I don't have an answer, per se, but I have some input:
There's a magic value to interval in an inputs.conf script: specifier (I think it's zero) that means "run once on startup". But then you'd have to restart the forwarder a lot.
If I were asked to implement this, I would put the logic in my script. I'd use a persistent state file (maybe in $SPLUNK_HOME/var/lib/splunk/persistentstorage) to keep track of the last stamp we'd seen, then use this script to inspect the file on each run. Most executions of the script are going to be no-ops "data hasn't changed!", but when it does execute the decoding loop, you'll want to make sure that the interval of the script is longer than the amount of time a single run would take. Splunk shouldn't start a second instance until the first is done, but I would schedule my runs so that I didn't have to rely upon that.
