I didn't think about this issue! But this would mean that it is unsafe to have an (external) process create a file, and then use it in my program - at least under Windows, which is very picky about this kind of stuff?

Here's the thing. If you want to delete a file that you've been accessing, from the script you've been accessing it from, you need close it first. If you've also run some commands via system that accessed that file, and they have left open, duped handles in the kernel, they will disappear as soon as you close your handle.

So normally, the problem does not arise. Indeed you can verify this by running your OP code with the binmode commented out.

It was only necessary to remove all the subprocesses from your example in order to confirm, by dint of 'it couldn't be anything else', that is was the perl process that was holding the open handle.

It then becomes clear that the "other process" is actually 'this process'; and thus the close is failing silently when binmode has added the ':unix' IO layer to the stack.