_function("")(a); // 1
    create_custom_function("a")(a); // 12
    create_custom_function("b")(a); // 13
}
function create_custom_function(nsname){
    default xml namespace = nsname;
    return function(a){ trace(a.data.@x); };
};

But again, I don't exactly understand how the chained MethodFrame iteration in avmplus relates to this, I don't know how to observe it.
For MVP, maybe just setting dxns on the current Activation would be enough? (and make a failing test for the more complex behavior)

Superseded by #22761