Var parameters have a defined value on entry. OUT parameters not, hence that the fuzzer also randomizes them.
That's reasonable and it is a nice feature to ensure that the caller gets a "return" value but, an out parameter isn't a variable allocated in the callee's stack. The -gt switch, as documented, is used to ensure that local variables (by definition, allocated on the callee's stack) are initialized before they are used by the callee, or to make it obvious that one or more of them weren't before they were used. It's nice that it does that for out parameters as well but, it is beyond its document purpose.
It's also a bit problematic to implement the feature that way because its proper (say, consistent) operation depends on how the compiler has chosen to pass the parameters and the sequence in which the "fuzzer" is going to scramble the values. The 32bit version of the program is a good example, the const is passed by value, the out by reference, the out gets scrambled but it doesn't scramble the parameter passed by value, that's why the 32bit version works, because a unique location has effectively been split into two separate locations. The 64bit version passes the parameters in registers which ends "working" in the sense that the unique value does get scrambled.
Anyway, as a result of looking into this problem, I noticed that the fuzzer uses 1 of 3 values (if I recall correctly) none of which is particularly interesting. A bunch of 5s sometimes, a bunch of As in other occasions. Microsoft chose to use $CC, which is the opcode for int 3. If the stack gets corrupted, when the ret is executed, it's likely, though, not guaranteed of course, that the code will end up executing a breakpoint, which is good because, it's consistent. A bunch of 5_s or a bunch of A_s could end up doing god knows what. For that reason, I think the FPC team should consider using $CC to scramble the values. Ends up being simpler too. Always use one value instead of picking one out of three.
Peace.
ETA:
If anything, the doc may need an update to mention this.
I was typing my reply while you posted yours. I wholeheartedly agree with your suggestion.