Two things spoken by Guido:

But millions of tuples are not uncommon. They're probably the only thing to worry about here.

Unfortunately, the visit API doesn't make it easy to watch this[*]; a tuple calls visit() on its items but learns nothing except whether it failed. (I've never seen a visit() implementation that could fail, so I'm not sure even why the return code exists.)

* "this" is the number of tracked objects visited during a GC scan, as described by MvL

If visit() does something sensible (like traversing a directed, potentially cyclic graph), then the only way it could fail is to not return. No doubt there is a clear record of who calls this function.

Why not we change the semantics of visit() in this case to provide the required information? Make visit() return a non-negative integer for "number of tracked objects seen". If someone can find a good reason for error codes, then there is the negative half of the integer number line. (/me is going out on a limb and hypothesizing that visit() is in C for all relevant cases.)