Comment # 3 on bug 1049186 from Jan Matejek

(In reply to Bernhard Wiedemann from comment #2)
> Then maybe the order of processing them makes a difference
> and we could avoid that by always processing them in the same order.

that's a possibility, but...

> Is there asynchronous garbage collection
> that could matter for the reference counts?

...but you're not supposed to GC references that live inside a
deterministically compiled code block.

It seems somewhat more likely to me that the filesystem ordering is a red
herring and the problem is actually caused by hash randomization (which changes
order of elements in dicts, this could lead to something like "first listed
element has one more reference because it's first")

but we'll see