Here's a very simple-minded patch (with microtest) that changes behavior
in a questionable way.

If it goes in as-is, dry_run=1 will not always succeed anymore. So it'd
be incompatible... but a 1:1 representation of a real move kinda
requires some failure mode.

It's possible to avoid raising when dry_run=1, logging a warning that a
real move would fail, and return the imaginary new full name. Or return
None, '', etc. Suggestions?

Perhaps studying a use-case makes things clearer. 

Looks like copy_file already has the same return value independent of
dry_run.