It'd be useful for debugging if
_autoload_registry_check_code() watchdogged when an autoload miss occurs. WATCHDOG_NOTICE seems like the appropriate level, since autoload is triggered as part of user-controlled class/interface checking logic (e.g., is_subclass_of(), instanceof etc.) and a miss therefore does not necessarily indicate a critical error.
Note that I've got an issue up against 8.x, which can hopefully also get backported to 7.x: