Actually, the Linux kernel holds a mutex (exclusive not shared lock) against the directory during readdir (for example) so re-using the kernel locking means that the processes must do their searching for a matching error (file) in single-file. This is probably part of why directories containing a large number of files are notoriously extremely slow in Unix. (And Perl's glob / readdir is notorious for at least sometimes being pathologically slow under Windows but I don't know what locking Windows uses for directory operations.)
And so the "no extra locking" claim is completely bogus. The extra locking case is "a ton of kernel mutex uses" not the "small number of kernel mutex uses to open a file and then get one shared lock".
I don't see how you justify that the race doesn't result in extra e-mails. I believe your analysis is mistaken there. And you may not care about a few extra e-mails (or tons of extra e-mails when the directory gets bloated and pathologically slow to use) but the person asking the "avoiding a race" question probably does. I can certainly understand caring about my boss getting duplicate e-mails after he assigned me the task of making sure we don't get duplicate e-mails.
But given 300 processes, I'd probably go with a separate, single process that de-dups errors rather than having 300 processes fighting over the list of errors (whether stored as lines in a file or files in a directory).